en_url stringlengths 47 153 | en_title stringlengths 12 118 | en_content stringlengths 16 19.4k | zh_url stringlengths 53 159 | zh_title stringlengths 4 70 | zh_content stringlengths 16 10.5k |
|---|---|---|---|---|---|
https://developer.nvidia.com/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | AI for Climate, Energy, and Ecosystem Resilience at NVIDIA GTC 2025 | From mitigating climate change to improving disaster response and environmental monitoring, AI is reshaping how we tackle critical global challenges. Advancements in fast, high-resolution climate forecasting, real-time monitoring, and digital twins are equipping scientists, policy-makers, and industry leaders with data... | https://developer.nvidia.com/zh-cn/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | NVIDIA GTC 2025 上的人工智能促进气候、能源和生态系统复原力 | 从减缓气候变化到改进灾害响应和环境监测,AI 正在重塑我们应对重大全球挑战的方式。快速、高分辨率的气候预报、实时监控和数字孪生技术的进步为科学家、政策制定者和行业领导者提供了数据驱动的工具,帮助他们了解、规划和应对一个变暖的星球。
在 3 月 17 日至 21 日举行的
NVIDIA GTC 2025
大会上,思想领袖、科学家、开发者和创新者将重点介绍 AI 如何帮助塑造更具可持续性和韧性的未来。以下会议展示了 AI 在气候预测、灾难缓解和保护工作中发挥的作用,帮助社区适应日益不可预测的世界。
加强极端天气预测和灾害响应
随着全球变暖加剧,极端天气事件变得更加严重和频繁,社区需要更快、更精确的自然灾害预测和响应策略。AI 正在改进洪... |
https://developer.nvidia.com/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | Automating GPU Kernel Generation with DeepSeek-R1 and Inference Time Scaling | As AI models extend their capabilities to solve more sophisticated challenges, a new
scaling law
known as
test-time scaling
or
inference-time scaling
is emerging. Also known as
AI reasoning
or
long-thinking
, this technique improves model performance by allocating additional computational resources during inference to ... | https://developer.nvidia.com/zh-cn/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | 使用 DeepSeek-R1 和推理时间缩放实现自动化 GPU 内核生成 | 随着 AI 模型扩展其功能以解决更复杂的挑战,一种称为“
测试时扩展
”或“
推理时扩展
”的新扩展法则正在出现。该技术也称为
AI 推理
或
长时思考
技术,通过在推理过程中分配额外的计算资源来评估多种可能的结果,然后选择最佳结果 (neural network),从而提高模型性能。这使得 AI 能够以类似于人类剖析复杂问题并单独解决这些问题以达成最终解决方案的方式,制定战略并系统化地解决复杂问题。
在本文中,我们将介绍 NVIDIA 工程师完成的一项实验,他们在推理过程中使用最新的开源模型之一 DeepSeek-R1 模型以及额外的计算能力来解决复杂的问题。该实验旨在自动生成 GPU 注意力内核,这些内核在数值上是正确的,并针对... |
https://developer.nvidia.com/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI Foundation Model Enhances Cancer Diagnosis and Tailors Treatment | A new study and AI model from researchers at Stanford University is streamlining cancer diagnostics, treatment planning, and prognosis prediction. Named MUSK (Multimodal transformer with Unified maSKed modeling), the research
aims to advance precision oncology, tailoring treatment plans to each patient based on their u... | https://developer.nvidia.com/zh-cn/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI 基础模型增强癌症诊断并实现个性化治疗 | 斯坦福大学研究人员的一项新研究和 AI 模型正在简化癌症诊断、治疗规划和预后预测。这项名为 MUSK (Multimodal transformer with Unified maSKed modeling) 的研究旨在提高精准肿瘤学,根据每位患者独特的医疗数据为其定制治疗计划。
“多模态基础模型是医学 AI 研究的新领域,”放射肿瘤学副教授兼研究高级作者
Ruijiang LI
说。“最近,我们为医学领域开发了视觉语言基础模型,尤其是在病理学领域。但是,现有研究使用的现有基础模型需要配对的图像 – 文本数据进行预训练。尽管我们付出了大量努力,最终打造出 1M 病理图像文本对,但它仍然不足以完全捕捉整个疾病谱系的多样性。”
在考虑患... |
https://developer.nvidia.com/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA Toolkit Now Available for NVIDIA Blackwell | The latest release of the
CUDA Toolkit
, version 12.8, continues to push accelerated computing performance in data sciences, AI, scientific computing, and computer graphics and simulation, using the latest NVIDIA CPUs and GPUs. This post highlights some of the new features and enhancements included with this release:
N... | https://developer.nvidia.com/zh-cn/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA 工具包现已支持 NVIDIA Blackwell 架构 | CUDA 工具包
的最新版本 (版本 12.8) 使用最新的 NVIDIA CPU 和 GPU,持续提升数据科学、AI、科学计算以及计算机图形和模拟领域的加速计算性能。本文重点介绍了此版本包含的一些新功能和增强功能:
NVIDIA Blackwell 架构支持
CUDA 图形处理条件节点增强功能
用于大语言模型(LLMs)的 Blackwell CUTLASS 内核
NVIDIA Nsight 开发者工具更新
数学库更新
cudaStreamGetDevice
编译器更新
加速 Python 更新
功能齐全的架构
NVIDIA Blackwell 架构支持
CUDA 工具包 12.8 是该工具包的第一个版本,在整个开发者工具套件 (... |
https://developer.nvidia.com/blog/recent-posts/ | Recent posts | No content found | https://developer.nvidia.com/zh-cn/blog/recent-posts/ | 最近文章 | No content found |
https://developer.nvidia.com/blog/high-performance-remote-io-with-nvidia-kvikio/ | High-Performance Remote IO With NVIDIA KvikIO | Workloads processing large amounts of data, especially those running on the cloud, will often use an object storage service (S3, Google Cloud Storage, Azure Blob Storage, etc.) as the data source. Object storage services can store and serve massive amounts of data, but getting the best performance can require tailoring... | https://developer.nvidia.com/zh-cn/blog/high-performance-remote-io-with-nvidia-kvikio/ | 借助 NVIDIA KvikIO 实现高性能远程 IO | 处理大量数据的工作负载 (尤其是在云端运行的工作负载) 通常会使用对象存储服务 (S3、Google Cloud Storage、Azure Blob Storage 等) 作为数据源。对象存储服务可以存储和提供海量数据,但要想获得最佳性能,可能需要根据远程对象存储的行为方式调整工作负载。本文适用于希望尽快将数据读或写到对象存储,以便 IO 不会限制工作负载的 RAPIDS 用户。
您对本地文件系统行为方式的一些了解可转换为远程对象存储,但它们本质上是不同的。这两者之间的最大区别 (至少对于数据分析工作负载而言) 可能在于,对象存储上的读取和写入操作具有越来越高的可变延迟。每个存储服务 (AWS、Azure) 都有自己的一套最佳实践... |
https://developer.nvidia.com/blog/latest-multimodal-addition-to-microsoft-phi-slms-trained-on-nvidia-gpus/ | Latest Multimodal Addition to Microsoft Phi SLMs Trained on NVIDIA GPUs | Large language models (LLMs) have permeated every industry and changed the potential of technology. However, due to their massive size they are not practical for the current resource constraints that many companies have.
The rise of
small language models
(SLMs) bridge quality and cost by creating models with a smaller ... | https://developer.nvidia.com/zh-cn/blog/latest-multimodal-addition-to-microsoft-phi-slms-trained-on-nvidia-gpus/ | 在 NVIDIA GPU 上训练的 Microsoft Phi SLM 的多模态最新进展 | 大语言模型(LLMs)已渗透到各行各业,并改变了技术潜力。但是,由于规模庞大,它们对于许多公司目前面临的资源限制来说并不切实际。
小语言模型
(SLMs)的兴起通过创建资源占用更小的模型,将质量和成本联系起来。SLMs 是语言模型的一个子集,这些模型倾向于专注于特定领域,并使用更简单的神经架构构建。随着模型的发展模仿人类感知周围环境的方式,模型必须接受多种形式的多模态数据。
Microsoft 宣布在 Phi 系列中
推出新一代开放式 SLM
,并新增两项功能:
Phi-4-mini
Phi-4-multimodal
Phi-4-multimodal 是第一个加入该系列的多模态模型,接受文本、音频和图像数据输入。
这些模型足够小,可... |
https://developer.nvidia.com/blog/building-a-simple-vlm-based-multimodal-information-retrieval-system-with-nvidia-nim/ | Building a Simple VLM-Based Multimodal Information Retrieval System with NVIDIA NIM | In today’s data-driven world, the ability to retrieve accurate information from even modest amounts of data is vital for developers seeking streamlined, effective solutions for quick deployments, prototyping, or experimentation. One of the key challenges in information retrieval is managing the diverse modalities in un... | https://developer.nvidia.com/zh-cn/blog/building-a-simple-vlm-based-multimodal-information-retrieval-system-with-nvidia-nim/ | 使用 NVIDIA NIM 构建基于 VLM 的简单多模态信息检索系统 | 在当今数据驱动的世界中,即使是从少量数据中检索准确信息的能力,对于寻求精简、有效的快速部署、原型设计或实验解决方案的开发者来说也至关重要。信息检索领域的主要挑战之一是管理非结构化数据集中的各种模式,包括文本、PDF、图像、表格、音频、视频等。
多模态 AI 模型通过同时处理多个数据模式来应对这一挑战,以不同的形式生成连贯一致的全面输出。
NVIDIA NIM
微服务可简化 AI
基础模型
在语言、
计算机视觉
、语音、生物学等领域的安全可靠部署。
NIM 微服务可随时随地部署在 NVIDIA 加速基础设施上,并提供行业标准 API,以快速集成应用和热门 AI 开发框架 (包括 LangChain 和 LlamaIndex)。
本文将... |
https://developer.nvidia.com/blog/tag/inference-performance/ | Inference Performance | No content found | https://developer.nvidia.com/zh-cn/blog/tag/inference-performance/ | Inference Performance | No content found |
https://developer.nvidia.com/blog/optimizing-qwen2-5-coder-throughput-with-nvidia-tensorrt-llm-lookahead-decoding/ | Optimizing Qwen2.5-Coder Throughput with NVIDIA TensorRT-LLM Lookahead Decoding | Large language models (LLMs)
that specialize in coding have been steadily adopted into developer workflows. From pair programming to self-improving
AI agents
, these models assist developers with various tasks, including enhancing code, fixing bugs, generating tests, and writing documentation.
To promote the developmen... | https://developer.nvidia.com/zh-cn/blog/optimizing-qwen2-5-coder-throughput-with-nvidia-tensorrt-llm-lookahead-decoding/ | 使用 NVIDIA TensorRT-LLM 前瞻性解码优化 Qwen2.5-Coder 吞吐量 | 专注于编码的
大语言模型(LLMs)
已稳步应用于开发者工作流程。从配对编程到自我改进的
AI 智能体
,这些模型可帮助开发者完成各种任务,包括增强代码、修复错误、生成测试和编写文档。
为促进开源 LLM 的开发,Qwen 团队最近发布了 Qwen2.5-Coder,这是一系列先进的 LLM,用于跨热门编程语言的代码生成、推理和修复。本文将探讨针对
NVIDIA TensorRT-LLM 支持
的 Qwen2.5-Coder 模型进行推理优化的优势,以及借助 NVIDIA NIM 轻松部署以提升变革潜力和编码效率的好处。
Qwen2.5-Coder 模型
Qwen2.5-Coder 模型在热门的学术基准测试中取得了出色的性能。
NV... |
https://developer.nvidia.com/blog/optimize-ai-inference-performance-with-nvidia-full-stack-solutions/ | Optimize AI Inference Performance with NVIDIA Full-Stack Solutions | The explosion of AI-driven applications has placed unprecedented demands on both developers, who must balance delivering cutting-edge performance with managing operational complexity and cost, and AI infrastructure.
NVIDIA is empowering developers with full-stack innovations—spanning chips, systems, and software—that r... | https://developer.nvidia.com/zh-cn/blog/optimize-ai-inference-performance-with-nvidia-full-stack-solutions/ | 借助 NVIDIA 全栈解决方案提升 AI 推理性能 | AI 驱动的应用的爆炸式发展对开发者提出了前所未有的要求,他们必须在提供先进的性能与管理运营复杂性和成本以及 AI 基础设施之间取得平衡。
NVIDIA 正在为开发者提供涵盖芯片、系统和软件的全栈创新,重新定义
AI 推理
的可能性,使其比以往更快、更高效、更具可扩展性。
轻松部署高吞吐量、低延迟推理
六年前,NVIDIA 着手打造 AI 推理服务器,专为构建高吞吐量、延迟关键型生产应用的开发者而设计。当时,许多开发者都在努力使用定制的、特定于框架的服务器,这些服务器增加了复杂性,增加了运营成本,并且难以满足严格的服务水平协议(service-level agreements)关于延迟和吞吐量的要求。
为解决这一问题,NVIDIA ... |
https://developer.nvidia.com/blog/nvidia-tensorrt-llm-now-supports-recurrent-drafting-for-optimizing-llm-inference/ | NVIDIA TensorRT-LLM Now Supports Recurrent Drafting for Optimizing LLM Inference | Recurrent drafting
(referred as ReDrafter) is a novel speculative decoding technique developed and open-sourced by Apple for large language model (LLM) inference now available with
NVIDIA TensorRT-LLM
. ReDrafter helps developers significantly boost LLM workload performance on NVIDIA GPUs.
