Skip to content
ANSWER · SELECTION

A100 or H100: which is cheaper per unit of compute?

At medians H100 costs 2.83x what A100 costs and delivers 3.17x the BF16 peak — 11% cheaper for H100 per unit of peak compute. That is not a selection answer: it depends on the price you can get and on whether your workload actually uses FP8 and NVLink.

Quote this: Simultaneous median asking prices: H100 SXM $2.388/GPU/h, A100 SXM4 $0.845/GPU/h. Per dollar of BF16 peak: $2.41/PFLOPS-hour for H100 vs $2.71 for A100. — CVIN, 2026-11-05, https://cvin.org/en/answers/a100-vs-h100

Prices: device-level listings on vast.ai / Clore.ai, latest per machine · as of 2026-11-05 · peak FLOPS are vendor dense specs

01

Compute per dollar

型号与价位单卡价/时BF16 峰值 TFLOPS$/PFLOPS·时每美元 TFLOPS
A100 SXM4 · P10$0.401 312$1.29 777
A100 SXM4 · median$0.845 312$2.71 369
H100 SXM · P10$2.134 990$2.16 464
H100 SXM · median$2.388 990$2.41 415

$/PFLOPS·时 越低越划算。这张表只能回答「预算买算力」这一个问题,回答不了「哪台机器跑得快」。

02

Sample and supply

型号在架机器有价观测区间(min–max)P10–P90
A100 SXM49999$0.309 – $2.004$0.401 – $1.337
H100 SXM5454$1.788 – $4.167$2.134 – $3.405

供给量本身是决策变量:H100 在架 54 台、A100 在架 99 台。要一次性拿 8 台以上 H100,可选的供给比 A100 少,实际能谈到的价往往偏离中位。

03

How it is computed

每美元算力 = 单卡时价 ÷ 峰值算力。峰值取厂商公开的 BF16 dense 数字(A100 312 / H100 990 TFLOPS,不含稀疏加速),因为稀疏加速在真实负载里很少能吃到。两边的价格都取同一批采集里的最新观测,保证是「同一时刻」的比较,而不是拿三个月前的 A100 价对今天的 H100 价。

代入自己的数:$/PFLOPS·时 = 单卡时价 × 1000 ÷ 峰值 TFLOPS

04

When the conclusion inverts

① 拿到 P10 价时结论会变。A100 的 P10 是 $0.401,每 PFLOPS·时 $1.29;H100 的 P10 档是 $2.16——A100 比 H100 便宜 40%。能不能拿到低价,比选哪张卡更影响账单。
② 吃 FP8 的负载会大幅偏向 H100。峰值表里没有 FP8;大 batch 推理、长序列训练用 FP8 后,H100 的有效吞吐常明显超过 BF16 峰值比。
③ 多卡通信密集的负载,NVLink 是变量。H100 SXM 的卡间带宽高于 A100,张量并行度高的训练在 A100 上会先撞通信墙,那时便宜就是假便宜。
05

Related

NaN