NVIDIA Rubin Shipping Specs Adjusted: Memory Bandwidth Down 11%, External Network Up 12.5%
English text is being prepared.
NVIDIA's September update to the Rubin shipping specs shows that the per-GPU NVFP4 inference compute of 50 PFLOPS, training of 35 PFLOPS, and 288GB HBM4 memory remain unchanged, while memory bandwidth drops from 22TB/s to 19.2TB/s, NVLink drops from 3.6TB/s to 3.0TB/s, and external network bandwidth rises from 400GB/s to 450GB/s. The rack-level NVL72 is adjusted proportionally in tandem, with total memory bandwidth dropping from 1580TB/s to 1400TB/s and external networking rising from 28.8TB/s to 32.4TB/s; the eight-GPU HGX NVL8 changes the most, with memory dropping from 2.3TB to 2.0TB, a 26% reduction in total bandwidth. The article states that the new specs apply to all Rubin hardware from the first day of shipment; NVIDIA did not explain the reason for the downgrade, and the author believes this is a system-level rebalancing in response to heavier cross-rack communication.
Why read this
AI score 60
The article compares the bandwidth changes in Rubin's shipping specs item by item and explains why memory, interconnect, and external networking were rebalanced.
Topics
Body
期货推广begin 期货推广end 秒拍begin 秒拍end 视频播放器start 视频播放器end 行情图begin 行情图end
来源:芝能汽车article_adlist[
芝能智芯出品 ]article_adlist
甲骨文宣布Vera Rubin NVL72在OCI云平台进入通用可用,Rubin开始真正给客户跑业务。
英伟达9月悄悄更新的出货规格被翻了出来:算力和显存容量是一样的,两条内部带宽一起下调,对外网络反而加宽。Rubin发布接近一年,官方一直标注数值可变。
新规格从出货第一天就适用于所有Rubin硬件,没有过渡期。
单颗Rubin GPU,NVFP4推理算力50 PFLOPS、训练35 PFLOPS,288GB HBM4由8组36GB堆栈组成,这些全部不动。
变化的是这些:
◎显存带宽从22TB/s降到19.2TB/s,对应引脚速率从10.75Gbps降到9.375Gbps,降幅约11%;
◎NVLink从3.6TB/s降到3.0TB/s,降17%;
◎面向外部网络的带宽从400GB/s升到450GB/s,升12.5%。
机架级等比例缩水:
◎NVL72的显存总带宽从1580TB/s降到1400TB/s,
◎NVLink从260TB/s降到216TB/s,
◎对外网络从28.8TB/s升到32.4TB/s,20.7TB显存容量保持不变。
八卡HGX NVL8差异最大:
◎显存从2.3TB砍到2.0TB,8组HBM4只用7组,单卡只剩约250GB、带宽约16.8TB/s;
◎显存总带宽从176TB/s掉到130TB/s,降幅26%,比NVL72的11%下降多得多;
◎NVLink从28.8TB/s降到24TB/s。
x86版、Vera版和基于NVL8的DGX全部适用。
Part 2
带宽为什么先变化
HBM4把单堆栈接口从1024位翻倍到2048位,JEDEC基准速率是8Gbps/引脚。
◎英伟达原定10.75Gbps远超标准,降到9.375Gbps仍在基准之上;
◎对手AMD的MI455X只跑7.6Gbps。
英伟达没有解释下调原因,似乎是把激进目标往回收,高带宽路线没有变,行业推测指向内存控制器速度、堆栈速度或良率,或三者叠加。
对外网络为何反向加宽?
模型训练跨成千上万颗GPU,推理把预填充、解码和缓存分散部署,跨机架通信越来越重。
显存、互连、网络三层的瓶颈会随部署规模互相转移,这次调整是一次系统级重新配比,对重度跨机架通信的大规模部署,这反而有利。article_adlist[
小结 ]article_adlistarticle_adlist[
英伟达对产品的规格调整果然是动态的。 ]article_adlist
特别声明:以上内容仅代表作者本人的观点或立场,不代表新浪财经头条的观点或立场。如因作品内容、版权或其他问题需要与新浪财经头条联系的,请于上述内容发布后的30天内进行。 news_keyword_pub,stock, 正文下iframe 正文下iframe 总声明 总声明 文末二维码 start 文末二维码 start
文末二维码 start 文末二维码 start 编辑姓名及工作代码 编辑姓名及工作代码