NVIDIA TensorRT-LLM
is a libr... | https://developer.nvidia.com/zh-cn/blog/nvidia-tensorrt-llm-now-supports-recurrent-drafting-for-optimizing-llm-inference/ | NVIDIA TensorRT-LLM 现支持 Recurrent Drafting,实现 LLM 推理优化 | Recurrent Drafting (简称 ReDrafter) 是苹果公司为大语言模型 (LLM) 推理开发并开源的一种新型推测解码技术,该技术现在可与
NVIDIA TensorRT-LLM
一起使用。ReDrafter 帮助开发者大幅提升了 NVIDIA GPU 上的 LLM 工作负载性能。
NVIDIA TensorRT-LLM
是一个 LLM 推理优化库,提供了一个易于使用的 Python API 来定义 LLM 和构建
NVIDIA TensorRT
引擎,这些引擎具有顶尖的优化功能,可在 GPU 上高效执行推理。优化功能包括自定义 Attention Kernel、Inflight Batching、Paged KV... |
https://developer.nvidia.com/blog/boost-llama-3-3-70b-inference-throughput-3x-with-nvidia-tensorrt-llm-speculative-decoding/ | Boost Llama 3.3 70B Inference Throughput 3x with NVIDIA TensorRT-LLM Speculative Decoding | Meta’s Llama collection of open
large language models (LLMs)
continues to grow with the recent addition of
Llama 3.3 70B
, a text-only instruction-tuned model. Llama 3.3 provides enhanced performance respective to the older Llama 3.1 70B model and can even match the capabilities of the larger, more computationally expe... | https://developer.nvidia.com/zh-cn/blog/boost-llama-3-3-70b-inference-throughput-3x-with-nvidia-tensorrt-llm-speculative-decoding/ | 借助 NVIDIA TensorRT-LLM 预测解码,将 Llama 3.3 的推理吞吐量提升 3 倍 | 随着近期新增的 Llama 3.3 70B
(一种纯文本指令调整模型),Meta 的开放
大语言模型 (LLMs)
集合将继续增长。Llama 3.3 相对于较旧的 Llama 3.1 70B 模型提供了增强的性能,甚至可以在数学、推理、编码和多语种支持等多项任务中匹配更大、计算成本更高的 Llama 3.1 405B 模型的功能。
NVIDIA TensorRT-LLM
是一款功能强大的推理引擎,可在最新的 LLM 上提供先进的性能,并整合了许多优化措施,可提供出色的 Llama 3.3 70B 推理吞吐量。其中包括在飞批处理、
KV 缓存
、
自定义 FP8 量化
、推测解码等,可实现快速、经济高效的 LLM 服务。
动态批处理... |
https://developer.nvidia.com/blog/spotlight-perplexity-ai-serves-400-million-search-queries-a-month-using-nvidia-inference-stack/ | Spotlight: Perplexity AI Serves 400 Million Search Queries a Month Using NVIDIA Inference Stack | The demand for AI-enabled services continues to grow rapidly, placing increasing pressure on IT and infrastructure teams. These teams are tasked with provisioning the necessary hardware and software to meet that demand while simultaneously balancing cost efficiency with optimal user experience. This challenge was faced... | https://developer.nvidia.com/zh-cn/blog/spotlight-perplexity-ai-serves-400-million-search-queries-a-month-using-nvidia-inference-stack/ | 聚焦:Perplexity AI 利用 NVIDIA 推理栈每月服务 4 亿个搜索请求 | 对 AI 赋能服务的需求持续快速增长,这给 IT 和基础设施团队带来了越来越大的压力。这些团队的任务是配置必要的硬件和软件来满足这一需求,同时平衡成本效益和最佳用户体验。Perplexity AI 的推理团队面临着这一挑战。Perplexity AI 是一款由 AI 驱动的搜索引擎,每月可处理超过 435 million 个查询。每个查询代表多个 AI 推理请求。
为了满足这一需求,Perplexity 推理团队转而使用
NVIDIA H100 Tensor Core GPUs
、
NVIDIA Triton Inference Server
和
NVIDIA TensorRT-LLM
来进行经济高效的
大语言模型(LLM)
部署。... |
https://developer.nvidia.com/blog/llama-3-2-full-stack-optimizations-unlock-high-performance-on-nvidia-gpus/ | Llama 3.2 Full-Stack Optimizations Unlock High Performance on NVIDIA GPUs | Meta recently released its
Llama 3.2
series of vision language models (VLMs), which come in 11B parameter and 90B parameter variants. These models are multimodal, supporting both text and image inputs. In addition, Meta has launched text-only
small language model (SLM)
variants of Llama 3.2 with 1B and 3B parameters. N... | https://developer.nvidia.com/zh-cn/blog/llama-3-2-full-stack-optimizations-unlock-high-performance-on-nvidia-gpus/ | Llama 3.2 全栈优化释放 NVIDIA GPU 的高性能 | Meta 最近发布了
Llama 3.2
系列视觉语言模型(VLM),其中包含 11B 参数和 90B 参数变体。这些模型是多模态模型,支持文本和图像输入。此外,Meta 还推出了 Llama 3.2 的纯文本
小语言模型(SLM)
变体,具有 1B 和 3B 参数。NVIDIA 已对 Llama 3.2 模型集合进行优化,以便在全球数百万个 GPU 上提供出色的性能和经济高效的服务 – 从我们功能强大的数据中心和云 GPU 到本地
NVIDIA RTX 工作站
,甚至是使用
NVIDIA Jetson
的低功耗边缘设备。
Llama 3.2 VLM 支持长达 128K 的文本标记,以及分辨率为 1120 x 1120 像素的单张图... |
https://developer.nvidia.com/blog/streamlining-ai-inference-performance-and-deployment-with-nvidia-tensorrt-llm-chunked-prefill/ | Streamlining AI Inference Performance and Deployment with NVIDIA TensorRT-LLM Chunked Prefill | In this blog post, we take a closer look at chunked prefill, a feature of
NVIDIA TensorRT-LLM
that increases GPU utilization and simplifies the deployment experience for developers. This builds on our previous
post
discussing how advanced KV cache optimization features in TensorRT-LLM improve performance up to 5x in us... | https://developer.nvidia.com/zh-cn/blog/streamlining-ai-inference-performance-and-deployment-with-nvidia-tensorrt-llm-chunked-prefill/ | 借助 NVIDIA TensorRT-LLM 分块预填充提高 AI 推理效率和简化部署 | 在本
博文
中,我们详细介绍了分块预填充,这是
NVIDIA TensorRT-LLM
的一项功能,可提高 GPU 利用率并简化开发者的部署体验。本文基于之前的博文,讨论了 TensorRT-LLM 中的高级 KV 缓存优化功能如何在需要系统预填充的用例中将性能提升高达 5 倍。
传统预填充和解码推理方法面临的挑战
当用户向模型提交请求时,它将经历两个不同的计算阶段:预填充和解码。每个阶段使用 GPU 计算资源的方式各不相同。
在预填充阶段,系统会处理所有输入令牌以计算 KV 缓存,然后使用该缓存生成输出的第一个令牌。此阶段的计算要求很高,可以有效利用 GPU 庞大的并行计算资源。
在解码阶段,系统会单独生成输出 tokens,使用... |
https://developer.nvidia.com/blog/5x-faster-time-to-first-token-with-nvidia-tensorrt-llm-kv-cache-early-reuse/ | 5x Faster Time to First Token with NVIDIA TensorRT-LLM KV Cache Early Reuse | In our previous
blog post
, we demonstrated how reusing the key-value (KV) cache by offloading it to CPU memory can accelerate time to first token (TTFT) by up to 14x on x86-based NVIDIA H100 Tensor Core GPUs and 28x on the NVIDIA GH200 Superchip. In this post, we shed light on KV cache reuse techniques and best practi... | https://developer.nvidia.com/zh-cn/blog/5x-faster-time-to-first-token-with-nvidia-tensorrt-llm-kv-cache-early-reuse/ | NVIDIA TensorRT-LLM KV 缓存早期重用实现首个令牌速度 5 倍提升 | 在上一篇
博客文章
中,我们展示了通过将键值 (KV) 缓存卸载到 CPU 内存来重复使用该缓存如何在基于 x86 的 NVIDIA H100 Tensor Core GPU 上将首次生成令牌 (Time To First Token, TTFT) 的速度提升高达 14 倍,在 NVIDIA GH200 Superchip 上可将其提升高达 28 倍。在本文中,我们介绍了可以进一步推动 TTFT 加速的 KV 缓存重复使用技术和最佳实践。
KV 缓存简介
许多任务(包括问答和代码生成)都迅速采用了 LLM 模型。为了生成响应,这些模型首先将用户的提示符转换为 tokens,然后将其转换为 dense vectors。随后会进行大量的... |
https://developer.nvidia.com/blog/3x-faster-allreduce-with-nvswitch-and-tensorrt-llm-multishot/ | 3x Faster AllReduce with NVSwitch and TensorRT-LLM MultiShot | Deploying generative AI workloads in production environments where user numbers can fluctuate from hundreds to hundreds of thousands – and where input sequence lengths differ with each request – poses unique challenges. To achieve low latency inference in these environments, multi-GPU setups are a must – irrespective o... | https://developer.nvidia.com/zh-cn/blog/3x-faster-allreduce-with-nvswitch-and-tensorrt-llm-multishot/ | NVSwitch 和 TensorRT-LLM MultiShot 共同加速 AllReduce 速度达 3 倍 | 在用户数量可能在数百到数十万之间波动,且输入序列长度随每个请求而变化的生产环境中,部署生成式 AI 工作负载会面临独特的挑战。要在这些环境中实现低延迟推理,无论 GPU 生成方式或显存容量如何,多 GPU 设置都是必需的。为了提高生产级设置中的推理性能,我们很高兴推出 TensorRT-LLM Multi-shot,这是一种新的多 GPU 通信协议,利用
NVIDIA NVLink Switch
可将通信速度大幅提升高达 3 倍。本博客概述了这一新功能,以及它如何帮助开发者和解决方案架构师克服传统多 GPU 通信方法的限制。
传统 AllReduce 算法面临的挑战
对于低延迟推理,无论单个 GPU 的显存容量如何,多 GPU 都至... |
https://developer.nvidia.com/blog/accelerating-scientific-literature-reviews-with-nvidia-nim-microservices-for-llms/ | Accelerating Scientific Literature Reviews with NVIDIA NIM Microservices for LLMs | A well-crafted systematic review is often the initial step for researchers exploring a scientific field. For scientists new to this field, it provides a structured overview of the domain. For experts, it refines their understanding and sparks new ideas. In 2024 alone, 218,650 review articles were indexed in the
Web of ... | https://developer.nvidia.com/zh-cn/blog/accelerating-scientific-literature-reviews-with-nvidia-nim-microservices-for-llms/ | 借助适用于 LLM 的 NVIDIA NIM 微服务加速科学文献审查 | 精心设计的系统审查通常是研究人员探索科学领域的第一步。对于刚刚接触该领域的科学家,它提供了该领域的结构化概述。对于专家而言,它可以细化他们的理解并激发新的想法。在 2024 年,
Web of Science
数据库收录了 218,650 篇评测文章,凸显了这些资源在研究中的重要性。
完成系统审查可显著增强研究人员的知识库及其学术影响。然而,传统的评论文章需要收集、阅读和总结有关特定主题的大量学术文章。由于这项人工练习非常耗时,处理过的文献的范围通常局限于几十篇或几百篇文章。跨学科内容 (通常超出研究人员的专业领域) 又增加了一层复杂性。
这些挑战使得创建全面、可靠且有影响力的系统审查变得越来越困难。
大语言模型 (LLMs) 的出... |
https://developer.nvidia.com/blog/vision-language-model-prompt-engineering-guide-for-image-and-video-understanding/ | Vision Language Model Prompt Engineering Guide for Image and Video Understanding | Vision language models
(VLMs) are evolving at a breakneck speed. In 2020, the first VLMs revolutionized the
generative AI
landscape by bringing visual understanding to
large language models
(LLMs) through the use of a
vision encoder
. These initial VLMs were limited in their abilities, only able to understand text and ... | https://developer.nvidia.com/zh-cn/blog/vision-language-model-prompt-engineering-guide-for-image-and-video-understanding/ | 图像和视频理解的视觉语言模型提示工程实践指南 | 视觉语言模型 (VLMs) 正在以极快的速度发展。2020 年,首批 VLMs 通过使用视觉编码器将视觉理解引入大语言模型 (LLMs) ,彻底改变了生成式 AI 格局。这些初始 VLMs 能力有限,只能理解文本和单张图像输入。
几年后,VLM 现在能够理解多图像和视频输入,以执行高级视觉语言任务,例如视觉问答 (VQA)、字幕、搜索和摘要。
图 1、NVIDIA NeMo 中的模型自定义工具套件
通过调整提示和模型权重,可以提高特定用例的 VLM 准确性。可以使用 PEFT 等高效技术微调模型权重,但仍需要足够的数据和计算。
但是,
prompt engineering
和上下文学习也可用于提高输出质量,并且与在运行时调整模型的文... |
https://developer.nvidia.com/blog/configurable-graph-based-task-solving-with-the-marco-multi-ai-agent-framework-for-chip-design/ | Configurable Graph-Based Task Solving with the Marco Multi-AI Agent Framework for Chip Design | Chip and hardware design presents numerous challenges stemming from its complexity and advancing technologies. These challenges result in longer turn-around time (TAT) for optimizing performance, power, area, and cost (PPAC) during synthesis, verification, physical design, and reliability loops.
Large language models (... | https://developer.nvidia.com/zh-cn/blog/configurable-graph-based-task-solving-with-the-marco-multi-ai-agent-framework-for-chip-design/ | 基于 Marco 多 AI 智能体框架的可配置图形任务求解在芯片设计中的应用 | 芯片和硬件设计因其复杂性和先进的技术而面临众多挑战。这些挑战导致在合成、验证、物理设计和可靠性循环期间优化性能、功耗、面积和成本(PPAC)的周转时间(TAT)更长。
大语言模型(LLMs)已显示出大规模理解和生成自然语言的出色能力,从而为各种领域带来许多潜在应用和优势。成功地将基于 LLMs 的 AI 智能体用于硬件设计,可以大幅减少 TAT,从而缩短产品周期、降低成本、提高设计可靠性,并降低代价高昂的错误风险。
图 1. Marco 框架概述
Marco:基于图形的可配置任务求解和多 AI 代理框架
我们介绍了拟议的 Marco 框架,其中包括基于图形的任务求解、子任务的智能体配置,以及每个 AI 智能体的实时技能/工具配置。
... |
https://developer.nvidia.com/blog/defining-llm-red-teaming/ | Defining LLM Red Teaming | There is an activity where people provide inputs to generative AI technologies, such as large language models (LLMs), to see if the outputs can be made to deviate from acceptable standards. This use of LLMs began in 2023 and has rapidly evolved to become a common industry practice and a cornerstone of
trustworthy AI
. ... | https://developer.nvidia.com/zh-cn/blog/defining-llm-red-teaming/ | 定义 LLM 红色团队 | 在一项活动中,人们为生成式 AI 技术 (例如大语言模型 (LLMs)) 提供输入,以确定输出是否会偏离可接受的标准。LLMs 的这种使用始于 2023 年,并已迅速发展成为一种常见的行业实践,也是值得信赖的 AI 的基石。如何标准化和定义 LLM 红色团队?
NVIDIA、华盛顿大学、Center for Human-Compatible AI 和哥本哈根 IT 大学的研究人员对红色团队的实际应用“
Summon a demon and bind it: A grounded theory of LLM red teaming
”(在 PLOS One 中发布) 进行了一项研究。
由于目标是定义和理解相对较新的活动,因此该研究采用... |
https://developer.nvidia.com/blog/agentic-autonomy-levels-and-security/ | Agentic Autonomy Levels and Security | Agentic workflows
are the next evolution in AI-powered tools. They enable developers to chain multiple AI models together to perform complex activities, enable AI models to use tools to access additional data or automate user actions, and enable AI models to operate autonomously, analyzing and performing complex tasks ... | https://developer.nvidia.com/zh-cn/blog/agentic-autonomy-levels-and-security/ | 代理自主级别与安全性 | 智能体工作流是 AI 赋能工具的下一次演进。它们使开发者能够将多个 AI 模型关联在一起以执行复杂的活动,使 AI 模型能够使用工具来访问其他数据或自动执行用户操作,并使 AI 模型能够自主运行,以尽可能减少人类参与或交互的方式分析和执行复杂的任务。
由于其强大的功能,代理工作流也存在风险因素。代理系统的核心最常见的模型仍然是各种 LLM,如果可以将不受信任的数据引入系统,这些模型仍然容易受到 prompt injection 的影响。
为帮助评估和缓解这些漏洞,NVIDIA 提供了一个 Agentic Autonomy 框架,我们将其用于以下用途:
了解与日益复杂的 AI 工作流相关的风险
帮助模拟和减轻此类工作流程带来的风险
介... |
https://developer.nvidia.com/blog/nvidia-deep-learning-institute-releases-new-generative-ai-teaching-kit/ | NVIDIA Deep Learning Institute Releases New Generative AI Teaching Kit | Generative AI, powered by advanced machine learning models and deep neural networks, is revolutionizing industries by generating novel content and driving innovation in fields like healthcare, finance, and entertainment.
NVIDIA is leading this transformation with its cutting-edge GPU architectures and software ecosyste... | https://developer.nvidia.com/zh-cn/blog/nvidia-deep-learning-institute-releases-new-generative-ai-teaching-kit/ | NVIDIA 深度学习培训中心发布全新生成式 AI 教学套件 | 生成式 AI 由先进的机器学习模型和深度神经网络提供动力支持,通过生成新内容并驱动医疗、金融和娱乐等领域的创新来变革各行各业。
NVIDIA 正凭借其先进的 GPU 架构和软件生态系统(例如
H100 Tensor Core GPU
和 CUDA 平台),引领这一转型,该平台可优化生成式模型的开发和部署。
NVIDIA NIM
可增强 AI 推理任务的效率和可扩展性,从而实现跨各种计算环境的快速部署和迭代,并加速生成式 AI 应用的发展。
生成式 AI 教育的重要性
随着生成式 AI 模型(例如 GANs 和 Transformers)变得越来越复杂,对能够开发、改进和合乎道德地部署这些技术的熟练专业人员的需求也在不断增长。生成式 ... |
https://developer.nvidia.com/blog/json-lines-reading-with-pandas-100x-faster-using-nvidia-cudf/ | JSON Lines Reading with pandas 100x Faster Using NVIDIA cuDF | JSON is a widely adopted format for text-based information working interoperably between systems, most commonly in web applications and large language models (LLMs). While the JSON format is human-readable, it is complex to process with data science and data engineering tools.
JSON data often takes the form of newline-... | https://developer.nvidia.com/zh-cn/blog/json-lines-reading-with-pandas-100x-faster-using-nvidia-cudf/ | 使用 NVIDIA cuDF,pandas 读取 JSON 行文件速度提升100倍 | JSON 是一种广泛采用的格式,用于在系统之间 (通常用于 Web 应用和大语言模型 (LLMs)) 以互操作方式运行的基于文本的信息。虽然 JSON 格式是人类可读的,但使用数据科学和数据工程工具进行处理十分复杂。
JSON
数据通常采用换行分隔的 JSON 行
(也称为 NDJSON) 的形式来表示数据集中的多个记录。将 JSON 行数据读入数据帧是数据处理中常见的第一步。
在本文中,我们比较了使用以下库将 JSON 行数据转换为数据帧的 Python API 的性能和功能:
pandas
DuckDB
pyarrow
RAPIDS cuDF pandas
加速器模式
我们使用 cudf.pandas 中的 JSON 读取器展示... |
https://developer.nvidia.com/blog/understanding-the-language-of-lifes-biomolecules-across-evolution-at-a-new-scale-with-evo-2/ | Understanding the Language of Life’s Biomolecules Across Evolution at a New Scale with Evo 2 | AI has evolved
from an experimental curiosity to a driving force within biological research. The convergence of deep learning algorithms, massive omics datasets, and automated laboratory workflows has allowed scientists to tackle problems once thought intractable—from rapid protein structure prediction to generative dr... | https://developer.nvidia.com/zh-cn/blog/understanding-the-language-of-lifes-biomolecules-across-evolution-at-a-new-scale-with-evo-2/ | 利用 Evo 2 以全新规模跨进化研究生命生物分子的语言 | AI 已从实验好奇心发展为生物学研究的驱动力。
深度学习算法、海量组学数据集和自动化实验室工作流程的融合使
科学家能够解决一度被认为棘手的问题
(从快速蛋白质结构预测到生成式药物设计),从而增加了科学家对 AI 素养的需求。在这一势头下,我们正处于下一次范式转变的边缘:专为生物学打造的强大 AI 基础模型的出现。
这些新模型有望将不同的数据源 (基因组序列、RNA 和蛋白质组谱,在某些情况下还包括科学文献) 统一为分子、细胞和系统级别上对生命的统一、一致的理解。学习生物学的语言和结构为变革性应用打开了大门,例如更智能的药物发现、合理的酶设计和疾病机制阐明。
在我们为下一波 AI 驱动的突破做好准备之际,这些基础模型显然不仅能够加速进... |
https://developer.nvidia.com/blog/using-networkx-jaccard-similarity-and-cugraph-to-predict-your-next-favorite-movie/ | Using NetworkX, Jaccard Similarity, and cuGraph to Predict Your Next Favorite Movie | As the amount of data available to everyone in the world increases, the ability for a consumer to make informed decisions becomes increasingly difficult. Fortunately, large datasets are a beneficial component for recommendation systems, which can make a sometimes overwhelming decision much easier.
