Quick answer
H100 vs A100: which should I rent?
H100 vs A100: rent the A100 80GB for memory-bound inference and LoRA fine-tuning of models that fit in 80 GB; rent the H100 for compute-bound training, FP8, or jobs that must finish fast. As of 29 October 2026, GPUwithCrypto rents the H100 at $2.32/hour and the A100 80GB at $0.92/hour, paid in crypto with no KYC. The H100 pays off only if it runs your job about 2.5 times faster.
How do the H100 and A100 specs compare?
Official NVIDIA figures, read on 29 October 2026. Tensor figures are dense peak TFLOPS. NVIDIA prints H100 Tensor values only with sparsity, so the H100 dense values are NVIDIA's figure ÷ 2, marked (d). Peak figures are not benchmarks.
| Spec | H100 SXM | H100 PCIe | A100 80GB SXM | A100 80GB PCIe |
|---|---|---|---|---|
| Architecture | Hopper | Hopper | Ampere | Ampere |
| Memory | 80 GB HBM3 | 80 GB HBM2e | 80 GB HBM2e | 80 GB HBM2e |
| Memory bandwidth | 3,350 GB/s | 2,000 GB/s | 2,039 GB/s | 1,935 GB/s |
| FP16 / BF16 Tensor, dense | 989.5 TFLOPS (d) | — (1) | 312 TFLOPS | 312 TFLOPS |
| FP8 Tensor, dense | 1,979 TFLOPS (d) | — | N/A (2) | N/A (2) |
| NVLink | 900 GB/s | 600 GB/s (bridge) | 600 GB/s | 600 GB/s (2-GPU bridge) |
| PCIe | Gen5 | Gen5 | Gen4 | Gen4 |
| Max power | 700 W | 350 W | 400 W | 300 W |
| Launch | 2022 | 2022 | 2020 | 2021 |
(d) NVIDIA prints 1,979 TFLOPS (BF16) and 3,958 TFLOPS (FP8) with sparsity; dense is half. The A100 sparse figure is 624 TFLOPS. (1) No current official page gives H100 PCIe Tensor values as dense or sparse; an NVIDIA blog cites 1,513 TFLOPS BF16 without saying which. (2) NVIDIA lists FP8 as not available on Ampere.
Which is faster, the H100 or the A100?
On paper, the H100 SXM does 3.17 times the dense BF16 math of an A100 (989.5 vs 312 TFLOPS) and has 1.64 times the memory bandwidth (3,350 vs 2,039 GB/s). These are NVIDIA's peak figures, not benchmarks. Use the two ratios as planning bounds, not as measured speed.
- Compute-bound work (training, fine-tuning at large batch, long prompt prefill) tracks the TFLOPS ratio more closely.
- Memory-bound work (token-by-token LLM decoding at small batch) tracks the bandwidth ratio more closely.
- FP8 runs on the H100's Tensor Cores at 1,979 TFLOPS dense. The A100 has no FP8 Tensor Cores, so an FP8 checkpoint gets no FP8 speed-up there.
The H100 PCIe is a different card. Its 2,000 GB/s is close to the 1,935 GB/s of an A100 80GB PCIe. Instances have 1 to 8 GPUs depending on the host. Before you start, the dashboard shows each host’s country, GPU count, form factor (SXM or PCIe) and interconnect.
Which models fit on an H100 or an A100?
The same models fit on both, because both have 80 GB. Rule of thumb: memory ≈ parameters (billions) × bytes per parameter × 1.2, with 2 bytes for BF16, 1 for FP8/INT8 and 0.5 for 4-bit. The 20% covers KV cache and activations at modest context and batch (EleutherAI). Long contexts need more: see the VRAM guide.
