RentServers.comBy the RentServers.com team · Updated
MI300X specifications
| Architecture | AMD CDNA 3 |
|---|---|
| Memory | 192 GB HBM3 |
| Memory bandwidth | 5.3 TB/s |
| FP8 (dense) | ~2,615 TFLOPS |
| BF16 (dense) | ~1,307 TFLOPS |
| Interconnect | Infinity Fabric, 896 GB/s peer |
| Max power | 750 W |
Figures for the OAM version unless noted, from AMD. Dense throughput; vendors' headline numbers with sparsity are double.
Where the MI300X fits
Memory-bound inference is the strongest case: a 70B model in FP16 fits on one MI300X with room for KV cache, where Hopper needs two GPUs. Serving stacks such as vLLM support ROCm, so many inference deployments move over with configuration changes rather than code changes.
Training works too, but check your framework, kernels and collective-communication library versions on ROCm before committing to a long term.
Pricing and supply
Fewer providers list the MI300X: GetDeploying tracked 4 (October 28, 2026), with a reserved median of $2.52 and an on-demand median of $4.42 per GPU-hour. Thin supply means medians swing more than for H100, so ask for several quotes.
What to check in an MI300X quote
- ROCm version on the provided image and how upgrades are handled.
- Fabric between nodes for multi-node jobs.
- Whether the provider has run your framework on MI300X before.
Rent or buy an MI300X?
There is no dated public purchase price for an 8x MI300X (OAM) system; it is quoted per deal. Renting 8 GPUs costs about $14,717 a month reserved.
No dated public purchase price found. Quoted per deal. The full comparison, including what owning really costs, is in rent vs buy GPUs; to buy outright, use our wholesale desk.
Bare metal or cloud for MI300X
MI300X runs PyTorch and vLLM through AMD's ROCm stack. Fewer providers list it, so prices swing more, but its 192 GB per GPU fits large models on fewer GPUs.
| System | Purchase price | Owned, per month | Rent reserved, per month | Rent on demand 24/7, per month | Breakeven use vs on demand | Payback vs on demand |
|---|---|---|---|---|---|---|
| 8x MI300X (OAM) | Quoted per deal | n/a | $14,717 | $25,813 | n/a | n/a |
Rental: provider medians for 8 GPUs x 730 hours, week of October 19, 2026, GetDeploying, GPU rental price history, CC BY 4.0. Purchase: estimates from public reseller price lists and market reports, dated per row; no dealer quote. Owned per month: purchase price over 36 months plus hosting at about $250 per kW per month all-in (published 2026 retail colocation pricing, an estimate) at the system's maximum power. Excludes network, spares, staff, financing and resale value.
Typical MI300X server specifications
| Typical system | 8x MI300X (OAM) |
|---|---|
| Memory per GPU | 192 GB HBM3 |
| System GPU memory | 1,536 GB |
| Memory bandwidth per GPU | 5.3 TB/s |
| GPU interconnect | Infinity Fabric, 896 GB/s peer bandwidth per GPU |
| GPU power | up to 750 W |
| System power | about 10 kW (Estimate: 8 x 750 W GPUs plus host CPUs, memory and fans) |
From AMD. Model names are trademarks of their owners.
MI300X price by term
| Term | Provider median, $/GPU-hr |
|---|---|
| On demand | $4.42 |
| 12 months | $2.52 |
Week of October 19, 2026, 4 providers on demand. Source: GetDeploying, GPU rental price history, CC BY 4.0.
Where to run it
Power and cooling decide where an MI300X cluster can live. Local power, density and quotes for 25 markets:
All 25 cities · Weekly GPU price tracker
MI300X questions
How much does it cost to rent an MI300X?
On GetDeploying's tracker (October 28, 2026): median $2.52 per GPU-hour reserved and $4.42 on demand, from 4 providers.
Does MI300X run CUDA code?
Not directly. It runs ROCm. PyTorch, vLLM and many popular libraries support ROCm, but custom CUDA kernels need porting, often with AMD's HIP tools.