Graphs are excellent ... | https://developer.nvidia.com/zh-cn/blog/using-networkx-jaccard-similarity-and-cugraph-to-predict-your-next-favorite-movie/ | 使用 NetworkX、Jaccard Similarity 和 cuGraph 预测您下一部最喜欢的电影 | 随着全球每个人可用的数据量不断增加,消费者做出明智决策的能力也越来越难。幸运的是,对于推荐系统而言,大型数据集是一个非常有用的组件,有时这会让某些情况下的决策变得更加容易。
对于为推荐系统提供支持的数据中固有的关系建模而言,图形是一个很好的选择,而 NetworkX 是许多数据科学家在 Python 中进行图形分析时非常喜欢的选择。NetworkX 易于学习和使用,拥有各种图形算法,并由庞大而友好的社区提供支持,并且在 Notebook、文档、Stack Overflow 和您喜欢的 LLM 中提供了大量示例。然而,令无数开发人员失望的是,他们使用 NetworkX 或甚至因为 NetworkX 而涉足图形分析,但众所周知,它在典型... |
https://developer.nvidia.com/blog/nvidia-open-gpu-datacenter-drivers-for-rhel9-signed-by-red-hat/ | NVIDIA Open GPU Datacenter Drivers for RHEL9 Signed by Red Hat | NVIDIA and Red Hat have partnered to bring continued improvements to the precompiled
NVIDIA Driver introduced in 2020
. Last month, NVIDIA
announced
that the open GPU driver modules will become the default recommended way to enable NVIDIA graphics hardware.
Today, NVIDIA announced that Red Hat is now compiling and sign... | https://developer.nvidia.com/zh-cn/blog/nvidia-open-gpu-datacenter-drivers-for-rhel9-signed-by-red-hat/ | 适用于 RHEL9 的 NVIDIA 开放 GPU 数据中心驱动,由 Red Hat 签署 | NVIDIA 和 Red Hat 合作,对
2020 年推出的预编译 NVIDIA Driver
进行持续改进。上个月,NVIDIA
宣布
,开放的 GPU 驱动模组将成为启用 NVIDIA 图形硬件的默认推荐方式。
今天,NVIDIA 宣布 Red Hat 正在编译并签署
NVIDIA 开放 GPU 内核模块
,以进一步简化 Red Hat Enterprise Linux 9 操作系统的使用,并为在数据中心使用 NVIDIA 硬件支持
机密计算
提供前进方向。
现在无需任何额外的密钥注册配置即可安装和使用经过签名的开放 GPU 驱动。与当前的旧版预编译软件包一样,在驱动更新时仅支持最新的 RHEL 9 内核。
从 RHEL 9.... |
https://developer.nvidia.com/blog/get-started-with-gpu-acceleration-for-data-science/ | Get Started with GPU Acceleration for Data Science | In data science, operational efficiency is key to handling increasingly complex and large datasets. GPU acceleration has become essential for modern workflows, offering significant performance improvements.
RAPIDS
is a suite of open-source libraries and frameworks developed by NVIDIA, designed to accelerate data scienc... | https://developer.nvidia.com/zh-cn/blog/get-started-with-gpu-acceleration-for-data-science/ | 适用于数据科学的 GPU 加速入门 | 在数据科学领域,运营效率是处理日益复杂和大型数据集的关键。GPU 加速已成为现代工作流程的关键,可显著提高性能。
RAPIDS
是由 NVIDIA 开发的一套开源库和框架,旨在使用 GPU 以尽可能减少代码更改来加速数据科学流程。RAPIDS 提供用于数据操作的
cuDF
、用于机器学习的
cuML
和用于图形分析的
cuGraph
等工具,可实现与现有 Python 库的无缝集成,使数据科学家更容易实现更快、更高效的处理。
本文分享了从 CPU 数据科学库过渡到 GPU 加速工作流程的技巧,特别适合经验丰富的数据科学家。
在桌面或云基础架构上设置 RAPIDS
开始使用 RAPIDS 非常简单,但它确实有几个依赖项。推荐的方法是遵... |
https://developer.nvidia.com/blog/how-to-use-openusd/ | How to Use OpenUSD | Universal Scene Description (OpenUSD)
is an open, extensible framework and ecosystem with APIs for composing, editing, querying, rendering, collaborating, and simulating within 3D virtual worlds.
This post explains how you can start using OpenUSD today with your existing assets and tools and what steps you can take to ... | https://developer.nvidia.com/zh-cn/blog/how-to-use-openusd/ | 如何使用 OpenUSD | 通用场景描述(Universal Scene Description)
是开放、可扩展的框架和生态系统,具有用于在 3D 虚拟世界中合成、编辑、查询、渲染、协作和仿真的 API。
本文介绍了如何使用现有资产和工具立即开始使用 OpenUSD,以及可以采取哪些步骤以迭代方式提升 USD 工作流程的水平。
要以交互方式全面介绍 OpenUSD 的基本概念、词汇和最佳实践,请探索
NVIDIA Deep Learning Institute
提供的
免费 Learn OpenUSD 学习路径
。
数据交换
OpenUSD 不会取代您现有的工具或数据格式。顾名思义,Universal Scene Description 提供了一种常用方法来... |
https://developer.nvidia.com/blog/nvidia-jetpack-6-2-brings-super-mode-to-nvidia-jetson-orin-nano-and-jetson-orin-nx-modules/ | NVIDIA JetPack 6.2 Brings Super Mode to NVIDIA Jetson Orin Nano and Jetson Orin NX Modules | The introduction of the
NVIDIA Jetson Orin Nano Super Developer Kit
sparked a new age of
generative AI
for small edge devices. The new
Super Mode
delivered an unprecedented generative AI performance boost of up to 1.7x on the developer kit, making it the most affordable generative AI supercomputer.
JetPack 6.2 is now a... | https://developer.nvidia.com/zh-cn/blog/nvidia-jetpack-6-2-brings-super-mode-to-nvidia-jetson-orin-nano-and-jetson-orin-nx-modules/ | NVIDIA JetPack 6.2 为 NVIDIA Jetson Orin Nano 和 Jetson Orin NX 模块引入超级模式 | NVIDIA Jetson Orin Nano 超级开发者套件
的推出为小型边缘设备开启了
生成式 AI
的新时代。新的
超级模式
在开发者套件上实现了前所未有的生成式 AI 性能提升,最高可达 1.7 倍,使其成为最经济实惠的生成式 AI 超级计算机。
JetPack 6.2 现已支持
Jetson Orin Nano
和
Jetson Orin NX
生产模组的超级模式,可将生成式 AI 模型的性能提升高达 2 倍。现在,您可以为新的和现有的机器人和边缘 AI 应用释放更多价值并降低总体拥有成本。
本文讨论了 Super Mode 的详细信息,包括新的功率模式、Jetson Orin Nano 和 Orin NX 模组上热门生成... |
https://developer.nvidia.com/blog/advancing-physical-ai-with-nvidia-cosmos-world-foundation-model-platform/ | Advancing Physical AI with NVIDIA Cosmos World Foundation Model Platform | As robotics and autonomous vehicles advance, accelerating development of
physical AI
—which enables
autonomous machines
to perceive, understand, and perform complex actions in the physical world—has become essential. At the center of these systems are
world foundation models (WFMs)
—AI models that simulate physical sta... | https://developer.nvidia.com/zh-cn/blog/advancing-physical-ai-with-nvidia-cosmos-world-foundation-model-platform/ | NVIDIA Cosmos World 基础模型平台助力物理 AI 进步 | 随着机器人和
自动驾驶汽车
的发展,加速
物理 AI
的发展变得至关重要,而物理 AI 使自主机器能够感知、理解并在现实世界中执行复杂的操作。这些系统的核心是
世界基础模型 (WFMs)
,即通过物理感知视频模拟物理状态的 AI 模型,使机器能够做出准确决策并与周围环境无缝交互。
NVIDIA Cosmos
平台可帮助开发者大规模为物理 AI 系统构建自定义世界模型。它为从数据管护、训练到定制的每个开发阶段提供开放世界基础模型和工具。
视频 1、NVIDIA Cosmos:适用于物理 AI 的世界基础模型平台
本文将介绍 Cosmos 及其加速物理 AI 开发的主要功能。Cosmos 是一种开源的 Python 库,用于加速物理 A... |
https://developer.nvidia.com/blog/accelerate-custom-video-foundation-model-pipelines-with-new-nvidia-nemo-framework-capabilities/ | Accelerate Custom Video Foundation Model Pipelines with New NVIDIA NeMo Framework Capabilities | Generative AI
has evolved from text-based models to multimodal models, with a recent expansion into video, opening up new potential uses across various industries. Video models can create new experiences for users or simulate scenarios for training autonomous agents at scale. They are helping revolutionize various indu... | https://developer.nvidia.com/zh-cn/blog/accelerate-custom-video-foundation-model-pipelines-with-new-nvidia-nemo-framework-capabilities/ | 通过 NVIDIA NeMo 框架的最新功能全链路加速视频基础模型的训练和微调 | 生成式 AI
已经从基于文本的模型发展到多模态模型,最近还扩展到了视频领域,为各行各业开辟了新的潜在用途。视频模型可以为用户创造新的体验,以及大规模模拟场景来训练自主代理。它们正在帮助包括机器人技术、自动驾驶汽车和娱乐业在内的多个行业实现变革。
由于视频数据的庞大和多样性,视频
基础模型
的开发面临着独特的挑战。这也凸显了可扩展的数据管理流程和训练框架的必要性,以此来理解时间和空间动态的模型。
我们发布了
NVIDIA NeMo 框架
中全新的视频基础模型功能,这是一个端到端的训练框架,用于预训练和微调您的专属视频基础模型。该框架包括高吞吐量的数据管理、高效的多模态数据加载功能、可扩展的模型训练,以及并行化的框架内推理。
高吞吐量视... |
https://developer.nvidia.com/blog/nvidia-cudss-advances-solver-technologies-for-engineering-and-scientific-computing/ | NVIDIA cuDSS Advances Solver Technologies for Engineering and Scientific Computing | NVIDIA cuDSS
is a first-generation sparse direct solver library designed to accelerate engineering and scientific computing. cuDSS is increasingly adopted in data centers and other environments and supports single-GPU, multi-GPU and multi-node (MGMN) configurations.
cuDSS has become a key tool for accelerating computer... | https://developer.nvidia.com/zh-cn/blog/nvidia-cudss-advances-solver-technologies-for-engineering-and-scientific-computing/ | NVIDIA cuDSS 提高工程和科学计算中的求解器技术 | NVIDIA cuDSS
是第一代稀疏直接求解器库,旨在加速工程和科学计算。cuDSS 正越来越多地应用于数据中心和其他环境,并支持单 GPU、多 GPU 和多节点(MGMN)配置。
cuDSS 已成为加速多个领域 (例如结构工程、流体动力学、电磁学、电路模拟、优化和 AI 辅助工程问题) 的计算机辅助工程 (CAE) 工作流程和科学计算的关键工具。
本文重点介绍了
cuDSS v0.4.0
和
cuDSS v0.5.0
中提供的一些关键性能和可用性功能 (如表 1 中总结的内容),cuDSS v0.4.0 可显著提升分解和求解步骤的性能,同时还引入了一些新功能,包括内存预测 API、自动混合内存选择和可变批量支持。cuDSS v0... |
https://developer.nvidia.com/blog/spotlight-university-of-tokyo-uses-nvidia-grace-hopper-for-groundbreaking-energy-efficient-seismic-research/ | Spotlight: University of Tokyo Uses NVIDIA Grace Hopper for Groundbreaking Energy-Efficient Seismic Research | Supercomputers are the engines of groundbreaking discoveries. From predicting extreme weather to advancing disease research and designing safer, more efficient infrastructures, these machines simulate complex systems that are impractical to test in the real world due to their size, cost, and material requirements.
Sinc... | https://developer.nvidia.com/zh-cn/blog/spotlight-university-of-tokyo-uses-nvidia-grace-hopper-for-groundbreaking-energy-efficient-seismic-research/ | 聚焦:东京大学使用 NVIDIA Grace Hopper 进行开创性的节能地震研究 | 超级计算机是推动突破性发现的引擎。从预测极端天气到推进疾病研究,以及设计更安全、更高效的基础设施,这些机器可以模拟因规模、成本和材料要求而无法在现实世界中进行测试的复杂系统。
自 1999 年推出 GPU 以来,NVIDIA 一直在不断突破加速
计算的极限
,这种方法使用专业硬件,通过在并行处理中捆绑频繁出现的任务来大幅加速工作。这种效率使超级计算机能够处理前所未有的计算挑战,同时减少每项任务消耗的能源。
如今,由 NVIDIA 提供支持的系统在高效超级计算领域处于领先地位,在 Green500 榜单前 10 名中有 8 名入选,而 Green500 榜单是高效超级计算机的行业基准测试。德国于利希超级计算中心(Jülich Supe... |
https://developer.nvidia.com/blog/spotlight-brli-and-toulouse-inp-develop-ai-based-flood-models-using-nvidia-modulus/ | Spotlight: BRLi and Toulouse INP Develop AI-Based Flood Models Using NVIDIA Modulus | Flooding poses a significant threat to 1.5 billion people, making it the most common cause of major natural disasters. Floods cause up to
$25 billion in global economic damage
every year. Flood forecasting is a critical tool in disaster preparedness and risk mitigation. Numerical methods have long been developed that p... | https://developer.nvidia.com/zh-cn/blog/spotlight-brli-and-toulouse-inp-develop-ai-based-flood-models-using-nvidia-modulus/ | 聚焦:BRLi 和 Toulouse INP 使用 NVIDIA Modulus 开发基于 AI 的洪水模型 | 洪水对 15 亿人构成了重大威胁,使其成为重大自然灾害的最常见原因。洪水每年给全球带来高达 25 亿美元的经济损失。洪水预测是防灾和减轻风险的重要工具。我们早已开发出能够准确模拟河流流域的数值方法。借助这些技术,像咨询公司
BRLi
的工程师可以研究不同的降水情景,并为当地利益相关者提供宝贵的洪水风险评估、缓解策略和灾害救济计划。准确的预测可以挽救生命、保护基础设施并减少经济损失。
然而,这些方法使用基于物理性质的数值模拟,而且计算量非常大。这种方法通常需要数小时才能模拟洪水事件,即使在许多并行处理器上也是如此。它不适合实时预测正在进行的事件。长期以来,这种限制一直是针对特定事件提供快速、可行的洪水警报的瓶颈,并阻碍了实时预报系统的... |
https://developer.nvidia.com/blog/render-path-traced-hair-in-real-time-with-nvidia-geforce-rtx-50-series-gpus/ | Render Path-Traced Hair in Real Time with NVIDIA GeForce RTX 50 Series GPUs | Hardware support for ray tracing triangle meshes was introduced as part of
NVIDIA RTX
in 2018. But ray tracing for hair and fur has remained a compute-intensive problem that has been difficult to further accelerate. That is, until now.
