The GPU database
21 accelerators that matter for AI, with the numbers that decide what they can run, and the cheapest on-demand price across the clouds we track.
Datacenter GPUs
| GPU | VRAM | BW TB/s | FP16 TF | From $/hr | Arch |
|---|---|---|---|---|---|
| MI325X | 256 GB | 6 | 1,307 | $2.40 | CDNA 3 '24 |
| B200 | 192 GB | 8 | 2,250 | $3.75 | Blackwell '24 |
| MI300X | 192 GB | 5.3 | 1,307 | $1.85 | CDNA 3 '23 |
| H200 | 141 GB | 4.8 | 990 | $2.30 | Hopper '24 |
| Gaudi 3 | 128 GB | 3.7 | 918 | $2.30 | Gaudi '24 |
| GH200 | 96 GB | 4 | 990 | $1.35 | Grace Hopper '23 |
| RTX PRO 6000 | 96 GB | 1.79 | 250 | $0.50 | Blackwell '25 |
| H100 SXM | 80 GB | 3.35 | 990 | $1.90 | Hopper '22 |
| H100 PCIe | 80 GB | 2 | 756 | $1.60 | Hopper '22 |
| A100 80GB | 80 GB | 2 | 312 | $0.87 | Ampere '20 |
| L40S | 48 GB | 0.864 | 362 | $0.80 | Ada Lovelace '23 |
| A100 40GB | 40 GB | 1.56 | 312 | $0.48 | Ampere '20 |
| V100 | 32 GB | 0.9 | 125 | $0.21 | Volta '17 |
| A10 | 24 GB | 0.6 | 125 | $0.75 | Ampere '21 |
| L4 | 24 GB | 0.3 | 121 | $0.32 | Ada Lovelace '23 |
| T4 | 16 GB | 0.32 | 65 | $0.10 | Turing '18 |
Consumer cards (marketplace rentals)
AI ASICs
Reading the table: for LLM inference, memory bandwidth (TB/s) predicts tokens/second better than FLOPS, decoding is memory-bound. FLOPS matter for training, prefill, and diffusion models. VRAM decides what fits at all: see the VRAM math.