…
H100 NVL vs H100 SXM
Cloud rental price, compute per dollar and specs, side by side.
Open in the interactive comparer →
H100 NVL$2.59lowest on-demand per GPU·h
$3.10 per PFLOP·h · 5 providersH100 SXM$1.66lowest on-demand per GPU·h
$1.68 per PFLOP·h · 28 providersWhat actually differs
- The H100 SXM has 1.2× more dense FP16/BF16 throughput than the H100 NVL (0.835 PF vs 0.989 PF).
- The H100 NVL has 1.2× more memory bandwidth than the H100 SXM (3,900 vs 3,350 GB/s), which usually limits LLM token generation. Memory: 94 GB vs 80 GB.
- The H100 SXM is 36% cheaper per hour than the H100 NVL at the lowest on-demand prices ($2.59 vs $1.66).
- Per unit of compute, the H100 SXM is the better deal: $1.68 vs $3.10 per PFLOP-hour of dense FP16/BF16.
| Spec / price | H100 NVL | H100 SXM |
|---|---|---|
| Memory | 94 GB HBM3better | 80 GB HBM3 |
| Memory bandwidth | 3,900 GB/sbetter | 3,350 GB/s |
| FP16 / BF16 (dense) | 0.835 PF | 0.989 PFbetter |
| FP8 (dense) | 1.67 PF | 1.98 PFbetter |
| FP4 (dense) | — | — |
| INT8 (dense) | 1.67 PF | 1.98 PFbetter |
| TF32 (dense) | 0.417 PF | 0.495 PFbetter |
| FP32 (dense) | 0.06 PF | 0.067 PFbetter |
| FP64 (dense) | 0.06 PF | 0.067 PFbetter |
| Power | 400 W | 700 W |
| Architecture | Hopper (2023) | Hopper (2022) |
| Lowest on-demand | $2.59 | $1.66better |
| Median on-demand | $2.92better | $3.19 |
| Lowest spot | $1.29 | $0.79better |
| FP16/BF16 $ per PFLOP·h | $3.10 | $1.68better |
| Providers | 5 | 28 |
H100 NVL vs H100 SXM: questions
Is the H100 NVL faster than the H100 SXM?
The H100 SXM has 1.2× more dense FP16/BF16 compute than the H100 NVL. The H100 NVL has 1.2× more memory bandwidth than the H100 SXM. Real-world speedups depend on the workload: training and prefill scale with compute, while LLM decoding scales mostly with memory bandwidth.
Which is cheaper per FLOP, the H100 NVL or the H100 SXM?
Right now the H100 SXM is cheaper per unit of compute: $1.68 vs $3.10 per PFLOP-hour at the lowest on-demand prices.
How much does it cost to rent an H100 NVL vs an H100 SXM for a month?
Running one GPU around the clock at the lowest on-demand price costs about $1,891 per month for the H100 NVL and $1,212 for the H100 SXM.
Which has more memory, the H100 NVL or the H100 SXM?
The H100 NVL has more memory: 94 GB vs 80 GB.