AI Practices AI实践 6h ago Updated 1h ago 更新于 1小时前 50

NVIDIA Vera Rubin and Blackwell Set a New Standard for Agentic AI Performance per Watt 英伟达Vera Rubin和Blackwell为Agentic AI的每瓦性能树立新标准

SemiAnalysis AgentX is an open-source benchmark in the InferenceX suite that evaluates agentic AI inference by replaying production-style coding agent sessions, capturing long-context prefill, KV-cache reuse, tool-call gaps, and dynamic concurrency NVIDIA Vera Rubin NVL72 achieves up to 30x higher AI-factory throughput per megawatt compared to GB300 NVL72 on AgentX workloads GB300 NVL72 delivers up to 80x throughput-per-megawatt gains over H200 NVL8 for large MoE models like Kimi K3 2.8T on agen SemiAnalysis AgentX基准测试通过回放生产级编码agent会话,准确捕捉长上下文prefill、KV-cache复用、tool-call间隙和动态并发等agentic AI推理特征 NVIDIA Vera Rubin NVL72在AgentX工作负载下实现比GB300 NVL72高达30倍的每兆瓦AI工厂吞吐量 GB300 NVL72在大型MoE模型(如Kimi K3 2.8T)上相比H200 NVL8实现高达80倍的吞吐量提升 性能优化来自系统级改进,包括MoE serving运行时(SGLang/TensorRT-LLM/vLLM)、DeepGEMM内核、MXFP4/MXF

72
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • SemiAnalysis AgentX is an open-source benchmark in the InferenceX suite that evaluates agentic AI inference by replaying production-style coding agent sessions, capturing long-context prefill, KV-cache reuse, tool-call gaps, and dynamic concurrency
  • NVIDIA Vera Rubin NVL72 achieves up to 30x higher AI-factory throughput per megawatt compared to GB300 NVL72 on AgentX workloads
  • GB300 NVL72 delivers up to 80x throughput-per-megawatt gains over H200 NVL8 for large MoE models like Kimi K3 2.8T on agentic inference
  • Key efficiency drivers include MoE serving runtimes (SGLang, TensorRT-LLM, vLLM), DeepGEMM-based kernels, mixed-precision formats (MXFP4, MXFP8), the NVIDIA Dynamo session-aware serving stack, and NVLink scale-up fabric connecting 72 GPUs
  • Real-world data shows agentic requests consume 15x more tokens than ordinary chat, with average prompt tokens per request growing fourfold across 100 trillion tokens of usage

Why It Matters

As AI agents become the dominant inference workload, traditional benchmarks using fixed sequence lengths no longer reflect real serving performance—making new evaluation frameworks like AgentX essential for hardware purchasing and optimization decisions. The dramatic throughput-per-watt gains across NVIDIA generations signal that energy efficiency, not just raw compute, is becoming the critical differentiator for AI infrastructure at scale.

Technical Details

  • AgentX Benchmark Methodology: Uses prerecorded Claude Code sessions replayed turn-by-turn via the AIPerf client, preserving context accumulation, input/output sequence lengths, reasoning time, and tool-call latency to reproduce realistic KV cache pressure
  • Key Metrics: Tokens per megawatt reported against four user-experience values—E2E normalized interactivity, standard interactivity, E2E latency, and time to first token (TTFT)—enabling trade-off analysis between throughput and responsiveness
  • System-Level Optimizations: Leverages MoE serving runtimes (SGLang, TensorRT-LLM, vLLM), DeepGEMM-based kernels, mixed-precision formats (MXFP4, MXFP8), the NVIDIA Dynamo session-aware serving stack, and high-bandwidth NVLink scale-up fabric for 72-GPU coordinated rack-scale inference
  • Generational Performance: GB300 NVL72 leads H200 by up to 40x tokens per megawatt on static 8K/1K DeepSeek-R1-0528 workloads, with gains extending to 80x for large MoE models on dynamic agentic workloads
  • Shift in Benchmarking: Fixed-sequence-length scenarios have been demoted to "maintenance mode" in InferenceX as they no longer represent production agent traffic, which features variable request lengths, interleaved tool use, and growing context

