H100 vs A100
Same 80 GB of memory, two generations apart. What separates them is bandwidth, FP8, and the rate each provider is charging today.
- Both hold 80 GB, so any model that fits on one fits on the other. Capacity is not the decision.
- The H100 SXM5 80GB publishes 3.2× the dense BF16 throughput but only 1.6× the memory bandwidth — two different ratios, and which one you feel depends on the workload.
- It rents for 1.74× the price cheapest-to-cheapest, across the 13 providers listing both.
- We have not measured either card on any workload. Everything above is the vendors' own peak figures beside today's published rates.
Violet marks the stronger figure. Throughput numbers are the vendors' own published dense peaks, read from H100 SXM5 80GB's product page and A100 SXM4 80GB's product page — not speeds we measured. Process node and transistor count are absent because they are not on the pages we read; architecture comes from our hand-compiled hardware registry, which has no fetch behind it.
On-demand, per GPU, USD, from providers publishing a comparable rate for both cards. 12 more list one of the two but not both. Spot and committed rates are excluded — they are not comparable to on-demand, and mixing them would turn a different product into what looks like a discount.
Memory is a tie, bandwidth is not
- Identical 80 GB on both, but 3,350 GB/s against 2,039 GB/s — the H100 SXM5 80GB moves weights 1.6× faster.
- That is a smaller gap than the 3.2× in BF16 compute. Work limited by how fast weights move will track the first number; work limited by arithmetic will track the second.
FP8 is present on one side and absent on the other
- The H100 SXM5 80GB publishes 1,979 dense TFLOPS at FP8. The A100 SXM4 80GB publishes no FP8 figure at all.
- That is an absent capability rather than a smaller number — the blank cell in the spec table is the silicon, not a gap in our data.
Where you rent moves more money than which card you pick
- The same H100 SXM5 80GB is published between $1.30 and $12.84 — a 9.9× spread across 24 providers for identical hardware.
- The gap between the two cards is 1.74×. Choosing the provider is worth more than choosing the card, and it is the part a spec sheet cannot tell you.
More pairings need both cards' specs fetched first.