| Model and precision | Memory needed | H100 80GB | A100 80GB |
|---|---|---|---|
| Llama 3.1 8B, BF16 inference | 19.2 GB | 1 GPU | 1 GPU |
| Llama 3.1 8B, LoRA fine-tune (planning) | 19.2 GB | 1 GPU | 1 GPU |
| Llama 3.1 8B, full fine-tune | 128–144 GB + activations | 2+ GPUs | 2+ GPUs |
| Llama 3.1 70B, 4-bit inference | 42 GB | 1 GPU | 1 GPU |
| Llama 3.1 70B, QLoRA fine-tune | 52.5 GB | 1 GPU | 1 GPU |
| Llama 3.1 70B, FP8 inference | 84 GB | 2 GPUs | 2 GPUs, no FP8 speed-up |
| Qwen2.5 72B, FP8 inference | 87.2 GB | 2 GPUs | 2 GPUs, no FP8 speed-up |
| Llama 3.1 70B, BF16 inference | 168 GB | 3 GPUs | 3 GPUs |
| Llama 3.1 405B, FP8 inference | 486 GB | 7 GPUs | 7 GPUs, no FP8 speed-up |
Naive fit = the number of 80 GB cards whose total memory holds the estimate; it ignores tensor-parallel overhead and memory reserved by the driver. LoRA planning = 2 bytes per parameter × 1.2. QLoRA ≈ 0.75 GB per billion parameters, calibrated on the QLoRA paper's 65B model on one 48 GB GPU. Full fine-tune = 16–18 bytes per parameter (EleutherAI, Hugging Face).
How much do the H100 and A100 cost per hour?
Per GPU-hour, USD. 'Lowest' and 'median' are self-serve on-demand list prices from GPUwithCrypto's price check of 29 October 2026; spot and marketplace prices are excluded from both. Vast.ai is a marketplace, so its figure is shown apart. Every GPU price is on the pricing page.
| GPU | GPUwithCrypto | Lowest on-demand elsewhere | Median on-demand | Providers checked | Vast.ai lowest offer |
|---|---|---|---|---|---|
| H100 80GB | $2.32/hour | $2.73 (Massed Compute) | $4.16 | 16 | $1.73 (H100 SXM) |
| A100 80GB | $0.92/hour | $1.09 (Thunder Compute) | $2.23 | 13 | $0.48 (SXM), $0.34 (PCIe) |
Prices checked on
Massed Compute and Thunder Compute do not state the form factor of these cards. Vast.ai figures: Vast.ai's public price feed, cheapest rentable offer on 29 October 2026. Taxes excluded.
What does a GB of VRAM cost per hour on each?
According to GPUwithCrypto's price check of 29 October 2026, a GB of A100 memory costs 2.5 times less per hour than a GB of H100 memory. Computed as hourly price ÷ 80 GB:
- A100 80GB: 1.36 cents per GB-hour at the lowest on-demand price ($1.09 ÷ 80), 2.79 cents at the median ($2.23 ÷ 80).
- H100 80GB: 3.41 cents per GB-hour at the lowest price ($2.73 ÷ 80), 5.20 cents at the median ($4.16 ÷ 80).
At the median the gap shrinks to 1.87 times. At GPUwithCrypto the same division gives $0.012 and $0.029, each GPU priced at least 15% under its named on-demand reference.
Per dollar at the lowest prices, the H100 SXM gives 27% more peak BF16 math (362 vs 286 TFLOPS per $/hour). The A100 SXM gives 52% more bandwidth per dollar (1,871 vs 1,227 GB/s per $/hour). If your job needs memory more than math, the A100 is the cheaper 80 GB.
What does one job cost on an H100 vs an A100?
The H100 is cheaper per job only when it runs the job faster than its price premium. That break-even is 2.50 times at the lowest prices we found ($2.73 ÷ $1.09) and 1.87 times at the median ($4.16 ÷ $2.23). Both GPUwithCrypto prices sit at least 15% under their named references (Massed Compute at $2.73, Thunder Compute at $1.09), so our break-even is also about 2.5 times.
Worked example (assumption, not a benchmark): a fine-tuning job takes 24 hours on one A100 80GB. On an H100 SXM it takes 7.6 hours if it gains the full 3.17× BF16 peak ratio, or 14.6 hours if it only gains the 1.64× bandwidth ratio.
