…
Cloud rental price, compute per dollar and specs, side by side.
Open in the interactive comparer →| Spec / price | H200 SXM | Instinct MI300X |
|---|---|---|
| Memory | 141 GB HBM3e | 192 GB HBM3better |
| Memory bandwidth | 4,800 GB/s | 5,300 GB/sbetter |
| FP16 / BF16 (dense) | 0.989 PF | 1.31 PFbetter |
| FP8 (dense) | 1.98 PF | 2.61 PFbetter |
| FP4 (dense) | — | — |
| INT8 (dense) | 1.98 PF | 2.61 PFbetter |
| TF32 (dense) | 0.495 PF | 0.654 PFbetter |
| FP32 (dense) | 0.067 PF | 0.163 PFbetter |
| FP64 (dense) | 0.067 PF | 0.163 PFbetter |
| Power | 700 W | 750 W |
| Architecture | Hopper (2024) | CDNA 3 (2023) |
| Lowest on-demand | $2.10 | $1.71better |
| Median on-demand | $3.99 | $2.99better |
| Lowest spot | $0.79better | $1.11 |
| FP16/BF16 $ per PFLOP·h | $2.12 | $1.31better |
| Providers | 22 | 10 |
The Instinct MI300X has 1.3× more dense FP16/BF16 compute than the H200 SXM. The Instinct MI300X has 1.1× more memory bandwidth than the H200 SXM. Real-world speedups depend on the workload: training and prefill scale with compute, while LLM decoding scales mostly with memory bandwidth.
Right now the Instinct MI300X is cheaper per unit of compute: $1.31 vs $2.12 per PFLOP-hour at the lowest on-demand prices.
Running one GPU around the clock at the lowest on-demand price costs about $1,533 per month for the H200 SXM and $1,248 for the Instinct MI300X.
The Instinct MI300X has more memory: 192 GB vs 141 GB.