NVIDIA GeForce 50 Series GPUs
include a major advancement in the acceleration of ray... | https://developer.nvidia.com/zh-cn/blog/render-path-traced-hair-in-real-time-with-nvidia-geforce-rtx-50-series-gpus/ | 使用 NVIDIA GeForce RTX 50 系列 GPU 实时渲染路径追踪头发 | 2018 年,
NVIDIA RTX
引入了对光线追踪三角形网格的硬件支持。但是,头发和毛皮的光线追踪仍然是一个计算密集型问题,一直难以进一步加速。那就是,直到现在。
NVIDIA GeForce 50 系列 GPUs
在加速毛发和毛皮光线追踪方面取得了重大进步:硬件光线追踪支持线性扫描球体 (LSS) 基元。这种新基元是向实时渲染高质量数字人迈出的一大步。逼真的人类渲染的应用领域与日俱增,包括 AI 虚拟形象、电影和游戏角色、图形研究、高性能科学计算应用、合成数据生成等。
LSS 现已在
NVIDIA OptiX
和
NVAPI SDK
中提供。
RTX Character Rendering SDK
是
NVIDIA RTX K... |
https://developer.nvidia.com/blog/nvidia-video-codec-sdk-13-0-powered-by-nvidia-blackwell/ | NVIDIA Video Codec SDK 13.0 Powered by NVIDIA Blackwell | The release of
NVIDIA Video Codec SDK 13.0
marks a significant upgrade, adding support for the latest-generation NVIDIA Blackwell GPUs. This version brings a wealth of improvements aimed at elevating both video encoding and decoding capabilities. From enhanced compression efficiency to better throughput and encoding qu... | https://developer.nvidia.com/zh-cn/blog/nvidia-video-codec-sdk-13-0-powered-by-nvidia-blackwell/ | NVIDIA 视频编解码器 SDK 13.0 由 NVIDIA Blackwell 驱动 | NVIDIA Video Codec SDK 13.0
的发布标志着重大升级,增加了对新一代 NVIDIA Blackwell GPU 的支持。此版本带来了大量改进,旨在提升视频编码和解码功能。从增强的压缩效率到更好的吞吐量和编码质量,SDK 13.0 可满足视频生态系统不断变化的需求。
以下是本次更新中引入的一些主要功能。
编码功能:
提高压缩效率:
以更低的比特率实现更好的视频质量。
新的 YUV 4:2:2 编码支持 (H.264 和 HEVC):
支持更广泛的用例,特别是在专业视频制作和广播中。
更高的位深编码:
在 H.264 中引入对 10 位编码的支持,以及将 8 位内容编码为 10 位的新硬件功能。
交错编码(H.2... |
https://developer.nvidia.com/blog/enabling-stereoscopic-and-3d-views-using-mv-hevc-in-nvidia-video-codec-sdk-13-0/ | Enabling Stereoscopic and 3D Views Using MV-HEVC in NVIDIA Video Codec SDK 13.0 | NVIDIA announces the implementation of Multi-View High Efficiency Video Coding (MV-HEVC) encoder in the latest NVIDIA Video Codec SDK release, version 13.0. This significant update marks a major leap forward in hardware-accelerated, multi-view video compression. It offers enhanced compression efficiency and quality for... | https://developer.nvidia.com/zh-cn/blog/enabling-stereoscopic-and-3d-views-using-mv-hevc-in-nvidia-video-codec-sdk-13-0/ | 在 NVIDIA 视频编解码器 SDK 13.0 中使用 MV-HEVC 启用立体和 3D 视图 | NVIDIA 宣布在最新的 NVIDIA 视频编解码器 SDK 版本 13.0 中实施多视图高效视频编码 (MV-HEVC) 编码器。这一重要更新标志着硬件加速的多视图视频压缩技术实现了重大飞跃。与 simulcast 编码相比,它为立体和 3D 视频应用提供了更高的压缩效率和质量。
了解 MV-HEVC
MV-HEVC 是高效视频编码 (HEVC) 标准的扩展,旨在高效压缩从不同有利位置拍摄的同一场景的多个视频视图。它解决了传统视频编码方法 (如 simulcast 编码) 的局限性,这些方法通常会导致多视图内容的高比特率,并且缺乏高效的视图间预测。
图 1、适用于 3D 内容的 MV-HEVC
NVIDIA 视频编解码器 SDK... |
https://developer.nvidia.com/blog/new-ai-model-offers-cellular-level-view-of-cancerous-tumor/ | New AI Model Offers Cellular-Level View of Cancerous Tumors | Researchers studying cancer unveiled a new AI model that provides cellular-level mapping and visualizations of cancer cells, which scientists hope can shed light on how—and why—certain inter-cellular relationships triggers cancers to grow.
BioTuring
, a San Diego-based startup,
announced
an AI model that can quickly cr... | https://developer.nvidia.com/zh-cn/blog/new-ai-model-offers-cellular-level-view-of-cancerous-tumor/ | 新的 AI 模型提供癌症细胞级视图 | 研究癌症的研究人员推出了一种新的 AI 模型,该模型提供了癌细胞的细胞级映射和可视化,科学家希望该模型能够揭示某些细胞间关系如何以及为什么会引发癌症生长。
位于圣地亚哥的初创公司 BioTuring
宣布推出
一种 AI 模型,该模型可以以单细胞分辨率快速创建癌症肿瘤的详细可视化图像。这种类型的颗粒数据显示了细胞的大小、形状、哪些基因被激活,关键是显示了组织样本中数百万个不同细胞的相对空间位置。
与速度慢得多的传统生物计算方法不同,新模型提供了有关肿瘤动力学以及癌细胞和免疫细胞相互作用的实时高分辨率见解。
BioTuring 首席执行官 Son Pham 表示:“人体大约有 30 万亿个细胞,如果进行大规模的肿瘤活检,就会发现有几百... |
https://developer.nvidia.com/blog/get-started-with-neural-rendering-using-nvidia-rtx-kit/ | Get Started with Neural Rendering Using NVIDIA RTX Kit | Neural rendering is the next era of computer graphics. By integrating neural networks into the rendering process, we can take dramatic leaps forward in performance, image quality, and interactivity to deliver new levels of immersion.
NVIDIA RTX Kit
is a suite of neural rendering technologies to ray-trace games with AI... | https://developer.nvidia.com/zh-cn/blog/get-started-with-neural-rendering-using-nvidia-rtx-kit/ | 开始使用 NVIDIA RTX 套件进行神经渲染 | 神经渲染是计算机图形的下一个时代。通过将神经网络集成到渲染过程中,我们可以在性能、图像质量和交互性方面实现巨大飞跃,从而将沉浸感提升到新的高度。
NVIDIA RTX Kit
是一套神经渲染技术,可借助 AI 进行光线追踪游戏、渲染具有宏大几何图形的场景,以及使用逼真视觉效果创建游戏角色。
RTX Kit 现已在
/NVIDIA RTX/RTXkit
GitHub 存储库中提供,并整合了几个新的和熟悉的 RTX 和 AI 组件。尽管可以单独使用每个组件,但 RTX Kit 提供了一个统一的位置来访问这些基本渲染 SDK。
姓名
说明
状态
RTX 神经着色器
(新增)
在着色器中训练和部署神经网络,为新一代资产生成解锁新的压缩和近似... |
https://developer.nvidia.com/blog/nvidia-rtx-mega-geometry-now-available-with-new-vulkan-samples/ | NVIDIA RTX Mega Geometry Now Available with New Vulkan Samples | Geometric detail in computer graphics has increased exponentially in the past 30 years. To render high quality assets with higher instance counts and greater triangle density, NVIDIA introduced RTX Mega Geometry. RTX Mega Geometry is available today through
NVIDIA RTX Kit
, a suite of rendering technologies to ray trac... | https://developer.nvidia.com/zh-cn/blog/nvidia-rtx-mega-geometry-now-available-with-new-vulkan-samples/ | NVIDIA RTX Mega Geometry 现已支持新的 Vulkan 示例程序 | 在过去 30 年里,计算机图形领域的几何细节呈指数级增长。为了渲染具有更高实例数量和三角形密度的高质量素材,NVIDIA 推出了 RTX Mega Geometry。RTX Mega Geometry 现已通过
NVIDIA RTX Kit
推出,这是一套渲染技术,可利用 AI 对游戏进行光线追踪、渲染具有宏大几何图形的场景,以及创建具有逼真视觉效果的游戏角色。
作为此版本的一部分,新的 Vulkan 示例已提供给所有开发者。这些开源示例展示了如何使用三角形集群更快地构建加速结构、渲染大量动画几何图形、具有流式传输细节级别(LoD)的路径追踪模型等。NVIDIA 还将发布两个库,帮助将几何图形处理到集群中。有关 RTX Mega G... |
https://developer.nvidia.com/blog/build-apps-with-neural-rendering-using-nvidia-nsight-developer-tools-on-geforce-rtx-50-series-gpus/ | Build Apps with Neural Rendering Using NVIDIA Nsight Developer Tools on GeForce RTX 50 Series GPUs | The next generation of NVIDIA graphics hardware has arrived. Powered by NVIDIA Blackwell,
GeForce RTX 50 Series GPUs
deliver groundbreaking new RTX features such as
DLSS 4 with Multi Frame Generation
, and
NVIDIA RTX Kit
with RTX Mega Geometry and RTX Neural Shaders. NVIDIA RTX Blackwell architecture introduces fifth-g... | https://developer.nvidia.com/zh-cn/blog/build-apps-with-neural-rendering-using-nvidia-nsight-developer-tools-on-geforce-rtx-50-series-gpus/ | 使用 NVIDIA Nsight 开发者工具和 GeForce RTX 50系列 GPU 构建神经渲染应用 | 新一代 NVIDIA 图形硬件已经面世。由 NVIDIA Blackwell 提供支持的
GeForce RTX 50 系列 GPU
可提供突破性的全新 RTX 功能,例如支持多帧生成的 DLSS 4,以及支持 RTX Mega Geometry 和 RTX Neural Shaders 的
NVIDIA RTX Kit
。NVIDIA RTX Blackwell 架构采用第五代 Tensor Cores 来驱动 AI 工作负载和第四代 RT Cores,三角形交叉速率是上一代的两倍。它配备超快的 GDDR7 内存,可大幅提升总内存带宽。
NVIDIA Nsight 开发者工具
支持图形开发者使用 GeForce RTX 50 系列... |
https://developer.nvidia.com/blog/how-to-integrate-nvidia-dlss-4-into-your-game-with-nvidia-streamline/ | How to Integrate NVIDIA DLSS 4 into Your Game with NVIDIA Streamline | NVIDIA DLSS 4 is the latest iteration of DLSS introduced with the
NVIDIA GeForce RTX 50 Series GPUs
. It includes several new features:
DLSS Multi Frame Generation:
Multi Frame Generation has been introduced for GeForce RTX 50 Series GPUs, allowing for generation of up to 3 frames per rendered frame.
DLSS Ray Reconstru... | https://developer.nvidia.com/zh-cn/blog/how-to-integrate-nvidia-dlss-4-into-your-game-with-nvidia-streamline/ | 如何借助 NVIDIA Streamline 将 NVIDIA DLSS 4 集成到您的游戏中 | NVIDIA DLSS 4 是
NVIDIA GeForce RTX 50 系列 GPU
中引入的 DLSS 的最新版本。它包含多项新功能:
DLSS 多帧生成:
已为 GeForce RTX 50 系列 GPUs 引入多帧生成,允许每渲染帧生成多达 3 帧。
DLSS Ray Reconstruction & DLSS Super Resolution (Beta):
:已更新,可使用新的 Transformer 模型,提高时间稳定性、动态细节和照明细节。
DLSS 帧生成
:帧生成模型经过更新,可在 GeForce RTX 50 和 RTX 40 系列 GPUs 上运行更快,并使用更少的 VRAM。
DLSS 4|
下面介绍如何... |
https://developer.nvidia.com/blog/nvidia-rtx-neural-rendering-introduces-next-era-of-ai-powered-graphics-innovation/ | NVIDIA RTX Neural Rendering Introduces Next Era of AI-Powered Graphics Innovation | NVIDIA today unveiled next-generation hardware for gamers, creators, and developers—the GeForce RTX 50 Series
desktop
and
laptop
GPUs. Alongside these GPUs, NVIDIA introduced
NVIDIA RTX Kit
, a suite of neural rendering technologies to ray trace games with AI, render scenes with immense geometry, and create game charac... | https://developer.nvidia.com/zh-cn/blog/nvidia-rtx-neural-rendering-introduces-next-era-of-ai-powered-graphics-innovation/ | NVIDIA RTX 神经渲染引领 AI 图形创新新时代 | NVIDIA 今天发布了面向游戏玩家、创作者和开发者的新一代硬件 – GeForce RTX 50 系列台式机和笔记本电脑 GPU。除这些 GPU 外,NVIDIA 还推出了
NVIDIA RTX Kit
,这是一套神经渲染技术,可使用 AI 对游戏进行光线追踪、渲染具有庞大几何图形的场景,以及创建具有逼真视觉效果的游戏角色。RTX Kit 可增强几何图形、纹理、材质和照明,为复杂的开放世界提供更高的视觉保真度和性能,同时减少伪影、稳定性和 VRAM 使用率。开发者可以将单个或所有 RTX Kit SDK 直接集成到其渲染管线中。
RTX 神经着色器将 AI 引入可编程着色器
25 年前,NVIDIA 推出了 GeForce 和可编... |
https://developer.nvidia.com/blog/accelerating-film-production-with-dell-ai-factory-and-nvidia/ | Accelerating Film Production with Dell AI Factory and NVIDIA | Filmmaking is an intricate and complex process that involves a diverse team of artists, writers, visual effects professionals, technicians, and countless other specialists. Each member brings their unique expertise to the table, collaborating to transform a simple idea into a captivating cinematic experience. From the ... | https://developer.nvidia.com/zh-cn/blog/accelerating-film-production-with-dell-ai-factory-and-nvidia/ | 借助 Dell AI Factory 和 NVIDIA 加速电影制作 | 电影制作是一个复杂而复杂的过程,涉及艺术家、作家、视觉效果专业人员、技术人员和无数其他专家组成的多元化团队。每位成员都将各自独特的专业知识带到桌面上,通过协作将简单的想法转化为引人入胜的电影体验。从故事的初始火花到最终剪辑,每一步都需要创造力、技能和对细节的细致关注。
然而,即使拥有如此优秀的团队,电影业也一直在努力应对众多挑战,这些挑战可能会阻碍生产力、增加预算和抑制创造力。这些挑战包括管理不断上涨的生产成本和适应迅速变化的技术,如 PyTorch、pandas 和 LangChain 等,以及驾不断变化的受众偏好和分销平台。同时努力在竞争日益激烈的市场中平衡艺术视觉与商业可行性,例如如何利用 NVIDIA 的 GPU 和 cuO... |
https://developer.nvidia.com/blog/deploying-nvidia-riva-multilingual-asr-with-whisper-and-canary-architectures-while-selectively-deactivating-nmt/ | Deploying NVIDIA Riva Multilingual ASR with Whisper and Canary Architectures While Selectively Deactivating NMT | NVIDIA has consistently developed automatic speech recognition (ASR) models that set the benchmark in the industry.
Earlier versions of
NVIDIA Riva
, a collection of GPU-accelerated speech and translation AI microservices for
ASR
,
TTS
, and
NMT
, support English-Spanish and English-Japanese code-switching ASR models b... | https://developer.nvidia.com/zh-cn/blog/deploying-nvidia-riva-multilingual-asr-with-whisper-and-canary-architectures-while-selectively-deactivating-nmt/ | 使用 Whisper 和 Canary 架构部署 NVIDIA Riva 多语种 ASR,同时选择性地停用 NMT | NVIDIA 持续开发自动语音识别 (ASR) 模型,这些模型在业内树立了基准。
早期版本的
NVIDIA Riva
是一种适用于
ASR
、
TTS
和
NMT
的 GPU 加速语音和翻译 AI 微服务的集合,支持基于
Conformer
架构的英语-西班牙语和英语-日语代码交换
ASR
模型,以及基于
Parakeet
架构的支持 EMEA 地区多种语言(即英国英语、欧洲西班牙语、法语、意大利语、标准德语和亚美尼亚语)的模型。
最近,NVIDIA 发布了
Riva 2.18.0 容器和 SDK
,以不断改进其语音 AI 模型。在此新版本中,我们现在提供以下服务:
支持 Parakeet,即多语言流式传输 ASR
支持 OpenA... |
https://developer.nvidia.com/blog/improving-translation-quality-with-domain-specific-fine-tuning-and-nvidia-nim/ | Improving Translation Quality with Domain-Specific Fine-Tuning and NVIDIA NIM | Translation plays an essential role in enabling companies to expand across borders, with requirements varying significantly in terms of tone, accuracy, and technical terminology handling. The emergence of
sovereign AI
has highlighted critical challenges in
large language models (LLMs)
, particularly their struggle to c... | https://developer.nvidia.com/zh-cn/blog/improving-translation-quality-with-domain-specific-fine-tuning-and-nvidia-nim/ | 利用特定领域的微调和 NVIDIA NIM 提高翻译质量 | 翻译在助力公司实现跨国业务扩展方面发挥着至关重要的作用,在语调、准确性和技术术语处理方面的要求差异很大。
主权 AI
的出现凸显了
大语言模型(LLM)
面临的严峻挑战,尤其是他们难以捕捉英语主导框架之外的细微文化和语言背景。随着全球通信变得日益复杂,组织必须仔细评估翻译解决方案,以平衡技术效率与文化敏感性和语言准确性。
在本文中,我们将探讨 LLMs 如何解决以下两种截然不同的英语到繁体中文翻译用例:
网站营销内容:精确翻译技术文本,同时保持自然的宣传基调。
在
线培训课程
:翻译 Jupyter Notebooks 等平台中使用的幻灯片文本和 markdown 内容,确保准确的技术翻译和正确的 markdown 格式 (如标题、... |
https://developer.nvidia.com/blog/announcing-nemotron-cc-a-trillion-token-english-language-dataset-for-llm-pretraining/ | Announcing Nemotron-CC: A Trillion-Token English Language Dataset for LLM Pretraining | NVIDIA is excited to announce the release of Nemotron-CC, a 6.3-trillion-token English language
Common Crawl
dataset for pretraining highly accurate
large language models (LLMs)
, including 1.9 trillion tokens of synthetically generated data. One of the keys to training state-of-the-art LLMs is a high-quality pretraini... | https://developer.nvidia.com/zh-cn/blog/announcing-nemotron-cc-a-trillion-token-english-language-dataset-for-llm-pretraining/ | 宣布推出 Nemotron-CC:用于 LLM 预训练的万亿级英语语言数据集 | NVIDIA 很高兴地宣布发布 Nemotron-CC,这是一种包含 6.3 万亿个令牌的 英语语言
Common Crawl
数据集 ,用于预训练高度准确的
大语言模型 (LLMs)
,其中包括 1.9 万亿个令牌的合成生成数据。高质量预训练数据集是训练先进 LLM 的关键之一,近期的顶级 LLM (例如
Meta Llama 系列
) 基于包含 15 万亿个令牌的大量数据进行了训练。
但我们对这 15 万亿个令牌的确切构成知之甚少。Nemotron-CC 旨在解决这一问题,并使更广泛的社区能够训练高度准确的 LLM。互联网爬行数据 (通常来自 Common Crawl) 通常是最大的令牌来源。近期开放的 Common Crawl... |
https://developer.nvidia.com/blog/build-a-generative-ai-medical-device-training-assistant-with-nvidia-nim-microservices/ | Build a Generative AI Medical Device Training Assistant with NVIDIA NIM Microservices | Innovation in medical devices continues to accelerate, with a record number authorized by the
FDA every year
. When these new or updated devices are introduced to clinicians and patients, they require training to use them properly and safely.
