Compute- and memory-per-dollar across 23 GPUs we price, using the cheapest live on-demand rate from each — a value view no single provider publishes. Updated July 2026.
Best value by compute (FP16 TFLOPS per $/hr)
Higher = more raw FP16 compute for every dollar-hour, at the cheapest on-demand price we track.
| # | GPU | FP16 | Cheapest $/hr | TFLOPS / $·hr |
|---|---|---|---|---|
| 1 | NVIDIA GeForce RTX 5090 3 prov. | 419 TFLOPS | $0.34 | 1,232 |
| 2 | NVIDIA GeForce RTX 3090 3 prov. | 142 TFLOPS | $0.12 | 1,183 |
| 3 | NVIDIA GeForce RTX 4090 3 prov. | 330 TFLOPS | $0.28 | 1,179 |
| 4 | NVIDIA RTX 6000 Ada 5 prov. | 362 TFLOPS | $0.47 | 770 |
| 5 | NVIDIA L40S 7 prov. | 366 TFLOPS | $0.55 | 665 |
| 6 | NVIDIA H100 SXM 7 prov. | 990 TFLOPS | $1.57 | 631 |
| 7 | NVIDIA RTX A6000 3 prov. | 155 TFLOPS | $0.30 | 517 |
| 8 | NVIDIA H200 SXM 3 prov. | 990 TFLOPS | $2.05 | 483 |
| 9 | NVIDIA A100 SXM (40GB) 2 prov. | 312 TFLOPS | $0.80 | 390 |
| 10 | NVIDIA L4 1 prov. | 121 TFLOPS | $0.39 | 310 |
| 11 | NVIDIA A100 SXM (80GB) 6 prov. | 312 TFLOPS | $1.10 | 284 |
| 12 | NVIDIA RTX PRO 6000 2 prov. | 252 TFLOPS | $1.71 | 147 |
Best value by memory (GB VRAM per $/hr)
Higher = more GPU memory per dollar-hour — the metric that matters for fitting large models.
| # | GPU | VRAM | Cheapest $/hr | GB / $·hr |
|---|---|---|---|---|
| 1 | NVIDIA GeForce RTX 3090 3 prov. | 24 GB | $0.12 | 200.0 |
| 2 | NVIDIA RTX A6000 3 prov. | 48 GB | $0.30 | 160.0 |
| 3 | NVIDIA RTX 6000 Ada 5 prov. | 48 GB | $0.47 | 102.1 |
| 4 | NVIDIA GeForce RTX 5090 3 prov. | 32 GB | $0.34 | 94.1 |
| 5 | NVIDIA L40S 7 prov. | 48 GB | $0.55 | 87.3 |
| 6 | NVIDIA GeForce RTX 4090 3 prov. | 24 GB | $0.28 | 85.7 |
| 7 | NVIDIA A100 SXM (80GB) 6 prov. | 80 GB | $1.10 | 72.7 |
| 8 | NVIDIA H200 SXM 3 prov. | 141 GB | $2.05 | 68.8 |
| 9 | NVIDIA L4 1 prov. | 24 GB | $0.39 | 61.5 |
| 10 | NVIDIA RTX PRO 6000 2 prov. | 96 GB | $1.71 | 56.1 |
| 11 | NVIDIA H100 SXM 7 prov. | 80 GB | $1.57 | 51.0 |
| 12 | NVIDIA A100 SXM (40GB) 2 prov. | 40 GB | $0.80 | 50.0 |