NVIDIA Rubin GPU: Driving the Era of Agentic AI

Speakers: Manas Mandal, Raj Dash, Rouslan Dimitrov(NVIDIA)
Venue: Hot Chips 2026 Conference
PDF: raw/papers/HC2026_NVIDIA_Rubin.pdf(非公开本地路径)

同日对照:Helios UALoE(也是 72-GPU scale-up)、Vera CPU。更新 Vera Rubin NVL72 与 NVLink fabric。

七芯片五机架

Vera Rubin 写成 full-stack AI factory:七芯片、五机架 — NVL72、Groq 3 LPX、Vera CPU rack、Vera BlueField-4 STX、Spectrum-6 SPX。角色:Foundation Platform / Extend Interactivity / Tool Calls and Sandboxes / Context Memory Storage / Scale-out Fabric。

Agentic 对标传统推理:输入从固定 1k/8k/32k 变成 32k & 100k & 400k;多轮 + tool calling。优化目标改成 token revenue:Δ tokens/watt、Δ TTFT、Δ MTBI、Δ useful life。

AgentX 图标 Unofficial Results. Pending SemiAnalysis Review:DeepSeek-v4-PRO,140K+ context;纵轴 TPS per MW 到 60M,横轴 TPS/User 到 300。NVL72 vs GB300 NVL72 标 2X / 10X / 30X(高交互端 30X)。给定 TPS/User 的绝对坐标 未知。

100 MW factory(不是单 GPU)

脚注:Specifications based on at scale AI factory using DSX with MaxLPS。2 ZFLOPS / 11 PB / 800 PB/s 是 100 MW factory,不是单卡:NVFP4 inference 2 ZFLOPS;NVFP4 training 1.4 ZFLOPS;HBM4 11 PB;HBM4 BW 800 PB/s。单卡 HBM4 GB / TB/s、GPU TDP、工艺节点:本 deck 未知。

Die 标签:Enhanced 5th Gen Tensor Cores;NVLink-C2C coherent CPU–GPU;x16 PCIe Gen6;HBM CTRL / distributed L2 / GPC;NVLink v6;Confidential Computing TEE-I/O capable。图上还有 NV-HBI、MIG CONTROL、GIGATHREAD ENGINE、NV-DEC。

2:4 稀疏与 Counted Write

稀疏路线:2023 FP8 weight sparsity → 2024/2025 NVFP4 → 2026 Rubin Adaptive Compression Sparsity。2:4 sparsity 比上代 NVFP4 sparsity 更通用;多数情况「no change to model and no finetuning」;inference runtime opt-in。注意力:BMM1 → LDTM.Sparsify → SoftMax → BMM2;「2x faster downstream SoftMax and BMM2」。精度演示是 Qwen-Image BF16 dense vs NVFP4 sparse(分数表 未知)。

Counted writes 替换 Blackwell 的 MEMBAR + atomic-flag GPU-to-GPU sync。延迟 µs 未知。

72 GPU scale-up domain;3.6 TB/s per GPU all-to-all BW。相对 Ethernet:「3x lower latency」;130 TFLOPS in-network compute;10x higher packet rate;另有 4X callout,分母 未知。

3rd-gen MGX:all-copper;80+ MGX partners;350+ factory sites / 30 countries;45°C liquid cooling;no retimers;800 VDC。TCS supply 45 °C / return 55 °C vs 传统 35 / 45 °C。LLM training 上 intelligent power smoothing 13% peak power reduction;叠加其它改进「up to 40% more GPUs per provisioned watt」。2nd-gen RAS Engine(RIST):seconds 级 zero-downtime GPU health checks;in-field SRAM repair;HBM bank re-mapper。

Citations

[1] raw/papers/HC2026_NVIDIA_Rubin.pdf(非公开本地路径) — Mandal / Dash / Dimitrov, Hot Chips 2026 [2] hc2026-nvidia-rubin.md — 结构化摘录