Once in use, clinicians or patients may need help troubleshooting issues. Med... | https://developer.nvidia.com/zh-cn/blog/build-a-generative-ai-medical-device-training-assistant-with-nvidia-nim-microservices/ | 利用 NVIDIA NIM 微服务构建生成式 AI 医疗设备训练助手 | 医疗设备创新不断加速,每年获得 FDA 批准的数量都创下新高。向临床医生和患者介绍这些全新或更新的设备时,他们需要接受培训,以正确、安全地使用这些设备。
使用后,临床医生或患者可能需要获得排查问题的帮助。医疗设备通常附有冗长且技术复杂的 Instructions for Use (IFU) 手册,这些手册描述了设备的正确用法。快速找到正确的信息可能很困难,在新设备上训练是一项耗时的任务。医疗设备代表通常会提供支持培训,但可能无法实时回答所有问题。这些问题可能会延迟医疗设备的使用和较新技术的采用,在某些情况下,还会导致错误使用。
使用生成式 AI 对医疗设备进行故障排除
检索增强型生成 (RAG) 使用深度学习模型 (包括大语言模型 ... |
https://developer.nvidia.com/blog/sandboxing-agentic-ai-workflows-with-webassembly/ | Sandboxing Agentic AI Workflows with WebAssembly | Agentic AI workflows
often involve the execution of
large language model (LLM)
-generated code to perform tasks like creating data visualizations. However, this code should be sanitized and executed in a safe environment to mitigate risks from
prompt injection
and errors in the returned code. Sanitizing Python with reg... | https://developer.nvidia.com/zh-cn/blog/sandboxing-agentic-ai-workflows-with-webassembly/ | 借助 WebAssembly 实现沙箱 Agentic AI 工作流 | 代理 AI 工作流通常
涉及执行由
大语言模型 (LLM)
生成的代码,以执行创建数据可视化等任务。但是,此代码应在安全环境中清理和执行,以降低提示
注入的风险
和返回代码中的错误。使用正则表达式和受限运行时清理 Python 是不够的,而且虚拟机的 Hypervisor 隔离需要大量的开发和资源。
本文将介绍如何使用
WebAssembly (Wasm) (一种基于堆栈的虚拟机的二进制指令格式)
,利用浏览器沙盒实现操作系统和用户隔离。这提高了应用的安全性,且不会产生重大开销。
确保代理工具的使用
LLM 应用开发的近期变化之一是公开工具,即 LLM 可以调用并使用响应的函数、应用或 API。例如,如果应用需要了解特定地点的天气,... |
https://developer.nvidia.com/blog/three-building-blocks-for-creating-ai-virtual-assistants-for-customer-service-with-an-nvidia-nim-agent-blueprint/ | Three Building Blocks for Creating AI Virtual Assistants for Customer Service with an NVIDIA AI Blueprint | In today’s fast-paced business environment, providing exceptional customer service is no longer just a nice-to-have—it’s a necessity. Whether addressing technical issues, resolving billing questions, or providing service updates, customers expect quick, accurate, and personalized responses at their convenience. However... | https://developer.nvidia.com/zh-cn/blog/three-building-blocks-for-creating-ai-virtual-assistants-for-customer-service-with-an-nvidia-nim-agent-blueprint/ | 借助 NVIDIA NIM 智能体蓝图创建客户服务 AI 虚拟助理的三大核心组件 | 在当今快节奏的商业环境中,提供卓越的客户服务已不再是一种好事情,而是一种必要条件。无论是解决技术问题、解决计费问题,还是提供服务更新,客户都期望在方便时得到快速、准确和个性化的响应。然而,实现这种服务水平面临重大挑战。
传统方法,例如静态脚本或手动流程,通常在提供个性化和实时支持方面存在不足。此外,许多客户服务运营依赖于敏感和零碎的数据,这些数据受严格的数据治理和隐私法规的约束。随着生成式 AI 的兴起,公司旨在通过提高运营效率、降低成本和最大限度地提高投资回报率(ROI)来革新客户服务。
将 AI 集成到现有系统会带来透明度、准确性和安全性方面的挑战,这些挑战可能会阻碍采用并中断工作流程。为此,公司正在利用由生成式 AI 提供支持... |
https://developer.nvidia.com/blog/hymba-hybrid-head-architecture-boosts-small-language-model-performance/ | Hymba Hybrid-Head Architecture Boosts Small Language Model Performance | Transformers, with their attention-based architecture, have become the dominant choice for language models (LMs) due to their strong performance, parallelization capabilities, and long-term recall through key-value (KV) caches. However, their quadratic computational cost and high memory demands pose efficiency challeng... | https://developer.nvidia.com/zh-cn/blog/hymba-hybrid-head-architecture-boosts-small-language-model-performance/ | Hymba 混合头架构提高小型语言模型性能 | Transformer 及其基于注意力的架构,凭借强大的性能、并行化功能以及通过键值 (KV) 缓存进行的长期召回,已成为语言模型 (LM) 的主要选择。然而,其二次计算成本和高内存需求带来了效率挑战。相比之下,Mamba 和 Mamba-2 等状态空间模型 (SSM) 可提供恒定的复杂性和高效的硬件优化,但难以处理记忆回收任务,从而影响其在常规基准测试中的性能。
NVIDIA 研究人员最近提出了
Hymba
,这是一系列小语言模型 (SLMs),采用混合 head 并行架构,将 Transformer Attention 机制与 SSMs 集成,以提高效率和性能。在 Hymba 中,attention heads 可实现高分辨率召... |
https://developer.nvidia.com/blog/advancing-robot-learning-perception-and-manipulation-with-latest-nvidia-isaac-release/ | Advancing Robot Learning, Perception, and Manipulation with Latest NVIDIA Isaac Release | At CES 2025, NVIDIA announced key updates to
NVIDIA Isaac
, a platform of accelerated libraries, application frameworks, and AI models that accelerate the development of AI robots.
NVIDIA Isaac streamlines the development of robotic systems from simulation to real-world deployment. In this post, we discuss all the new ... | https://developer.nvidia.com/zh-cn/blog/advancing-robot-learning-perception-and-manipulation-with-latest-nvidia-isaac-release/ | 利用最新的 NVIDIA Isaac 版本推进机器人学习、感知和操控技术 | 在 CES 2025 上,NVIDIA 宣布了对
NVIDIA Isaac
的重要更新。NVIDIA Isaac 是一个由加速库、应用框架和 AI 模型组成的平台,可加速 AI 机器人的开发。
NVIDIA Isaac 简化了从仿真到实际部署的机器人系统开发过程。在本文中,我们将讨论 NVIDIA Isaac 的所有新进展:
Isaac Sim
Isaac Lab
Isaac Manipulator
Isaac Perceptor
Isaac Sim 4.5 的新增功能
NVIDIA Isaac Sim
是一款基于
NVIDIA Omniverse
构建的参考应用,使您能够在基于物理的虚拟环境中开发、模拟和测试 AI 驱动的机器人。... |
https://developer.nvidia.com/blog/ai-vision-helps-green-recycling-plants/ | AI Vision Helps Green Recycling Plants | Each year, the world recycles only around 13% of its two billion-plus tons of municipal waste. By 2050, the world’s annual municipal waste will reach 3.88B tons.
But the global recycling industry is far from efficient. Annually, as much as $120B of potentially recoverable plastic—let alone paper or metals—ends up in la... | https://developer.nvidia.com/zh-cn/blog/ai-vision-helps-green-recycling-plants/ | AI 视觉技术助力绿色回收工厂智能化 | 全球每年仅回收约 13% 的 20 多亿吨城市垃圾。到 2050 年,全球每年的城市废弃物量将达到 3.88 亿吨。
但全球回收行业的效率远远不够。每年,价值高达 120 亿美元的可回收塑料 (更不用说纸张或金属了) 最终进入垃圾填埋场,而不是使用回收材料制造的新产品。
英国的初创公司
Greyparrot
开发了一款由 AI 驱动的小型设备,可提供“废弃物智能”功能,旨在帮助回收工厂变得更高效、更环保。
这款名为 Greyparrot Analyzer 的两英尺方形时尚设备使用嵌入式摄像头来识别和区分流经回收工厂的传送带上的材料。
该分析器使用机器学习物体检测模型(ODM)连接到回收工厂,并帮助改造其基础设施。ODM 根据数千万张... |
https://developer.nvidia.com/blog/five-takeaways-from-nvidia-6g-developer-day-2024/ | Five Takeaways from NVIDIA 6G Developer Day 2024 | NVIDIA 6G Developer Day 2024
brought together members of the 6G research and development community to share insights and learn new ways of engaging with NVIDIA 6G research tools. More than 1,300 academic and industry researchers from across the world attended the virtual event. It featured presentations from NVIDIA, ET... | https://developer.nvidia.com/zh-cn/blog/five-takeaways-from-nvidia-6g-developer-day-2024/ | 2024 年 NVIDIA 6G 开发者日的 5 大关键收获 | 2024 年
NVIDIA 6G 开发者日让 6G 研发社区成员共聚一堂
,分享见解,并学习使用 NVIDIA 6G 研究工具的新方式。来自世界各地的 1,300 多名学术和行业研究人员参加了此次虚拟活动。会上,NVIDIA、ETH Zürich、Keysight、Northeastern University、Samsung、Softbank 和 University of Oulu 发表了演讲。本文将探讨本次活动的五个要点。
1、6G 将采用 AI 原生技术并实施 AI-RAN
预计 6G 将驾 AI 浪潮,为消费者和企业释放新的潜力,并改变电信基础设施。这是 NVIDIA
高级副总裁 Ronnie Vasishta 发表的主题... |
https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/ | NVIDIA Jetson Orin Nano Developer Kit Gets a “Super” Boost | The
generative AI
landscape is rapidly evolving, with new
large language models
(LLMs), visual language models (VLMs), and vision language action (VLA) models emerging daily. To stay at the forefront of this transformative era, developers need a platform powerful enough to seamlessly deploy the latest models from the c... | https://developer.nvidia.com/zh-cn/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/ | NVIDIA Jetson Orin Nano 开发者套件得到“超级”提升 | 生成式 AI
格局正在迅速发展,新的
大语言模型 (LLMs)
、视觉语言模型 (VLMs) 和视觉语言动作 (VLAs) 模型每天都在涌现。为了在这个变革时代站在前沿,开发者需要一个足够强大的平台,通过使用 CUDA 的优化推理和开放 ML 框架,将最新模型从云端无缝部署到边缘。
为了支持机器人和多模态智能体中生成式 AI 工作负载的新兴应用,NVIDIA 正在通过软件更新来更新
NVIDIA Jetson Orin Nano 开发者套件
,以 249 美元的惊人新价格将性能提升高达 1.7 倍。为表彰 Jetson Orin Nano 开发者套件在性能和可访问性方面的出色提升,我们将 Jetson Orin Nano 开发者套件... |
https://developer.nvidia.com/blog/nvidia-jetpack-6-1-boosts-performance-and-security-through-camera-stack-optimizations-and-introduction-of-firmware-tpm/ | NVIDIA JetPack 6.1 Boosts Performance and Security through Camera Stack Optimizations and Introduction of Firmware TPM | NVIDIA JetPack
has continuously evolved to offer cutting-edge software tailored to the growing needs of
edge AI and robotic developers
. With each release, JetPack has enhanced its performance, introduced new features, and optimized existing tools to deliver increased value to its users. This means that your existing
J... | https://developer.nvidia.com/zh-cn/blog/nvidia-jetpack-6-1-boosts-performance-and-security-through-camera-stack-optimizations-and-introduction-of-firmware-tpm/ | NVIDIA JetPack 6.1 通过摄像头堆栈优化和固件 TPM 实现性能和安全性提升 | NVIDIA JetPack
不断演进,为满足
边缘 AI 和机器人开发者
的不断增长需求提供最新的软件。每个版本中,JetPack 都会增强性能,引入新功能,并优化现有工具,为用户提供更大的价值。这意味着,您现有的基于
Jetson Orin
的产品可以通过升级到最新版本的 JetPack 体验到性能优化。
JetPack 6 于 2023 年第三季度发布,标志着这一进程中的一个重要里程碑。它引入了强大的功能,如
自带内核
(Bring Your Own Kernel),提供了灵活性,可以引入任何 LTS Linux 内核,并扩展了对更广泛的 Linux 发行版的支持,为不同的开发环境提供了更大的灵活性。
它采用模块化设计,无需升... |
https://developer.nvidia.com/blog/nvidia-doca-2-9-enhances-ai-and-cloud-computing-infrastructure-with-new-performance-and-security-features/ | NVIDIA DOCA 2.9 Enhances AI and Cloud Computing Infrastructure with New Performance and Security Features | NVIDIA DOCA enhances the capabilities of NVIDIA networking platforms by providing a comprehensive software framework for developers to leverage hardware acceleration, boosting performance, security, and efficiency. Its ecosystem of APIs, libraries, and tools streamlines development for data center infrastructure, enabl... | https://developer.nvidia.com/zh-cn/blog/nvidia-doca-2-9-enhances-ai-and-cloud-computing-infrastructure-with-new-performance-and-security-features/ | NVIDIA DOCA 2.9 借助新性能和安全功能强化人工智能和云计算基础设施 | NVIDIA DOCA 通过为开发者提供全面的软件框架来利用硬件加速来增强 NVIDIA 网络平台的功能,从而提高性能、安全性和效率。它的 API、库和工具生态系统可简化数据中心基础设施的开发,实现工作负载卸载、加速和隔离,以支持现代、高效的数据中心。
如今,许多主要的 CSP 和 NVIDIA 云合作伙伴(NCP)都在使用 DOCA,并为快速开发和部署创新解决方案提供了标准化平台。这种方法可缩短上市时间、降低成本,并使合作伙伴能够在使用 NVIDIA 硬件加速的同时专注于自身的核心优势。
DOCA 培育的开放生态系统促进了协作和互操作性,创造了各种互补解决方案。合作伙伴可以探索 AI 驱动的网络、高级安全性和高性能存储等新兴市场,... |
https://developer.nvidia.com/blog/ai-powered-devices-track-howls-to-save-wolves/ | AI-Powered Devices Track Howls to Save Wolves | A new cell-phone-sized device—which can be deployed in vast, remote areas—is using AI to identify and geolocate wildlife to help conservationists track endangered species, including wolves around Yellowstone National Park.
