# A100 vs H100: Speed, Memory and Cost per Job

> A100 or H100 for training and inference? Compare memory, bandwidth, FP8 support and real rental prices, and work out when the faster H100 is cheaper per job.

- Canonical URL: https://computesuppliers.com/a100-vs-h100/
- Source: [NVIDIA H100 and A100 specifications](https://www.nvidia.com/en-us/data-center/h100/)
- Checked: 2026-10-26
- Structured data: https://computesuppliers.com/data/a100-vs-h100.json

## The specs that matter

| | A100 80 GB (SXM) | H100 80 GB (SXM) |
|---|---|---|
| Memory | 80 GB HBM2e | 80 GB HBM3 |
| Memory bandwidth | about 2.0 TB/s | about 3.35 TB/s |
| FP8 (Transformer Engine) | no | yes |
| Lowest on-demand we found | $1.35 (PCIe), $1.60 (SXM) | $2.50 (PCIe), $3.20 (SXM) |

Specs: [NVIDIA A100](https://www.nvidia.com/en-us/data-center/a100/), [NVIDIA H100](https://www.nvidia.com/en-us/data-center/h100/). Prices: [Hyperstack](https://www.hyperstack.cloud/gpu-pricing), 26 October 2026.

## Cost per job, not per hour

The H100 costs about twice as much per hour. It's cheaper per job whenever it finishes the job in less than half the time.

| If the H100 is… | Cost of the job on H100 vs A100 (at $3.20 vs $1.60) |
|---|---|
| 1.5× faster | 33% more expensive |
| 2× faster | the same |
| 3× faster | 33% cheaper |

How much faster depends on the workload. Memory-bound LLM inference gains roughly in line with bandwidth (about 1.7×). Training that can use FP8 gains more. Benchmarks vary widely, so time a short run on both before a long job.

## Which to choose

- **A100:** budget fine-tuning, LoRA and QLoRA, inference that's not latency-critical, anything that fits in 80 GB and doesn't use FP8.
- **H100:** large training runs, FP8 training and inference, multi-GPU jobs where NVLink speed matters, tight deadlines.

Price both with your hours in the [job cost calculator](https://computesuppliers.com/gpu-cost-calculator/).

---
Compute Suppliers · Contact: contact@computesuppliers.com