跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

加速卡Accelerators

NVIDIA B300(Blackwell Ultra)NVIDIA B300 (Blackwell Ultra)

回答「Blackwell Ultra 单卡在 288GB 档给多少 dense NVFP4 与带宽」。15 PF 是 dense NVFP4,不是稀疏数字;8 TB/s 与 B200 同级,容量也一样是 288GB,差别在 NVFP4 算力与能效。
不主张:15 PF 不能与 Rubin 的 50 PF NVFP4 推理数字直接横比;不是整机 GB300 NVL72 的算力;1400W 是最高 TGP,实际功耗随负载与配置变化。
Answers what Blackwell Ultra offers at the 288GB tier in dense NVFP4 and bandwidth. 15 PF is dense NVFP4, not sparse; 8 TB/s matches B200, and capacity is also 288GB — the difference is NVFP4 compute and efficiency.
It does not claim: 15 PF is not directly comparable with Rubin’s 50 PF NVFP4 inference figure; it is not the GB300 NVL72 rack number; 1,400W is max TGP, and real power varies with workload and configuration.

关键参数Key specs

显存 / HBM
288 GB HBM3e
带宽 / bandwidth
8 TB/s
NVFP4 dense
15 PFLOPS
SM / Tensor Cores
160 SM / 640 Tensor Cores
晶体管 / transistors
208B
NVLink
1.8 TB/s(NVLink 5)
最高功耗 / max TGP
1,400W

官方原文怎么说What the source says

delivers 15 petaFLOPS of dense NVFP4 compute through 640 fifth-generation Tensor Cores across 160 Streaming Multiprocessors
1 https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era
HBM3E memory capacity reaches 288 GB with 8 TB/s bandwidth
2 https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era
Max power (TGP) Up to 700W Up to 1,200W Up to 1,400W
3 https://developer.nvidia.com/blog/inside-nvidia-blackwell-ultra-the-chip-powering-the-ai-factory-era

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-blackwell-ultra-the-chip-powering-the-ai-factory-era
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