The battery-powered devices—dubbed GrizCams—are designed by a small Montana startup, Grizzly Syst... | https://developer.nvidia.com/zh-cn/blog/ai-powered-devices-track-howls-to-save-wolves/ | AI 驱动的设备追踪嚎叫以拯救狼群 | 一种可部署在广大偏远地区的手机大小的新型设备正在使用人工智能识别和地理定位野生动物,以帮助保护主义者追踪濒危物种,包括黄石国家公园周围的狼群。
这款名为 GrizCams 的电池供电设备由蒙大拿州的一家小型初创公司 Grizzly Systems 设计。他们与生物学家合作,在整个大黄石生态系统中部署一系列设备,以录制有关狼或狼群叫的时间和地点的音频和视频。
一旦完全部署,这些数据可以帮助科学家和环保人士更好地了解狼的行为,并制定新的策略来阻止狼攻击牲畜。
环保人士每隔几个月就会在远程录音机上从 SD 卡中检索音频数据。这些数据由使用数 TB 的狼嚎叫数据训练的 AI 模型输入和分析。该模型是一种卷积神经网络,可将音频转换为频谱图,然... |
https://developer.nvidia.com/blog/ai-accurately-forecasts-extreme-weather-up-to-23-days-ahead/ | AI Accurately Forecasts Extreme Weather Up to 23 Days Ahead | New research from the University of Washington is refining AI weather models using deep learning for more accurate predictions and longer-term forecasts. The study, published in
Geophysical Research Letters
, shows how adjusting initial atmospheric data enables advanced AI models to extend current forecast limits. As e... | https://developer.nvidia.com/zh-cn/blog/ai-accurately-forecasts-extreme-weather-up-to-23-days-ahead/ | AI 提前 23 天实现极端天气高精度预测 | 华盛顿大学的新研究正在使用深度学习完善 AI 天气模型,以实现更准确的预测和更长期的预测。这项研究发表在
《地球物理研究通讯》
上,展示了调整初始大气数据如何使先进的 AI 模型能够扩展当前的预测极限。随着极端天气因气候变化而变得越来越严重和频繁,让政府、企业、公众和应急响应人员有更多的时间为洪水、热浪或飓风等自然灾害做好准备,有助于减少生命和财产的损失。
华盛顿大学博士生 Trent Vonich 说:“如果为完美的天气模型提供略微不完善的初始条件,误差会随着时间的推移而加剧,并导致预测不准确。”“在对地球大气等混沌系统进行建模时尤其如此。最近人们非常关注制作更好的模型,同时在某种程度上忽略了一个事实,即完美的模型只是问题的一半。... |
https://developer.nvidia.com/blog/llm-model-pruning-and-knowledge-distillation-with-nvidia-nemo-framework/ | LLM Model Pruning and Knowledge Distillation with NVIDIA NeMo Framework | Model pruning and knowledge distillation are powerful cost-effective strategies for obtaining smaller language models from an initial larger sibling.
Pruning:
Either drop layers (depth-pruning) or drop neurons, attention heads, and embedding channels (width-pruning).
Knowledge distillation:
Transfer knowledge from a la... | https://developer.nvidia.com/zh-cn/blog/llm-model-pruning-and-knowledge-distillation-with-nvidia-nemo-framework/ | 使用 NVIDIA NeMo 框架进行 LLM 模型剪枝和知识蒸馏 | 模型剪枝和知识蒸馏是功能强大且经济高效的策略,用于从最初较大的同级获得较小的语言模型。
剪枝:丢弃图层 (深度剪枝) 或丢弃神经元、注意力头和嵌入通道 (宽度剪枝)。
知识蒸馏:
将知识从大型教师模型转移到较小的学生模型,目标是创建更高效、更小、运行速度更快且资源密集型更低的模型。
在一篇“
如何剪枝和蒸馏 Llama-3.1 8B
”博文中,讨论了使用
大语言模型(LLM)
的最佳实践,该模型将深度、宽度、注意力和 MLP 剪枝与基于蒸馏的知识重新训练相结合。
在本文中,我们提供了一个关于 NVIDIA NeMo 框架中基于简单数据集的剪枝和蒸馏工作流的演练教程。本教程使用
Meta-Llama-3.1-8B
作为教师模型,目标模... |
https://developer.nvidia.com/blog/nvidia-dgx-cloud-introduces-ready-to-use-templates-to-benchmark-ai-platform-performance/ | NVIDIA DGX Cloud Introduces Ready-To-Use Templates to Benchmark AI Platform Performance | In the rapidly evolving landscape of AI systems and workloads, achieving optimal model training performance extends far beyond chip speed. It requires a comprehensive evaluation of the entire stack, from compute to networking to model framework.
Navigating the complexities of AI system performance can be difficult. The... | https://developer.nvidia.com/zh-cn/blog/nvidia-dgx-cloud-introduces-ready-to-use-templates-to-benchmark-ai-platform-performance/ | NVIDIA DGX 云推出即用型模板,更好地评估 AI 平台性能 | 在快速发展的 AI 系统和工作负载环境中,实现出色的模型训练性能远远超出芯片速度。这需要对整个堆栈进行全面评估,从计算到网络,再到模型框架。
了解复杂的 AI 系统性能可能比较困难。从精度到并行,您可以对许多应用程序进行更改,但目前需要付出大量努力和掌握专业知识才能有效实施这些更改。
NVIDIA DGX Cloud
Benchmarking Recipes 为提高训练性能提供了指导,通过分享每个工作负载的良好性能以及实现方法的最佳实践。
如需了解更多信息,请参阅
NVIDIA DGX Cloud Benchmarking Recipes
。
全面评估 AI 系统
DGX 云 Benchmarking Recipes 是一款端到端基... |
https://developer.nvidia.com/blog/nvidia-grace-cpu-integrates-with-the-arm-software-ecosystem/ | NVIDIA Grace CPU Integrates with the Arm Software Ecosystem | The NVIDIA Grace CPU is transforming data center design by offering a new level of power-efficient performance. Built specifically for data center scale, the Grace CPU is designed to handle demanding workloads while consuming less power.
NVIDIA believes in the benefit of leveraging GPUs to accelerate every workload. Ho... | https://developer.nvidia.com/zh-cn/blog/nvidia-grace-cpu-integrates-with-the-arm-software-ecosystem/ | NVIDIA Grace CPU 与 Arm 软件生态系统集成 | NVIDIA Grace CPU 将节能高效的性能提升到了新的高度,正在改变数据中心的设计。Grace CPU 专为数据中心规模打造,旨在处理要求严苛的工作负载,同时降低功耗。
NVIDIA 相信利用 GPU 加速各种工作负载的好处。但是,并非所有工作负载都得到加速。对于涉及复杂的分支代码 (例如图形分析) 的工作负载尤其如此,这些代码通常用于欺诈检测、运营优化和社交网络分析等热门用例。
随着数据中心面临越来越大的功率限制,加速尽可能多的工作负载并在尽可能高效的计算上运行其余工作负载变得至关重要。Grace CPU 经过优化,可处理加速任务和仅使用 CPU 的任务,在功耗与传统 CPU 相同的情况下提供高达 2 倍的性能。
Grac... |
https://developer.nvidia.com/blog/openai-triton-on-nvidia-blackwell-boosts-ai-performance-and-programmability/ | OpenAI Triton on NVIDIA Blackwell Boosts AI Performance and Programmability | Matrix multiplication and attention mechanisms are the computational backbone of modern AI workloads. While libraries like
NVIDIA cuDNN
provide highly optimized implementations, and frameworks such as
CUTLASS
offer deep customization, many developers and researchers need a middle ground that combines performance with p... | https://developer.nvidia.com/zh-cn/blog/openai-triton-on-nvidia-blackwell-boosts-ai-performance-and-programmability/ | NVIDIA Blackwell 上的 OpenAI Triton 提升 AI 性能和可编程性 | 矩阵乘法和注意力机制是现代 AI 工作负载的计算支柱。虽然库如
NVIDIA cuDNN
提供高度优化的实现,而框架如
CUTLASS
提供深度自定义,但许多开发者和研究人员都需要将性能与可编程性结合起来的中间点。
NVIDIA Blackwell 架构
上的
开源 Triton 编译器
通过直观的编程模型展示 Blackwell 的高级功能来满足这一需求。
得益于 NVIDIA 与 OpenAI 的持续合作,Triton 编译器现已支持 NVIDIA Blackwell 架构。这可确保开发者和研究人员可以通过基于 Python 的编译器 (例如 Triton) 轻松使用 Blackwell 架构中最新的出色功能。
NVIDIA B... |
https://developer.nvidia.com/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | AI for Climate, Energy, and Ecosystem Resilience at NVIDIA GTC 2025 | From mitigating climate change to improving disaster response and environmental monitoring, AI is reshaping how we tackle critical global challenges. Advancements in fast, high-resolution climate forecasting, real-time monitoring, and digital twins are equipping scientists, policy-makers, and industry leaders with data... | https://developer.nvidia.com/zh-cn/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | NVIDIA GTC 2025 上的人工智能促进气候、能源和生态系统复原力 | 从减缓气候变化到改进灾害响应和环境监测,AI 正在重塑我们应对重大全球挑战的方式。快速、高分辨率的气候预报、实时监控和数字孪生技术的进步为科学家、政策制定者和行业领导者提供了数据驱动的工具,帮助他们了解、规划和应对一个变暖的星球。
在 3 月 17 日至 21 日举行的
NVIDIA GTC 2025
大会上,思想领袖、科学家、开发者和创新者将重点介绍 AI 如何帮助塑造更具可持续性和韧性的未来。以下会议展示了 AI 在气候预测、灾难缓解和保护工作中发挥的作用,帮助社区适应日益不可预测的世界。
加强极端天气预测和灾害响应
随着全球变暖加剧,极端天气事件变得更加严重和频繁,社区需要更快、更精确的自然灾害预测和响应策略。AI 正在改进洪... |
https://developer.nvidia.com/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | Automating GPU Kernel Generation with DeepSeek-R1 and Inference Time Scaling | As AI models extend their capabilities to solve more sophisticated challenges, a new
scaling law
known as
test-time scaling
or
inference-time scaling
is emerging. Also known as
AI reasoning
or
long-thinking
, this technique improves model performance by allocating additional computational resources during inference to ... | https://developer.nvidia.com/zh-cn/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | 使用 DeepSeek-R1 和推理时间缩放实现自动化 GPU 内核生成 | 随着 AI 模型扩展其功能以解决更复杂的挑战,一种称为“
测试时扩展
”或“
推理时扩展
”的新扩展法则正在出现。该技术也称为
AI 推理
或
长时思考
技术,通过在推理过程中分配额外的计算资源来评估多种可能的结果,然后选择最佳结果 (neural network),从而提高模型性能。这使得 AI 能够以类似于人类剖析复杂问题并单独解决这些问题以达成最终解决方案的方式,制定战略并系统化地解决复杂问题。
在本文中,我们将介绍 NVIDIA 工程师完成的一项实验,他们在推理过程中使用最新的开源模型之一 DeepSeek-R1 模型以及额外的计算能力来解决复杂的问题。该实验旨在自动生成 GPU 注意力内核,这些内核在数值上是正确的,并针对... |
https://developer.nvidia.com/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI Foundation Model Enhances Cancer Diagnosis and Tailors Treatment | A new study and AI model from researchers at Stanford University is streamlining cancer diagnostics, treatment planning, and prognosis prediction. Named MUSK (Multimodal transformer with Unified maSKed modeling), the research
aims to advance precision oncology, tailoring treatment plans to each patient based on their u... | https://developer.nvidia.com/zh-cn/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI 基础模型增强癌症诊断并实现个性化治疗 | 斯坦福大学研究人员的一项新研究和 AI 模型正在简化癌症诊断、治疗规划和预后预测。这项名为 MUSK (Multimodal transformer with Unified maSKed modeling) 的研究旨在提高精准肿瘤学,根据每位患者独特的医疗数据为其定制治疗计划。
“多模态基础模型是医学 AI 研究的新领域,”放射肿瘤学副教授兼研究高级作者
Ruijiang LI
说。“最近,我们为医学领域开发了视觉语言基础模型,尤其是在病理学领域。但是,现有研究使用的现有基础模型需要配对的图像 – 文本数据进行预训练。尽管我们付出了大量努力,最终打造出 1M 病理图像文本对,但它仍然不足以完全捕捉整个疾病谱系的多样性。”
在考虑患... |
https://developer.nvidia.com/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA Toolkit Now Available for NVIDIA Blackwell | The latest release of the
CUDA Toolkit
, version 12.8, continues to push accelerated computing performance in data sciences, AI, scientific computing, and computer graphics and simulation, using the latest NVIDIA CPUs and GPUs. This post highlights some of the new features and enhancements included with this release:
N... | https://developer.nvidia.com/zh-cn/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA 工具包现已支持 NVIDIA Blackwell 架构 | CUDA 工具包
的最新版本 (版本 12.8) 使用最新的 NVIDIA CPU 和 GPU,持续提升数据科学、AI、科学计算以及计算机图形和模拟领域的加速计算性能。本文重点介绍了此版本包含的一些新功能和增强功能:
NVIDIA Blackwell 架构支持
CUDA 图形处理条件节点增强功能
用于大语言模型(LLMs)的 Blackwell CUTLASS 内核
NVIDIA Nsight 开发者工具更新
数学库更新
cudaStreamGetDevice
编译器更新
加速 Python 更新
功能齐全的架构
NVIDIA Blackwell 架构支持
CUDA 工具包 12.8 是该工具包的第一个版本,在整个开发者工具套件 (... |
https://developer.nvidia.com/blog/recent-posts/ | Recent posts | No content found | https://developer.nvidia.com/zh-cn/blog/recent-posts/ | 最近文章 | No content found |
https://developer.nvidia.com/blog/high-performance-remote-io-with-nvidia-kvikio/ | High-Performance Remote IO With NVIDIA KvikIO | Workloads processing large amounts of data, especially those running on the cloud, will often use an object storage service (S3, Google Cloud Storage, Azure Blob Storage, etc.) as the data source. Object storage services can store and serve massive amounts of data, but getting the best performance can require tailoring... | https://developer.nvidia.com/zh-cn/blog/high-performance-remote-io-with-nvidia-kvikio/ | 借助 NVIDIA KvikIO 实现高性能远程 IO | 处理大量数据的工作负载 (尤其是在云端运行的工作负载) 通常会使用对象存储服务 (S3、Google Cloud Storage、Azure Blob Storage 等) 作为数据源。对象存储服务可以存储和提供海量数据,但要想获得最佳性能,可能需要根据远程对象存储的行为方式调整工作负载。本文适用于希望尽快将数据读或写到对象存储,以便 IO 不会限制工作负载的 RAPIDS 用户。
您对本地文件系统行为方式的一些了解可转换为远程对象存储,但它们本质上是不同的。这两者之间的最大区别 (至少对于数据分析工作负载而言) 可能在于,对象存储上的读取和写入操作具有越来越高的可变延迟。每个存储服务 (AWS、Azure) 都有自己的一套最佳实践... |
https://developer.nvidia.com/blog/latest-multimodal-addition-to-microsoft-phi-slms-trained-on-nvidia-gpus/ | Latest Multimodal Addition to Microsoft Phi SLMs Trained on NVIDIA GPUs | Large language models (LLMs) have permeated every industry and changed the potential of technology. However, due to their massive size they are not practical for the current resource constraints that many companies have.
The rise of
small language models
(SLMs) bridge quality and cost by creating models with a smaller ... | https://developer.nvidia.com/zh-cn/blog/latest-multimodal-addition-to-microsoft-phi-slms-trained-on-nvidia-gpus/ | 在 NVIDIA GPU 上训练的 Microsoft Phi SLM 的多模态最新进展 | 大语言模型(LLMs)已渗透到各行各业,并改变了技术潜力。但是,由于规模庞大,它们对于许多公司目前面临的资源限制来说并不切实际。
小语言模型
(SLMs)的兴起通过创建资源占用更小的模型,将质量和成本联系起来。SLMs 是语言模型的一个子集,这些模型倾向于专注于特定领域,并使用更简单的神经架构构建。随着模型的发展模仿人类感知周围环境的方式,模型必须接受多种形式的多模态数据。
Microsoft 宣布在 Phi 系列中
推出新一代开放式 SLM
,并新增两项功能:
Phi-4-mini
Phi-4-multimodal
Phi-4-multimodal 是第一个加入该系列的多模态模型,接受文本、音频和图像数据输入。
这些模型足够小,可... |
https://developer.nvidia.com/blog/building-a-simple-vlm-based-multimodal-information-retrieval-system-with-nvidia-nim/ | Building a Simple VLM-Based Multimodal Information Retrieval System with NVIDIA NIM | In today’s data-driven world, the ability to retrieve accurate information from even modest amounts of data is vital for developers seeking streamlined, effective solutions for quick deployments, prototyping, or experimentation. One of the key challenges in information retrieval is managing the diverse modalities in un... | https://developer.nvidia.com/zh-cn/blog/building-a-simple-vlm-based-multimodal-information-retrieval-system-with-nvidia-nim/ | 使用 NVIDIA NIM 构建基于 VLM 的简单多模态信息检索系统 | 在当今数据驱动的世界中,即使是从少量数据中检索准确信息的能力,对于寻求精简、有效的快速部署、原型设计或实验解决方案的开发者来说也至关重要。信息检索领域的主要挑战之一是管理非结构化数据集中的各种模式,包括文本、PDF、图像、表格、音频、视频等。
多模态 AI 模型通过同时处理多个数据模式来应对这一挑战,以不同的形式生成连贯一致的全面输出。
NVIDIA NIM
微服务可简化 AI
基础模型
在语言、
计算机视觉
、语音、生物学等领域的安全可靠部署。
NIM 微服务可随时随地部署在 NVIDIA 加速基础设施上,并提供行业标准 API,以快速集成应用和热门 AI 开发框架 (包括 LangChain 和 LlamaIndex)。
本文将... |
https://developer.nvidia.com/blog/tag/inference-performance/ | Inference Performance | No content found | https://developer.nvidia.com/zh-cn/blog/tag/inference-performance/ | Inference Performance | No content found |
https://developer.nvidia.com/blog/optimizing-qwen2-5-coder-throughput-with-nvidia-tensorrt-llm-lookahead-decoding/ | Optimizing Qwen2.5-Coder Throughput with NVIDIA TensorRT-LLM Lookahead Decoding | Large language models (LLMs)
that specialize in coding have been steadily adopted into developer workflows. From pair programming to self-improving
AI agents
, these models assist developers with various tasks, including enhancing code, fixing bugs, generating tests, and writing documentation.
