NVIDIA A100 40GB SXM vs NVIDIA H100 PCIe 80GB
A100 40GB SXM has 50% less memory than H100 PCIe 80GB (40 GB vs 80 GB). Its memory bandwidth is 22% less (1.555 vs 2 TB/s). In the cloud, the median on-demand A100 40GB SXM price is $1.86 per GPU-hour against $3.04 for the H100 PCIe 80GB. Per gigabyte of GPU memory, that is $0.047 vs $0.038 per hour.
Specifications
Manufacturer figures as recorded by Kovara. Higher memory and bandwidth are highlighted; neither is a benchmark of your workload.
| A100 40GB SXM | H100 PCIe 80GB | |
|---|---|---|
| Manufacturer | NVIDIA | NVIDIA |
| Architecture | Ampere | Hopper |
| Memory | 40 GB HBM2 | 80 GB HBM2E |
| Memory bandwidth | 1.555 TB/s | 2 TB/s |
| Power (TDP) | 400 W | 350 W |
| Form factor | SXM | PCIe dual-slot |
Cloud prices today
Public offers Kovara tracks, normalised to US dollars per GPU-hour. Prices use on-demand terms where available; spot and reserved offers are on each GPU’s price pages. Lower is highlighted.
| A100 40GB SXM | H100 PCIe 80GB | |
|---|---|---|
| Lowest price | $1.15 | $2.49 |
| Median price | $1.86 | $3.04 |
| Highest price | $1.99 | $3.29 |
| Median per GB of memory | $0.047 | $0.038 |
| Public offers (all terms) | 5 | 6 |
| Providers | 3 | 5 |
Where to rent each
Lowest tracked price per provider. Open a provider to see every offer, region and term.
Questions
Which is cheaper to rent, the A100 40GB SXM or the H100 PCIe 80GB?
Right now the cheapest on-demand A100 40GB SXM Kovara tracks is $1.15 per GPU-hour and the cheapest H100 PCIe 80GB is $2.49. Medians are $1.86 and $3.04. The cheaper GPU per hour is not always cheaper per job: a faster or larger GPU can finish sooner or need fewer units.
Which has more memory?
The H100 PCIe 80GB has more memory: 80 GB against 40 GB. Memory decides which models and batch sizes fit on one GPU.
Are these benchmark results?
No. This page compares manufacturer specifications and public prices. Real performance depends on the model, precision, software and interconnect. Kova can help you estimate cost for a specific workload.