GPU comparison
MI300X vs H200: Specs, monthly price and which to rent
MI300X carries 192 GB of HBM3 at 5.3 TB/s against 141 GB at 4.8 TB/s on the H200, and 2.6 PF of dense FP8 against 1.98 PF, for about 35% less per month. The H200 wins when you depend on CUDA-only kernels, TensorRT-LLM or recipes tuned for Hopper; for serving and fine-tuning large models on ROCm, MI300X gives more memory and more compute per dollar.
- Updated Nov 4, 2026
- Specs: vendor datasheets, dense Tensor figures
- Prices: live CoreGPU catalog
| Spec | AMD Instinct MI300X | NVIDIA H200 SXM |
|---|---|---|
| Architecture | CDNA 3 | Hopper |
| GPU memory | 192 GB HBM3 | 141 GB HBM3e |
| Memory bandwidth | 5.3 TB/s | 4.8 TB/s |
| FP8 dense | 2.6 PF | 1.98 PF |
| FP4 dense | n/a | n/a |
| Interconnect | Infinity Fabric · 896 GB/s | NVLink 4 · 900 GB/s |
| Power (TDP) | 750 W | 700 W |
| Price per GPU / month | $1,749 | $2,699 |
| Price per GB of memory | $9.11 | $19.14 |
| Price per FP8 PFLOPS | $673 | $1,363 |
| Availability | 41 available | 34 available |
Green marks the better value on each line. Prices are per GPU per month at CoreGPU, at least 25% below the market median.
- Serving 70B to 180B models with fewer GPUs
- Memory-bound decoding at 5.3 TB/s
- ROCm stacks: PyTorch, vLLM, SGLang
- CUDA-only kernels and TensorRT-LLM
- Recipes tuned for Hopper
- Teams standardised on NVIDIA tooling
MI300X vs H200: monthly price by server size
| Server | MI300X | H200 | Difference |
|---|---|---|---|
| 1 GPU | $1,749/mo | $2,699/mo | $950 less on the MI300X |
| 2 GPUs | $3,498/mo | $5,398/mo | $1,900 less on the MI300X |
| 4 GPUs | $6,996/mo | $10,796/mo | $3,800 less on the MI300X |
| 8 GPUs | $13,992/mo | $21,592/mo | $7,600 less on the MI300X |
Flat monthly price per GPU, the same at every server size. USD, taxes excluded.
Largest model that fits in GPU memory
| Weights | 1× MI300X | 1× H200 | 8× MI300X | 8× H200 |
|---|---|---|---|---|
| BF16 | 80B | 58B | 640B | 470B |
| FP8 | 160B | 117B | 1.3T | 940B |
| FP4 | 320B | 235B | 2.6T | 1.9T |
Parameters whose weights fit with 20% headroom for activations and runtime. Long contexts and large batches need more room for the KV cache: see how much GPU memory you need.
More GPU comparisons
MI300X vs H200: common questions
Already a customer? Open a ticket from the console: engineers reply within one business day.
How support worksIs the MI300X faster than the H200?
On dense FP8 Tensor throughput the MI300X is faster: 2.6 PF against 1.98 PF per GPU. Memory bandwidth is 5.3 TB/s on the MI300X and 4.8 TB/s on the H200, which matters most for LLM decoding.
How much does the MI300X cost per month compared with the H200?
At CoreGPU the MI300X is $1,749 and the H200 is $2,699 per GPU per month, so an 8-GPU server costs $13,992 or $21,592 a month. The MI300X is 35% cheaper per GPU. Both are at least 25% below the market median, billed monthly from a prepaid balance.
Which is better for LLM inference, the MI300X or the H200?
For large models and long contexts the MI300X has the edge with 192 GB of memory per GPU, so a model needs fewer GPUs. Choose the MI300X for serving 70B to 180B models with fewer GPUs; the H200 for cUDA-only kernels and TensorRT-LLM.
Can I rent the MI300X and the H200 by the month?
Yes. Both are dedicated GPUs you rent by the month, from 1 to 8 per server, with root SSH access. Right now the MI300X is 41 available and the H200 is 34 available. You pay in BTC, ETH, XMR, USDT (TRC-20) or SOL.