GH200 vs MI325X
GH200 starts at $2.29/GPU-hour on-demand; MI325X starts at $3.80/GPU-hour on-demand. Compare the exact variants, memory and published specifications below. Peak compute figures do not establish application speed.
GH200
MI325X
Spec sheet
| GH200 | MI325X | |
|---|---|---|
| Spec variant | GH200 | MI325X |
| Architecture | Hopper | CDNA 3 |
| VRAM (spec variant) | 96 GB | 256 GB |
| FP16 TFLOPS (dense) | 989 | 1307 |
| Memory bandwidth | 4000 GB/s | 6000 GB/s |
| TDP | 700 W | 1000 W |
| Interconnect | NVLink-C2C 900GB/s | Infinity Fabric |
| Launched | 2023 | 2024 |
Rental price per manufacturer-rated FP16 compute: GH200 $2.32 per PFLOP-hour vs MI325X $2.91. Formula: cheapest $/hr ÷ dense FP16 PFLOPS (1 PFLOP = 1000 TFLOPS). This uses GH200 and MI325X, whose datasheets may differ from the variants above. It is not measured job throughput. Specs: GH200 manufacturer, MI325X manufacturer.
What they cost to rent (2026-10-02)
| GH200 | MI325X | |
|---|---|---|
| Cheapest on-demand $/hr | $2.29 | $3.80 |
| Cheapest spot/community $/hr | – | $2.00 |
| Cheapest source | Lambda | DigitalOcean |
| Providers offering it | 3 | 2 |
Lowest recorded on-demand rate per configuration; highlighted = cheaper. Spot/marketplace tiers shown on the individual pages.
FAQ
GH200 is cheaper: the GH200 rents from $2.29 per GPU-hour and the MI325X from $3.80 on-demand, updated 2026-10-02.
The largest recorded GH200 variant has 96 GB and the largest MI325X variant has 256 GB. Other variants may have less memory. The MI325X has more.
Compare more rental options: the screener · price movers · methodology.