Compare · prices checked 2026-09-03
H100 SXM vs A100 SXM4: specs, price per hour, which to rent
NVIDIA H100 SXM (80 GB, $1.587/hr) against NVIDIA A100 SXM4 (80 GB, $0.583/hr): public specs side by side, live fixed prices from our sheet, what fits in each card's VRAM, and a verdict written for real workloads — both are rentable right now.
H100 SXM vs A100 SXM4 specifications
Public NVIDIA figures (dense, non-sparsity). The last column is H100 SXM relative to A100 SXM4.
| Spec | H100 SXM | A100 SXM4 | Difference |
|---|---|---|---|
| Architecture | Hopper (2022) | Ampere (2020) | — |
| VRAM | 80 GB HBM3 | 80 GB HBM2e | same |
| Memory bandwidth | 3,350 GB/s | 2,039 GB/s | +64% |
| FP16 tensor (dense) | 990 TFLOPS | 312 TFLOPS | +217% |
| FP32 | 67.0 TFLOPS | 19.5 TFLOPS | +244% |
| CUDA cores | 16,896 | 6,912 | +144% |
| TDP | 700 W | 400 W | +75% |
| PCIe · NVLink | Gen 5.0 · NVLink | Gen 4.0 · NVLink | — |
| PowerScore (RTX 3090 = 100) | 697 | 220 | +217% |
| Max GPUs per machine | 8× | 8× | — |
H100 SXM vs A100 SXM4 price per hour
Fixed rates from our sheet — every on-demand price is the marketplace median × 0.70, rounded down. Monthly = 730 hours.
| Rate | H100 SXM | A100 SXM4 | Cheaper |
|---|---|---|---|
| On-demand, per GPU-hour | $1.587 | $0.583 | A100 SXM4 (−63%) |
| Interruptible, per GPU-hour | $0.793 | $0.291 | A100 SXM4 |
| Reserved (3 mo), per GPU-hour | $1.031 | $0.378 | A100 SXM4 |
| On-demand, per month | $1,159 | $426 | A100 SXM4 |
| Market median (reference) | $2.27 | $0.83 | — |
| $ per 1,000 FP16 TFLOP-hours | $1.60 | $1.87 | H100 SXM (better value) |
| $ per GB of VRAM per hour | $0.0198 | $0.0073 | A100 SXM4 (better value) |
Try a full month with storage and bandwidth in the GPU cost calculator.
What fits in VRAM: 80 GB vs 80 GB
| Workload | H100 SXM | A100 SXM4 |
|---|---|---|
| Largest LLM in FP16, one card | ~32B | ~32B |
| Largest LLM at 4-bit, one card | ~123B | ~123B |
| Flux dev (FP8, ~17 GB) | fits | fits |
| Wan 2.x 14B video (offloaded) | yes | yes |
| 70B 4-bit LLM on one card | yes | yes |
Rules of thumb: ~2.4 GB per billion parameters in FP16 all-in, ~0.62 GB in 4-bit. Full tables in the VRAM guide.
Verdict: which should you rent?
The H100 SXM trains transformers roughly 2–3× faster than the A100 SXM4 thanks to FP8 and 3.35 TB/s of HBM3; the A100 costs far less per hour and still wins dollars per step for ≤13B models, LoRA farms and FP64 HPC. Rent the H100 for deadlines and FP8 serving, the A100 for volume experiments.
- Cheaper per hour: A100 SXM4 ($0.583 vs $1.587, −63%).
- More VRAM: H100 SXM (80 GB vs 80 GB).
- More FP16 throughput: H100 SXM (about 3.2×).
- Best value per TFLOP-hour: H100 SXM.
- Best value per GB of VRAM: A100 SXM4.
- Multi-GPU: H100 SXM with NVLink · A100 SXM4 with NVLink.
Related comparisons
- H100 SXM vs H200 $1.587 vs $3.058 per hour Compare
- H100 SXM vs B200 $1.587 vs $4.204 per hour Compare
- H100 SXM vs H100 PCIE $1.587 vs $2.147 per hour Compare
- A100 SXM4 vs A100 PCIE $0.583 vs $0.662 per hour Compare
- RTX 5090 vs A100 SXM4 $0.318 vs $0.583 per hour Compare
- H100 NVL vs H100 SXM $1.877 vs $1.587 per hour Compare
H100 SXM vs A100 SXM4: FAQ
Deeper reading: H100 vs H200 vs B200, RTX 4090 vs RTX 5090, how cloud GPU pricing works.
Is the H100 SXM faster than the A100 SXM4?
On dense FP16 tensor throughput the H100 SXM leads by about 3.2× (990 vs 312 TFLOPS). Memory bandwidth matters as much for inference: H100 SXM 3,350 GB/s vs A100 SXM4 2,039 GB/s.
Which is cheaper to rent, the H100 SXM or the A100 SXM4?
The A100 SXM4: $0.583/hr on-demand versus $1.587/hr — 63% less. Interruptible rates are $0.793 (H100 SXM) and $0.291 (A100 SXM4). Per TFLOP-hour the better value is the H100 SXM.
Which has more VRAM and what does that change?
The H100 SXM has 80 GB versus 80 GB. In LLM terms that is roughly a 32B FP16 model (or ~123B in 4-bit) on one card against 32B FP16 (~123B 4-bit). If the model does not fit, speed is irrelevant.
Can I rent both on PowerGPU right now?
Yes — 69 × H100 SXM and 43 × A100 SXM4 are online as this page renders, deployable in about 30 seconds, billed per second, paid in crypto with no KYC.