To promote the developmen... | https://developer.nvidia.com/zh-cn/blog/optimizing-qwen2-5-coder-throughput-with-nvidia-tensorrt-llm-lookahead-decoding/ | 使用 NVIDIA TensorRT-LLM 前瞻性解码优化 Qwen2.5-Coder 吞吐量 | 专注于编码的
大语言模型(LLMs)
已稳步应用于开发者工作流程。从配对编程到自我改进的
AI 智能体
,这些模型可帮助开发者完成各种任务,包括增强代码、修复错误、生成测试和编写文档。
为促进开源 LLM 的开发,Qwen 团队最近发布了 Qwen2.5-Coder,这是一系列先进的 LLM,用于跨热门编程语言的代码生成、推理和修复。本文将探讨针对
NVIDIA TensorRT-LLM 支持
的 Qwen2.5-Coder 模型进行推理优化的优势,以及借助 NVIDIA NIM 轻松部署以提升变革潜力和编码效率的好处。
Qwen2.5-Coder 模型
Qwen2.5-Coder 模型在热门的学术基准测试中取得了出色的性能。
NV... |
https://developer.nvidia.com/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | AI for Climate, Energy, and Ecosystem Resilience at NVIDIA GTC 2025 | From mitigating climate change to improving disaster response and environmental monitoring, AI is reshaping how we tackle critical global challenges. Advancements in fast, high-resolution climate forecasting, real-time monitoring, and digital twins are equipping scientists, policy-makers, and industry leaders with data... | https://developer.nvidia.com/zh-cn/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | NVIDIA GTC 2025 上的人工智能促进气候、能源和生态系统复原力 | 从减缓气候变化到改进灾害响应和环境监测,AI 正在重塑我们应对重大全球挑战的方式。快速、高分辨率的气候预报、实时监控和数字孪生技术的进步为科学家、政策制定者和行业领导者提供了数据驱动的工具,帮助他们了解、规划和应对一个变暖的星球。
在 3 月 17 日至 21 日举行的
NVIDIA GTC 2025
大会上,思想领袖、科学家、开发者和创新者将重点介绍 AI 如何帮助塑造更具可持续性和韧性的未来。以下会议展示了 AI 在气候预测、灾难缓解和保护工作中发挥的作用,帮助社区适应日益不可预测的世界。
加强极端天气预测和灾害响应
随着全球变暖加剧,极端天气事件变得更加严重和频繁,社区需要更快、更精确的自然灾害预测和响应策略。AI 正在改进洪... |
https://developer.nvidia.com/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | Automating GPU Kernel Generation with DeepSeek-R1 and Inference Time Scaling | As AI models extend their capabilities to solve more sophisticated challenges, a new
scaling law
known as
test-time scaling
or
inference-time scaling
is emerging. Also known as
AI reasoning
or
long-thinking
, this technique improves model performance by allocating additional computational resources during inference to ... | https://developer.nvidia.com/zh-cn/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | 使用 DeepSeek-R1 和推理时间缩放实现自动化 GPU 内核生成 | 随着 AI 模型扩展其功能以解决更复杂的挑战,一种称为“
测试时扩展
”或“
推理时扩展
”的新扩展法则正在出现。该技术也称为
AI 推理
或
长时思考
技术,通过在推理过程中分配额外的计算资源来评估多种可能的结果,然后选择最佳结果 (neural network),从而提高模型性能。这使得 AI 能够以类似于人类剖析复杂问题并单独解决这些问题以达成最终解决方案的方式,制定战略并系统化地解决复杂问题。
在本文中,我们将介绍 NVIDIA 工程师完成的一项实验,他们在推理过程中使用最新的开源模型之一 DeepSeek-R1 模型以及额外的计算能力来解决复杂的问题。该实验旨在自动生成 GPU 注意力内核,这些内核在数值上是正确的,并针对... |
https://developer.nvidia.com/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI Foundation Model Enhances Cancer Diagnosis and Tailors Treatment | A new study and AI model from researchers at Stanford University is streamlining cancer diagnostics, treatment planning, and prognosis prediction. Named MUSK (Multimodal transformer with Unified maSKed modeling), the research
aims to advance precision oncology, tailoring treatment plans to each patient based on their u... | https://developer.nvidia.com/zh-cn/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI 基础模型增强癌症诊断并实现个性化治疗 | 斯坦福大学研究人员的一项新研究和 AI 模型正在简化癌症诊断、治疗规划和预后预测。这项名为 MUSK (Multimodal transformer with Unified maSKed modeling) 的研究旨在提高精准肿瘤学,根据每位患者独特的医疗数据为其定制治疗计划。
“多模态基础模型是医学 AI 研究的新领域,”放射肿瘤学副教授兼研究高级作者
Ruijiang LI
说。“最近,我们为医学领域开发了视觉语言基础模型,尤其是在病理学领域。但是,现有研究使用的现有基础模型需要配对的图像 – 文本数据进行预训练。尽管我们付出了大量努力,最终打造出 1M 病理图像文本对,但它仍然不足以完全捕捉整个疾病谱系的多样性。”
在考虑患... |
https://developer.nvidia.com/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA Toolkit Now Available for NVIDIA Blackwell | The latest release of the
CUDA Toolkit
, version 12.8, continues to push accelerated computing performance in data sciences, AI, scientific computing, and computer graphics and simulation, using the latest NVIDIA CPUs and GPUs. This post highlights some of the new features and enhancements included with this release:
N... | https://developer.nvidia.com/zh-cn/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA 工具包现已支持 NVIDIA Blackwell 架构 | CUDA 工具包
的最新版本 (版本 12.8) 使用最新的 NVIDIA CPU 和 GPU,持续提升数据科学、AI、科学计算以及计算机图形和模拟领域的加速计算性能。本文重点介绍了此版本包含的一些新功能和增强功能:
NVIDIA Blackwell 架构支持
CUDA 图形处理条件节点增强功能
用于大语言模型(LLMs)的 Blackwell CUTLASS 内核
NVIDIA Nsight 开发者工具更新
数学库更新
cudaStreamGetDevice
编译器更新
加速 Python 更新
功能齐全的架构
NVIDIA Blackwell 架构支持
CUDA 工具包 12.8 是该工具包的第一个版本,在整个开发者工具套件 (... |
https://developer.nvidia.com/blog/recent-posts/ | Recent posts | No content found | https://developer.nvidia.com/zh-cn/blog/recent-posts/ | 最近文章 | No content found |
https://developer.nvidia.com/blog/high-performance-remote-io-with-nvidia-kvikio/ | High-Performance Remote IO With NVIDIA KvikIO | Workloads processing large amounts of data, especially those running on the cloud, will often use an object storage service (S3, Google Cloud Storage, Azure Blob Storage, etc.) as the data source. Object storage services can store and serve massive amounts of data, but getting the best performance can require tailoring... | https://developer.nvidia.com/zh-cn/blog/high-performance-remote-io-with-nvidia-kvikio/ | 借助 NVIDIA KvikIO 实现高性能远程 IO | 处理大量数据的工作负载 (尤其是在云端运行的工作负载) 通常会使用对象存储服务 (S3、Google Cloud Storage、Azure Blob Storage 等) 作为数据源。对象存储服务可以存储和提供海量数据,但要想获得最佳性能,可能需要根据远程对象存储的行为方式调整工作负载。本文适用于希望尽快将数据读或写到对象存储,以便 IO 不会限制工作负载的 RAPIDS 用户。
您对本地文件系统行为方式的一些了解可转换为远程对象存储,但它们本质上是不同的。这两者之间的最大区别 (至少对于数据分析工作负载而言) 可能在于,对象存储上的读取和写入操作具有越来越高的可变延迟。每个存储服务 (AWS、Azure) 都有自己的一套最佳实践... |
https://developer.nvidia.com/blog/latest-multimodal-addition-to-microsoft-phi-slms-trained-on-nvidia-gpus/ | Latest Multimodal Addition to Microsoft Phi SLMs Trained on NVIDIA GPUs | Large language models (LLMs) have permeated every industry and changed the potential of technology. However, due to their massive size they are not practical for the current resource constraints that many companies have.
The rise of
small language models
(SLMs) bridge quality and cost by creating models with a smaller ... | https://developer.nvidia.com/zh-cn/blog/latest-multimodal-addition-to-microsoft-phi-slms-trained-on-nvidia-gpus/ | 在 NVIDIA GPU 上训练的 Microsoft Phi SLM 的多模态最新进展 | 大语言模型(LLMs)已渗透到各行各业,并改变了技术潜力。但是,由于规模庞大,它们对于许多公司目前面临的资源限制来说并不切实际。
小语言模型
(SLMs)的兴起通过创建资源占用更小的模型,将质量和成本联系起来。SLMs 是语言模型的一个子集,这些模型倾向于专注于特定领域,并使用更简单的神经架构构建。随着模型的发展模仿人类感知周围环境的方式,模型必须接受多种形式的多模态数据。
Microsoft 宣布在 Phi 系列中
推出新一代开放式 SLM
,并新增两项功能:
Phi-4-mini
Phi-4-multimodal
Phi-4-multimodal 是第一个加入该系列的多模态模型,接受文本、音频和图像数据输入。
这些模型足够小,可... |
https://developer.nvidia.com/blog/building-a-simple-vlm-based-multimodal-information-retrieval-system-with-nvidia-nim/ | Building a Simple VLM-Based Multimodal Information Retrieval System with NVIDIA NIM | In today’s data-driven world, the ability to retrieve accurate information from even modest amounts of data is vital for developers seeking streamlined, effective solutions for quick deployments, prototyping, or experimentation. One of the key challenges in information retrieval is managing the diverse modalities in un... | https://developer.nvidia.com/zh-cn/blog/building-a-simple-vlm-based-multimodal-information-retrieval-system-with-nvidia-nim/ | 使用 NVIDIA NIM 构建基于 VLM 的简单多模态信息检索系统 | 在当今数据驱动的世界中,即使是从少量数据中检索准确信息的能力,对于寻求精简、有效的快速部署、原型设计或实验解决方案的开发者来说也至关重要。信息检索领域的主要挑战之一是管理非结构化数据集中的各种模式,包括文本、PDF、图像、表格、音频、视频等。
多模态 AI 模型通过同时处理多个数据模式来应对这一挑战,以不同的形式生成连贯一致的全面输出。
NVIDIA NIM
微服务可简化 AI
基础模型
在语言、
计算机视觉
、语音、生物学等领域的安全可靠部署。
NIM 微服务可随时随地部署在 NVIDIA 加速基础设施上,并提供行业标准 API,以快速集成应用和热门 AI 开发框架 (包括 LangChain 和 LlamaIndex)。
本文将... |
https://developer.nvidia.com/blog/tag/inference-performance/ | Inference Performance | No content found | https://developer.nvidia.com/zh-cn/blog/tag/inference-performance/ | Inference Performance | No content found |
https://developer.nvidia.com/blog/optimizing-qwen2-5-coder-throughput-with-nvidia-tensorrt-llm-lookahead-decoding/ | Optimizing Qwen2.5-Coder Throughput with NVIDIA TensorRT-LLM Lookahead Decoding | Large language models (LLMs)
that specialize in coding have been steadily adopted into developer workflows. From pair programming to self-improving
AI agents
, these models assist developers with various tasks, including enhancing code, fixing bugs, generating tests, and writing documentation.
To promote the developmen... | https://developer.nvidia.com/zh-cn/blog/optimizing-qwen2-5-coder-throughput-with-nvidia-tensorrt-llm-lookahead-decoding/ | 使用 NVIDIA TensorRT-LLM 前瞻性解码优化 Qwen2.5-Coder 吞吐量 | 专注于编码的
大语言模型(LLMs)
已稳步应用于开发者工作流程。从配对编程到自我改进的
AI 智能体
,这些模型可帮助开发者完成各种任务,包括增强代码、修复错误、生成测试和编写文档。
为促进开源 LLM 的开发,Qwen 团队最近发布了 Qwen2.5-Coder,这是一系列先进的 LLM,用于跨热门编程语言的代码生成、推理和修复。本文将探讨针对
NVIDIA TensorRT-LLM 支持
的 Qwen2.5-Coder 模型进行推理优化的优势,以及借助 NVIDIA NIM 轻松部署以提升变革潜力和编码效率的好处。
Qwen2.5-Coder 模型
Qwen2.5-Coder 模型在热门的学术基准测试中取得了出色的性能。
NV... |
https://developer.nvidia.com/blog/optimize-ai-inference-performance-with-nvidia-full-stack-solutions/ | Optimize AI Inference Performance with NVIDIA Full-Stack Solutions | The explosion of AI-driven applications has placed unprecedented demands on both developers, who must balance delivering cutting-edge performance with managing operational complexity and cost, and AI infrastructure.
NVIDIA is empowering developers with full-stack innovations—spanning chips, systems, and software—that r... | https://developer.nvidia.com/zh-cn/blog/optimize-ai-inference-performance-with-nvidia-full-stack-solutions/ | 借助 NVIDIA 全栈解决方案提升 AI 推理性能 | AI 驱动的应用的爆炸式发展对开发者提出了前所未有的要求,他们必须在提供先进的性能与管理运营复杂性和成本以及 AI 基础设施之间取得平衡。
NVIDIA 正在为开发者提供涵盖芯片、系统和软件的全栈创新,重新定义
AI 推理
的可能性,使其比以往更快、更高效、更具可扩展性。
轻松部署高吞吐量、低延迟推理
六年前,NVIDIA 着手打造 AI 推理服务器,专为构建高吞吐量、延迟关键型生产应用的开发者而设计。当时,许多开发者都在努力使用定制的、特定于框架的服务器,这些服务器增加了复杂性,增加了运营成本,并且难以满足严格的服务水平协议(service-level agreements)关于延迟和吞吐量的要求。
为解决这一问题,NVIDIA ... |
https://developer.nvidia.com/blog/nvidia-tensorrt-llm-now-supports-recurrent-drafting-for-optimizing-llm-inference/ | NVIDIA TensorRT-LLM Now Supports Recurrent Drafting for Optimizing LLM Inference | Recurrent drafting
(referred as ReDrafter) is a novel speculative decoding technique developed and open-sourced by Apple for large language model (LLM) inference now available with
NVIDIA TensorRT-LLM
. ReDrafter helps developers significantly boost LLM workload performance on NVIDIA GPUs.