Industry Insight

  • AI infrastructure buyers should prioritize throughput-per-megawatt metrics on agentic workloads rather than static benchmark scores, as the gap between generations widens significantly under real agent traffic patterns
  • The fourfold growth in prompt tokens and 15x increase in per-request token consumption for agentic workloads signals that KV-cache management and context reuse will be increasingly critical system design concerns
  • Hardware vendors that integrate session-aware serving stacks and efficient MoE execution at the rack scale will gain a decisive advantage as agentic AI becomes the leading inference workload across the industry

TL;DR

  • SemiAnalysis AgentX基准测试通过回放生产级编码agent会话,准确捕捉长上下文prefill、KV-cache复用、tool-call间隙和动态并发等agentic AI推理特征
  • NVIDIA Vera Rubin NVL72在AgentX工作负载下实现比GB300 NVL72高达30倍的每兆瓦AI工厂吞吐量
  • GB300 NVL72在大型MoE模型(如Kimi K3 2.8T)上相比H200 NVL8实现高达80倍的吞吐量提升
  • 性能优化来自系统级改进,包括MoE serving运行时(SGLang/TensorRT-LLM/vLLM)、DeepGEMM内核、MXFP4/MXFP8混合精度、NVIDIA Dynamo会话感知服务栈及NVLink 72 GPU互联
  • AI agent推理已从单轮交互演变为多步骤工作流,OpenRouter数据显示平均prompt tokens增长4倍,单次agent请求消耗token量是普通chat的15倍

为什么值得看

本文揭示了agentic AI时代硬件性能评估的新范式,传统固定序列长度基准已无法反映真实工作负载,AgentX为AI基础设施选型提供了更贴近生产环境的评估标准。对AI从业者而言,理解Vera Rubin和Blackwell在能效比上的代际优势,有助于在AI工厂部署中做出更优的硬件投资决策。

技术解析

  • AgentX基准测试方法:基于InferenceX开源套件,通过AIPerf客户端回放预录制的Claude Code会话,保留原始会话的上下文累积、输入/输出序列长度、推理时间和tool-call延迟,真实复现KV-cache容量压力场景
  • 性能指标体系:核心指标为每兆瓦吞吐量(tokens per megawatt),配合四个用户体验维度:E2E归一化交互性(总输出token/端到端时间)、标准交互性(生成阶段token率)、E2E延迟(单请求总耗时)、TTFT(首token时间)
  • 系统级优化技术栈:包括SGLang/TensorRT-LLM/vLLM等MoE serving运行时、DeepGEMM-based kernels、MXFP4/MXFP8混合精度格式、NVIDIA Dynamo会话感知服务栈、NVLink高带宽scale-up互联(72 GPU机架级协同)
  • 代际性能对比:GB300 NVL72在静态8K/1K序列测试中已领先H200 NVL8达40倍,在动态agentic工作负载下进一步扩展至80倍(针对Kimi K3 2.8T等大规模MoE模型)

行业启示

  • 基准测试范式转变:固定序列长度测试已进入"维护模式",AI基础设施评估需转向捕捉长上下文、KV-cache复用、工具调用间隙和动态并发的真实工作负载基准,AgentX代表了这一趋势
  • 能效比成为AI工厂核心指标:随着agentic AI成为主导工作负载,每兆瓦吞吐量直接决定运营成本,NVIDIA Vera Rubin和Blackwell的代际优势为云厂商和AI工厂的硬件采购提供明确参考
  • 基础设施需适配agent工作负载特征:AI agent的多步骤推理、工具调用和上下文累积特性要求-serving栈具备会话感知能力,传统单轮交互优化策略已不适用,需重新评估推理引擎和硬件选型

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

GPU GPU Inference 推理 Benchmark 基准测试 Agent Agent Chip 芯片