跳到主要内容Skip to main content

← 返回硬件算力← Back to Compute

云算力Cloud compute

AWS EC2 P6e-GB200 UltraServerAWS EC2 P6e-GB200 UltraServer

云上租一整个 72 卡 NVLink 域,不是租单卡;360 PFLOPS 是 FP8 无稀疏口径,13.4 TB 是域内 HBM 合计。适合万亿参数级训练与推理;本站不采价格,预算要按 AWS 官方定价页另算。与 P5en 相比,AWS 声称算力 20 倍、显存 11 倍、NVLink 聚合带宽 15 倍。Rents a whole 72-GPU NVLink domain rather than a single card. 360 PFLOPS is FP8 without sparsity and 13.4 TB is aggregate HBM in the domain. It targets trillion-parameter training and inference; this site does not collect prices, so budget from the official AWS pricing page. AWS claims 20x compute, 11x memory and 15x aggregate NVLink bandwidth versus P5en.

关键参数Key specs

GPU 数 / GPUs
up to 72 Blackwell GPUs per UltraServer
显存 / HBM
13.4 TB HBM3e total
FP8 算力 / FP8
360 PFLOPS (without sparsity)
NVLink
130 TB/s
对外网络 / scale-out
28.8 Tb/s EFAv4
售卖单位 / unit
UltraServer

官方原文怎么说What the source says

customers can access up to 72 Blackwell GPUs within one NVLink domain
1 https://aws.amazon.com/ec2/instance-types/p6/
360 petaflops of FP8 compute (without sparsity) and 13.4 TB of total high- bandwidth memory (HBM3e)
2 https://aws.amazon.com/ec2/instance-types/p6/
130 terabytes per second of low-latency NVLink connectivity between GPUs
3 https://aws.amazon.com/ec2/instance-types/p6/

资料与许可Source and licence

资料时间Source checked
2026-11-17
许可 / 类型Licence / kind
official-spec-page
许可说明Licence evidence
https://aws.amazon.com/ec2/instance-types/p6/
类目Category
云算力Cloud compute

去官网看原页Open the official page →

← 返回硬件算力← Back to Compute