…
Cloud rental price, compute per dollar and specs, side by side.
Open in the interactive comparer →| Spec / price | H100 PCIe | H100 SXM |
|---|---|---|
| Memory | 80 GB HBM2e | 80 GB HBM3 |
| Memory bandwidth | 2,000 GB/s | 3,350 GB/sbetter |
| FP16 / BF16 (dense) | 0.756 PF | 0.989 PFbetter |
| FP8 (dense) | 1.51 PF | 1.98 PFbetter |
| FP4 (dense) | — | — |
| INT8 (dense) | 1.51 PF | 1.98 PFbetter |
| TF32 (dense) | 0.378 PF | 0.495 PFbetter |
| FP32 (dense) | 0.051 PF | 0.067 PFbetter |
| FP64 (dense) | 0.051 PF | 0.067 PFbetter |
| Power | 350 W | 700 W |
| Architecture | Hopper (2022) | Hopper (2022) |
| Lowest on-demand | $1.66better | $1.82 |
| Median on-demand | $3.22better | $3.51 |
| Lowest spot | $2.00 | $1.46better |
| FP16/BF16 $ per PFLOP·h | $2.20 | $1.83better |
| Providers | 13 | 30 |
The H100 SXM has 1.3× more dense FP16/BF16 compute than the H100 PCIe. The H100 SXM has 1.7× more memory bandwidth than the H100 PCIe. Real-world speedups depend on the workload: training and prefill scale with compute, while LLM decoding scales mostly with memory bandwidth.
Right now the H100 SXM is cheaper per unit of compute: $1.83 vs $2.20 per PFLOP-hour at the lowest on-demand prices.
Running one GPU around the clock at the lowest on-demand price costs about $1,212 per month for the H100 PCIe and $1,325 for the H100 SXM.
Both have 80 GB of VRAM; the H100 SXM has more bandwidth.