GPU models
8 GPUs with specs from the vendor datasheets and the lowest reviewed cloud rate for each. Open a GPU to compare every offer like for like.
8 of 8 GPUsRates reviewed 7 Nov 2026 · Specs checked 8 Nov 2026
| GPU | Memory | Max bandwidth | FP8 dense | From / GPU-hour | Rates | Sellers |
|---|---|---|---|---|---|---|
| NVIDIA L40SAda Lovelace · 1 variant | 48 GB | 0.86 TB/s | 733 TFLOPS | $1.09Runpod | 2 | 6 |
| NVIDIA A100Ampere · 2 variants | 80 GB | 2.04 TB/s | Not published | $1.79Runpod | 1 | 5 |
| NVIDIA H100Hopper · 3 variants | 80–94 GB | 3.9 TB/s | 1,978.9 TFLOPS | $2.89Runpod | 5 | 8 |
| NVIDIA H200Hopper · 2 variants | 141 GB | 4.8 TB/s | 1,979 TFLOPS | $4.47DigitalOcean | 4 | 7 |
| NVIDIA B200Blackwell · 1 variant | 180 GB | 7.7 TB/s | 4,500 TFLOPS | $7.99Runpod | 3 | 6 |
| AMD Instinct MI300XCDNA 3 · 1 variant | 192 GB | 5.3 TB/s | 2,614.9 TFLOPS | No reviewed rate | 0 | 2 |
| NVIDIA B300 (Blackwell Ultra)Blackwell Ultra · 1 variant | 270 GB | 7.7 TB/s | 4,500 TFLOPS | No reviewed rate | 0 | 6 |
| NVIDIA GeForce RTX 4090Ada Lovelace · 1 variant | 24 GB | 1.01 TB/s | 330.3 TFLOPS | No reviewed rate | 0 | 2 |
Rates are on-demand examples from providers' public price pages, converted to USD per GPU-hour. They are not live quotes, and stock is not verified. Specs come from each vendor's datasheet; compute figures are dense, without sparsity. A100 has no FP8 support.