NVIDIA TensorRT-LLM
is a libr... | https://developer.nvidia.com/zh-cn/blog/nvidia-tensorrt-llm-now-supports-recurrent-drafting-for-optimizing-llm-inference/ | NVIDIA TensorRT-LLM 现支持 Recurrent Drafting,实现 LLM 推理优化 | Recurrent Drafting (简称 ReDrafter) 是苹果公司为大语言模型 (LLM) 推理开发并开源的一种新型推测解码技术,该技术现在可与
NVIDIA TensorRT-LLM
一起使用。ReDrafter 帮助开发者大幅提升了 NVIDIA GPU 上的 LLM 工作负载性能。
NVIDIA TensorRT-LLM
是一个 LLM 推理优化库,提供了一个易于使用的 Python API 来定义 LLM 和构建
NVIDIA TensorRT
引擎,这些引擎具有顶尖的优化功能,可在 GPU 上高效执行推理。优化功能包括自定义 Attention Kernel、Inflight Batching、Paged KV... |
https://developer.nvidia.com/blog/boost-llama-3-3-70b-inference-throughput-3x-with-nvidia-tensorrt-llm-speculative-decoding/ | Boost Llama 3.3 70B Inference Throughput 3x with NVIDIA TensorRT-LLM Speculative Decoding | Meta’s Llama collection of open
large language models (LLMs)
continues to grow with the recent addition of
Llama 3.3 70B
, a text-only instruction-tuned model. Llama 3.3 provides enhanced performance respective to the older Llama 3.1 70B model and can even match the capabilities of the larger, more computationally expe... | https://developer.nvidia.com/zh-cn/blog/boost-llama-3-3-70b-inference-throughput-3x-with-nvidia-tensorrt-llm-speculative-decoding/ | 借助 NVIDIA TensorRT-LLM 预测解码,将 Llama 3.3 的推理吞吐量提升 3 倍 | 随着近期新增的 Llama 3.3 70B
(一种纯文本指令调整模型),Meta 的开放
大语言模型 (LLMs)
集合将继续增长。Llama 3.3 相对于较旧的 Llama 3.1 70B 模型提供了增强的性能,甚至可以在数学、推理、编码和多语种支持等多项任务中匹配更大、计算成本更高的 Llama 3.1 405B 模型的功能。
NVIDIA TensorRT-LLM
是一款功能强大的推理引擎,可在最新的 LLM 上提供先进的性能,并整合了许多优化措施,可提供出色的 Llama 3.3 70B 推理吞吐量。其中包括在飞批处理、
KV 缓存
、
自定义 FP8 量化
、推测解码等,可实现快速、经济高效的 LLM 服务。
动态批处理... |
https://developer.nvidia.com/blog/spotlight-perplexity-ai-serves-400-million-search-queries-a-month-using-nvidia-inference-stack/ | Spotlight: Perplexity AI Serves 400 Million Search Queries a Month Using NVIDIA Inference Stack | The demand for AI-enabled services continues to grow rapidly, placing increasing pressure on IT and infrastructure teams. These teams are tasked with provisioning the necessary hardware and software to meet that demand while simultaneously balancing cost efficiency with optimal user experience. This challenge was faced... | https://developer.nvidia.com/zh-cn/blog/spotlight-perplexity-ai-serves-400-million-search-queries-a-month-using-nvidia-inference-stack/ | 聚焦:Perplexity AI 利用 NVIDIA 推理栈每月服务 4 亿个搜索请求 | 对 AI 赋能服务的需求持续快速增长,这给 IT 和基础设施团队带来了越来越大的压力。这些团队的任务是配置必要的硬件和软件来满足这一需求,同时平衡成本效益和最佳用户体验。Perplexity AI 的推理团队面临着这一挑战。Perplexity AI 是一款由 AI 驱动的搜索引擎,每月可处理超过 435 million 个查询。每个查询代表多个 AI 推理请求。
为了满足这一需求,Perplexity 推理团队转而使用
NVIDIA H100 Tensor Core GPUs
、
NVIDIA Triton Inference Server
和
NVIDIA TensorRT-LLM
来进行经济高效的
大语言模型(LLM)
部署。... |
https://developer.nvidia.com/blog/llama-3-2-full-stack-optimizations-unlock-high-performance-on-nvidia-gpus/ | Llama 3.2 Full-Stack Optimizations Unlock High Performance on NVIDIA GPUs | Meta recently released its
Llama 3.2
series of vision language models (VLMs), which come in 11B parameter and 90B parameter variants. These models are multimodal, supporting both text and image inputs. In addition, Meta has launched text-only
small language model (SLM)
variants of Llama 3.2 with 1B and 3B parameters. N... | https://developer.nvidia.com/zh-cn/blog/llama-3-2-full-stack-optimizations-unlock-high-performance-on-nvidia-gpus/ | Llama 3.2 全栈优化释放 NVIDIA GPU 的高性能 | Meta 最近发布了
Llama 3.2
系列视觉语言模型(VLM),其中包含 11B 参数和 90B 参数变体。这些模型是多模态模型,支持文本和图像输入。此外,Meta 还推出了 Llama 3.2 的纯文本
小语言模型(SLM)
变体,具有 1B 和 3B 参数。NVIDIA 已对 Llama 3.2 模型集合进行优化,以便在全球数百万个 GPU 上提供出色的性能和经济高效的服务 – 从我们功能强大的数据中心和云 GPU 到本地
NVIDIA RTX 工作站
,甚至是使用
NVIDIA Jetson
的低功耗边缘设备。
Llama 3.2 VLM 支持长达 128K 的文本标记,以及分辨率为 1120 x 1120 像素的单张图... |
https://developer.nvidia.com/blog/streamlining-ai-inference-performance-and-deployment-with-nvidia-tensorrt-llm-chunked-prefill/ | Streamlining AI Inference Performance and Deployment with NVIDIA TensorRT-LLM Chunked Prefill | In this blog post, we take a closer look at chunked prefill, a feature of
NVIDIA TensorRT-LLM
that increases GPU utilization and simplifies the deployment experience for developers. This builds on our previous
post
discussing how advanced KV cache optimization features in TensorRT-LLM improve performance up to 5x in us... | https://developer.nvidia.com/zh-cn/blog/streamlining-ai-inference-performance-and-deployment-with-nvidia-tensorrt-llm-chunked-prefill/ | 借助 NVIDIA TensorRT-LLM 分块预填充提高 AI 推理效率和简化部署 | 在本
博文
中,我们详细介绍了分块预填充,这是
NVIDIA TensorRT-LLM
的一项功能,可提高 GPU 利用率并简化开发者的部署体验。本文基于之前的博文,讨论了 TensorRT-LLM 中的高级 KV 缓存优化功能如何在需要系统预填充的用例中将性能提升高达 5 倍。
传统预填充和解码推理方法面临的挑战
当用户向模型提交请求时,它将经历两个不同的计算阶段:预填充和解码。每个阶段使用 GPU 计算资源的方式各不相同。
在预填充阶段,系统会处理所有输入令牌以计算 KV 缓存,然后使用该缓存生成输出的第一个令牌。此阶段的计算要求很高,可以有效利用 GPU 庞大的并行计算资源。
在解码阶段,系统会单独生成输出 tokens,使用... |
https://developer.nvidia.com/blog/5x-faster-time-to-first-token-with-nvidia-tensorrt-llm-kv-cache-early-reuse/ | 5x Faster Time to First Token with NVIDIA TensorRT-LLM KV Cache Early Reuse | In our previous
blog post
, we demonstrated how reusing the key-value (KV) cache by offloading it to CPU memory can accelerate time to first token (TTFT) by up to 14x on x86-based NVIDIA H100 Tensor Core GPUs and 28x on the NVIDIA GH200 Superchip. In this post, we shed light on KV cache reuse techniques and best practi... | https://developer.nvidia.com/zh-cn/blog/5x-faster-time-to-first-token-with-nvidia-tensorrt-llm-kv-cache-early-reuse/ | NVIDIA TensorRT-LLM KV 缓存早期重用实现首个令牌速度 5 倍提升 | 在上一篇
博客文章
中,我们展示了通过将键值 (KV) 缓存卸载到 CPU 内存来重复使用该缓存如何在基于 x86 的 NVIDIA H100 Tensor Core GPU 上将首次生成令牌 (Time To First Token, TTFT) 的速度提升高达 14 倍,在 NVIDIA GH200 Superchip 上可将其提升高达 28 倍。在本文中,我们介绍了可以进一步推动 TTFT 加速的 KV 缓存重复使用技术和最佳实践。
KV 缓存简介
许多任务(包括问答和代码生成)都迅速采用了 LLM 模型。为了生成响应,这些模型首先将用户的提示符转换为 tokens,然后将其转换为 dense vectors。随后会进行大量的... |
https://developer.nvidia.com/blog/3x-faster-allreduce-with-nvswitch-and-tensorrt-llm-multishot/ | 3x Faster AllReduce with NVSwitch and TensorRT-LLM MultiShot | Deploying generative AI workloads in production environments where user numbers can fluctuate from hundreds to hundreds of thousands – and where input sequence lengths differ with each request – poses unique challenges. To achieve low latency inference in these environments, multi-GPU setups are a must – irrespective o... | https://developer.nvidia.com/zh-cn/blog/3x-faster-allreduce-with-nvswitch-and-tensorrt-llm-multishot/ | NVSwitch 和 TensorRT-LLM MultiShot 共同加速 AllReduce 速度达 3 倍 | 在用户数量可能在数百到数十万之间波动,且输入序列长度随每个请求而变化的生产环境中,部署生成式 AI 工作负载会面临独特的挑战。要在这些环境中实现低延迟推理,无论 GPU 生成方式或显存容量如何,多 GPU 设置都是必需的。为了提高生产级设置中的推理性能,我们很高兴推出 TensorRT-LLM Multi-shot,这是一种新的多 GPU 通信协议,利用
NVIDIA NVLink Switch
可将通信速度大幅提升高达 3 倍。本博客概述了这一新功能,以及它如何帮助开发者和解决方案架构师克服传统多 GPU 通信方法的限制。
传统 AllReduce 算法面临的挑战
对于低延迟推理,无论单个 GPU 的显存容量如何,多 GPU 都至... |
https://developer.nvidia.com/blog/accelerating-scientific-literature-reviews-with-nvidia-nim-microservices-for-llms/ | Accelerating Scientific Literature Reviews with NVIDIA NIM Microservices for LLMs | A well-crafted systematic review is often the initial step for researchers exploring a scientific field. For scientists new to this field, it provides a structured overview of the domain. For experts, it refines their understanding and sparks new ideas. In 2024 alone, 218,650 review articles were indexed in the
Web of ... | https://developer.nvidia.com/zh-cn/blog/accelerating-scientific-literature-reviews-with-nvidia-nim-microservices-for-llms/ | 借助适用于 LLM 的 NVIDIA NIM 微服务加速科学文献审查 | 精心设计的系统审查通常是研究人员探索科学领域的第一步。对于刚刚接触该领域的科学家,它提供了该领域的结构化概述。对于专家而言,它可以细化他们的理解并激发新的想法。在 2024 年,
Web of Science
数据库收录了 218,650 篇评测文章,凸显了这些资源在研究中的重要性。
完成系统审查可显著增强研究人员的知识库及其学术影响。然而,传统的评论文章需要收集、阅读和总结有关特定主题的大量学术文章。由于这项人工练习非常耗时,处理过的文献的范围通常局限于几十篇或几百篇文章。跨学科内容 (通常超出研究人员的专业领域) 又增加了一层复杂性。
这些挑战使得创建全面、可靠且有影响力的系统审查变得越来越困难。
大语言模型 (LLMs) 的出... |
https://developer.nvidia.com/blog/vision-language-model-prompt-engineering-guide-for-image-and-video-understanding/ | Vision Language Model Prompt Engineering Guide for Image and Video Understanding | Vision language models
(VLMs) are evolving at a breakneck speed. In 2020, the first VLMs revolutionized the
generative AI
landscape by bringing visual understanding to
large language models
(LLMs) through the use of a
vision encoder
. These initial VLMs were limited in their abilities, only able to understand text and ... | https://developer.nvidia.com/zh-cn/blog/vision-language-model-prompt-engineering-guide-for-image-and-video-understanding/ | 图像和视频理解的视觉语言模型提示工程实践指南 | 视觉语言模型 (VLMs) 正在以极快的速度发展。2020 年,首批 VLMs 通过使用视觉编码器将视觉理解引入大语言模型 (LLMs) ,彻底改变了生成式 AI 格局。这些初始 VLMs 能力有限,只能理解文本和单张图像输入。
几年后,VLM 现在能够理解多图像和视频输入,以执行高级视觉语言任务,例如视觉问答 (VQA)、字幕、搜索和摘要。
图 1、NVIDIA NeMo 中的模型自定义工具套件
通过调整提示和模型权重,可以提高特定用例的 VLM 准确性。可以使用 PEFT 等高效技术微调模型权重,但仍需要足够的数据和计算。
但是,
prompt engineering
和上下文学习也可用于提高输出质量,并且与在运行时调整模型的文... |
https://developer.nvidia.com/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | AI for Climate, Energy, and Ecosystem Resilience at NVIDIA GTC 2025 | From mitigating climate change to improving disaster response and environmental monitoring, AI is reshaping how we tackle critical global challenges. Advancements in fast, high-resolution climate forecasting, real-time monitoring, and digital twins are equipping scientists, policy-makers, and industry leaders with data... | https://developer.nvidia.com/zh-cn/blog/ai-for-climate-energy-and-ecosystem-resilience-at-nvidia-gtc-2025/ | NVIDIA GTC 2025 上的人工智能促进气候、能源和生态系统复原力 | 从减缓气候变化到改进灾害响应和环境监测,AI 正在重塑我们应对重大全球挑战的方式。快速、高分辨率的气候预报、实时监控和数字孪生技术的进步为科学家、政策制定者和行业领导者提供了数据驱动的工具,帮助他们了解、规划和应对一个变暖的星球。
在 3 月 17 日至 21 日举行的
NVIDIA GTC 2025
大会上,思想领袖、科学家、开发者和创新者将重点介绍 AI 如何帮助塑造更具可持续性和韧性的未来。以下会议展示了 AI 在气候预测、灾难缓解和保护工作中发挥的作用,帮助社区适应日益不可预测的世界。
加强极端天气预测和灾害响应
随着全球变暖加剧,极端天气事件变得更加严重和频繁,社区需要更快、更精确的自然灾害预测和响应策略。AI 正在改进洪... |
https://developer.nvidia.com/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | Automating GPU Kernel Generation with DeepSeek-R1 and Inference Time Scaling | As AI models extend their capabilities to solve more sophisticated challenges, a new
scaling law
known as
test-time scaling
or
inference-time scaling
is emerging. Also known as
AI reasoning
or
long-thinking
, this technique improves model performance by allocating additional computational resources during inference to ... | https://developer.nvidia.com/zh-cn/blog/automating-gpu-kernel-generation-with-deepseek-r1-and-inference-time-scaling/ | 使用 DeepSeek-R1 和推理时间缩放实现自动化 GPU 内核生成 | 随着 AI 模型扩展其功能以解决更复杂的挑战,一种称为“
测试时扩展
”或“
推理时扩展
”的新扩展法则正在出现。该技术也称为
AI 推理
或
长时思考
技术,通过在推理过程中分配额外的计算资源来评估多种可能的结果,然后选择最佳结果 (neural network),从而提高模型性能。这使得 AI 能够以类似于人类剖析复杂问题并单独解决这些问题以达成最终解决方案的方式,制定战略并系统化地解决复杂问题。
在本文中,我们将介绍 NVIDIA 工程师完成的一项实验,他们在推理过程中使用最新的开源模型之一 DeepSeek-R1 模型以及额外的计算能力来解决复杂的问题。该实验旨在自动生成 GPU 注意力内核,这些内核在数值上是正确的,并针对... |
https://developer.nvidia.com/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI Foundation Model Enhances Cancer Diagnosis and Tailors Treatment | A new study and AI model from researchers at Stanford University is streamlining cancer diagnostics, treatment planning, and prognosis prediction. Named MUSK (Multimodal transformer with Unified maSKed modeling), the research
aims to advance precision oncology, tailoring treatment plans to each patient based on their u... | https://developer.nvidia.com/zh-cn/blog/ai-foundation-model-enhances-cancer-diagnosis-and-tailors-treatment/ | AI 基础模型增强癌症诊断并实现个性化治疗 | 斯坦福大学研究人员的一项新研究和 AI 模型正在简化癌症诊断、治疗规划和预后预测。这项名为 MUSK (Multimodal transformer with Unified maSKed modeling) 的研究旨在提高精准肿瘤学,根据每位患者独特的医疗数据为其定制治疗计划。
“多模态基础模型是医学 AI 研究的新领域,”放射肿瘤学副教授兼研究高级作者
Ruijiang LI
说。“最近,我们为医学领域开发了视觉语言基础模型,尤其是在病理学领域。但是,现有研究使用的现有基础模型需要配对的图像 – 文本数据进行预训练。尽管我们付出了大量努力,最终打造出 1M 病理图像文本对,但它仍然不足以完全捕捉整个疾病谱系的多样性。”
在考虑患... |
https://developer.nvidia.com/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA Toolkit Now Available for NVIDIA Blackwell | The latest release of the
CUDA Toolkit
, version 12.8, continues to push accelerated computing performance in data sciences, AI, scientific computing, and computer graphics and simulation, using the latest NVIDIA CPUs and GPUs. This post highlights some of the new features and enhancements included with this release:
N... | https://developer.nvidia.com/zh-cn/blog/cuda-toolkit-12-8-delivers-nvidia-blackwell-support/ | CUDA 工具包现已支持 NVIDIA Blackwell 架构 | CUDA 工具包
的最新版本 (版本 12.8) 使用最新的 NVIDIA CPU 和 GPU,持续提升数据科学、AI、科学计算以及计算机图形和模拟领域的加速计算性能。本文重点介绍了此版本包含的一些新功能和增强功能:
NVIDIA Blackwell 架构支持
CUDA 图形处理条件节点增强功能
用于大语言模型(LLMs)的 Blackwell CUTLASS 内核
NVIDIA Nsight 开发者工具更新
数学库更新
cudaStreamGetDevice
编译器更新
加速 Python 更新
功能齐全的架构
NVIDIA Blackwell 架构支持
CUDA 工具包 12.8 是该工具包的第一个版本,在整个开发者工具套件 (... |
https://developer.nvidia.com/blog/recent-posts/ | Recent posts | No content found | https://developer.nvidia.com/zh-cn/blog/recent-posts/ | 最近文章 | No content found |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.