NVIDIA Rubin CPXNVIDIA Rubin CPX
未公布单价No public price未公布单价No public price · 官方页未给公开单价;这类产品通常按整机、云实例或询价成交,不拿经销商价冒充官方价。No public unit price on the official page; these usually sell as systems, cloud instances or on request, so dealer quotes are not presented as official.
回答「百万 token 上下文推理该看哪类卡」。Rubin CPX 是专为长上下文推理切的单 die GPU:30 PF NVFP4、128GB GDDR7,注意力能力官方称是 GB300 NVL72 的 3 倍;整机柜 Vera Rubin NVL144 CPX 给 8 EF AI 算力、100TB 快存、1.7PB/s 带宽。它不等于训练卡:GDDR7 容量大但带宽低于 HBM4,官方定位是推理与视频编解码。可用时点是 2026 年底。Answers which class of GPU to look at for million-token context inference. Rubin CPX is a monolithic GPU cut for long-context inference: 30 PF at NVFP4 and 128GB of GDDR7, with NVIDIA claiming 3x the attention capability of GB300 NVL72; a full Vera Rubin NVL144 CPX rack delivers 8 exaflops, 100TB of fast memory and 1.7PB/s. It is not a training card: GDDR7 gives capacity, not HBM4-class bandwidth, and NVIDIA positions it for inference plus video encode/decode. Availability is the end of 2026.
关键参数Key specs
- architecture
- NVIDIA Rubin
- NVFP4 compute
- up to 30 PFLOPS
- memory
- 128 GB GDDR7
- attention
- 3× GB300 NVL72 (vendor claim)
- rack platform
- Vera Rubin NVL144 CPX
- rack AI compute
- 8 exaflops
- rack fast memory
- 100 TB
- rack bandwidth
- 1.7 PB/s
- availability
- end of 2026
官方原文怎么说What the source says
Rubin CPX delivers up to 30 petaflops of compute with NVFP4 precision
It features 128GB of cost-efficient GDDR7 memory
packs 8 exaflops of AI performance and 100TB of fast memory in a single rack
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-19
- 许可 / 类型Licence / kind
- official-spec-page
- 许可说明Licence evidence
- https://nvidianews.nvidia.com/news/nvidia-unveils-rubin-cpx-a-new-class-of-gpu-designed-for-massive-context-inference
- 类目Category
- 加速卡Accelerators
- 状态Status
- 已发布 / 未出货Announced
- 价格口径Price basis
- 未公布单价No public price · not-published