Cost of the same job (24 A100-hours assumed)
| Price basis | A100 80GB, 24 h | H100, 7.6 h (3.17× faster) | H100, 14.6 h (1.64× faster) |
|---|---|---|---|
| GPUwithCrypto | $22.08 | $17.63 | $33.87 |
| Lowest on-demand elsewhere | $26.16 | $20.75 | $39.86 |
| Median on-demand | $53.52 | $31.62 | $60.74 |
Prices checked on
Hours are an assumption; speed ratios are NVIDIA SXM peak figures. GPUwithCrypto bills per minute, so a 7.6-hour run is billed 7.6 hours.
When should I rent the H100 instead of the A100?
Rent the H100 when at least one of these is true:
- You train or fine-tune at large batch. Compute-bound work tracks the 3.17× peak ratio more closely, and that ratio is above the 2.5× break-even.
- You serve FP8 models. Only the H100 has FP8 Tensor Cores (1,979 TFLOPS dense).
- You need several GPUs talking fast. H100 SXM NVLink runs at 900 GB/s, against 600 GB/s on the A100.
- A deadline matters more than the bill.
Rent the A100 80GB when the model fits in 80 GB and the work is memory-bound: chat inference at small batch, LoRA and QLoRA runs, evaluations, notebooks. If 80 GB is too small, read H100 vs H200. For a cheaper 48 GB card, read L40S vs A100.
Questions about the H100 and the A100
Is the H100 worth it over the A100?
Only when it runs your job more than about 2.5 times faster, the ratio of the lowest hourly prices on 29 October 2026. NVIDIA's peak figures give the H100 SXM 3.17 times the dense BF16 math and 1.64 times the bandwidth. Compute-bound training clears that bar at peak; bandwidth-bound inference does not.
How much does an H100 cost per hour compared with an A100?
On 29 October 2026 the lowest self-serve on-demand price we found was $2.73 for an H100 80GB and $1.09 for an A100 80GB; the medians were $4.16 and $2.23. GPUwithCrypto charges $2.32/hour and $0.92/hour, paid in crypto.
Can an A100 run FP8 models?
The memory math is the same, but NVIDIA lists FP8 as not available on Ampere, so the A100 has no FP8 Tensor Core speed-up. The H100 runs FP8 at 1,979 dense TFLOPS, half of NVIDIA's 3,958 sparse figure.
Can I run Llama 3.1 70B on one H100 or A100?
Only in 4-bit: about 42 GB with the × 1.2 rule, which fits one 80 GB card. In FP8 it needs about 84 GB (two cards) and in BF16 about 168 GB (three cards). One H200 with 141 GB holds the FP8 version.
Can I rent an H100 or A100 with crypto and no KYC?
Yes. GPUwithCrypto rents both by the minute, paid in BTC, USDT, USDC, ETH, SOL, LTC, with no ID, no selfie and no card. Of the 33 GPU providers we checked on 29 October 2026, 14 accept crypto, and RunPod asks for KYC before the first crypto payment.
Rent an H100 or an A100 with crypto
H100 $2.32/hour · A100 80GB $0.92/hour · billed per minute · no KYC.
Sources (12)
- NVIDIA H100 Tensor Core GPU — NVIDIA (read 2026-10-29)
- NVIDIA H100 PCIe product brief (PB-11133-001) — NVIDIA (read 2026-10-29)
- NVIDIA announces Hopper architecture (HBM3) — NVIDIA (read 2026-10-29)
- NVIDIA A100 Tensor Core GPU — NVIDIA (read 2026-10-29)
- NVIDIA A100 datasheet — NVIDIA (read 2026-10-29)
- Transformer Math 101 — EleutherAI (read 2026-10-29)
- Model memory anatomy — Hugging Face (read 2026-10-29)
- QLoRA: Efficient Finetuning of Quantized LLMs — arXiv (read 2026-10-29)
- Llama 3.1 70B model card — Meta / Hugging Face (read 2026-10-29)
- Massed Compute pricing vm.massedcompute.com/pricing — Massed Compute (read 2026-10-29)
- Thunder Compute pricing thundercompute.com/pricing — Thunder Compute (read 2026-10-29)
- Vast.ai public GPU pricing feed — Vast.ai (read 2026-10-29)