跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators

NVIDIA RubinNVIDIA Rubin

回答「Rubin 单卡把 HBM4 做到多少容量与带宽、NVFP4 推理给到多少」。288GB 与 B300 相同,但 HBM4 把带宽从 8 TB/s 拉到 22 TB/s,NVLink 6 从 1.8 TB/s 拉到 3.6 TB/s;这三项决定长上下文与高并发的实际上限。
不主张:50 PF 是 NVFP4 推理性能,不是所有精度;不能与 AMD MXFP4 数字直接横比;不是整机 Vera Rubin NVL72 的合计;不是实测吞吐。
Answers Rubin’s per-GPU HBM4 capacity and bandwidth, and its NVFP4 inference figure. 288GB matches B300, but HBM4 lifts bandwidth from 8 TB/s to 22 TB/s and NVLink 6 from 1.8 TB/s to 3.6 TB/s; those three set the practical ceiling for long context and concurrency.
It does not claim: 50 PF is NVFP4 inference performance, not every precision; it is not directly comparable with AMD MXFP4 figures; it is not the Vera Rubin NVL72 rack total; it is not measured throughput.

关键参数Key specs

显存 / HBM
288 GB HBM4(12-Hi stacks)
带宽 / bandwidth
22 TB/s
NVFP4 推理 / inference
up to 50 PFLOPS
晶体管 / transistors
336B
SM / Tensor Cores
224 SM / 896 Tensor Cores
NVLink 6
3.6 TB/s
形态 / form
GPU;Vera Rubin NVL72 机架

官方原文怎么说What the source says

Rubin integrates up to 288 GB of HBM4 memory, driven by dedicated HBM controllers and 12-Hi stacks, to deliver up to 22 TB/s of peak bandwidth.
1 https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai
Its 336 billion transistors, 224 streaming multiprocessors (SMs), and 896 Tensor Cores
2 https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai
up to 50 petaflops of NVFP4 inference performance
3 https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai
NVLink 6 provides 3,600 GB/s scale-up bandwidth
4 https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai

资料与许可Source and licence

资料时间Source checked
2026-11-18
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://developer.nvidia.com/blog/inside-nvidia-rubin-gpu-architecture-powering-the-era-of-agentic-ai
类目Category
加速卡Accelerators

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute