MI300X vs H200
Manufacturer specifications of the MI300X and the H200, side by side, per GPU. A value the manufacturer does not publish for the exact model is marked as such, never estimated.
AMD Instinct MI300X 192GB
CDNA 3
- GPU memory
- 192 GB
- Memory bandwidth
- 5,300 GB/s
Live price on EverXCloud · from $4.16 per GPU-hour
Suited workloads
NVIDIA H200 141GB
Hopper
- GPU memory
- 141 GB
- Memory bandwidth
- 4,800 GB/s
Live price on EverXCloud · from $4.19 per GPU-hour
Suited workloads
Key differences
- The MI300X has 1.4× the GPU memory of the H200 (192 GB vs 141 GB).
- The MI300X has 1.1× the memory bandwidth of the H200 (5,300 vs 4,800 GB/s), which usually sets the pace of LLM token generation.
- The MI300X has 1.3× the dense BF16 tensor throughput of the H200 (1,307 vs 989 TFLOPS), the precision of most training runs.
- The MI300X has 1.3× the dense FP8 tensor throughput of the H200 (2,614 vs 1,979 TFLOPS).
Specifications
| Specification | MI300X | H200 |
|---|---|---|
| GPU memory | 192 GB (higher) | 141 GB |
| Memory type | HBM3 | HBM3e |
| Memory bandwidth | 5,300 GB/s (higher) | 4,800 GB/s |
| FP16 tensor | 1,307 TFLOPS (higher) | 989 TFLOPS |
| BF16 tensor | 1,307 TFLOPS (higher) | 989 TFLOPS |
| FP8 tensor | 2,614 TFLOPS (higher) | 1,979 TFLOPS |
| Max power | 750 W | 700 W |
| Architecture | CDNA 3 | Hopper |
| Interconnect | AMD Infinity Fabric (128 GB/s per link) · PCIe 5.0 | NVLink 4 (900 GB/s) · PCIe 5.0 |
The higher value of each measured row is shown in bold. Manufacturer specifications, per GPU. Tensor throughput is shown without sparsity. Values the manufacturer does not publish for this exact model are left out.
What fits in memory, on one GPU
| Workload | MI300X | H200 |
|---|---|---|
| LLM inference, 16-bit | ≈ 75B parameters | ≈ 55B parameters |
| LLM inference, 4-bit | ≈ 250B parameters | ≈ 180B parameters |
| Full training, mixed precision | ≈ 9B parameters | ≈ 7B parameters |
Cautious estimates counting 80% of memory for model weights; context length, batch size and framework change the real limit. Several GPUs together hold larger models.
MI300X vs H200: frequently asked questions
Which has more memory, the MI300X or the H200?
The MI300X has 192 GB of GPU memory, against 141 GB for the H200.
Which is faster for LLM inference, the MI300X or the H200?
Generating tokens is usually bound by memory bandwidth, and the MI300X has 1.1× the bandwidth of the H200. Memory size then decides which models fit. Measure your own model before a large run.
Can I rent the MI300X and the H200 by the hour?
Yes, when sellers list them on the EverXCloud spot exchange: you buy GPU hours at the live price from a prepaid balance in euros, US dollars or pounds. Prices are shown excluding tax and EverXCloud is the merchant of record for every rental.
Be ready on launch day
Create your account now: you can top up your balance and rent the GPU that fits your workload as soon as the market opens.
Have GPU servers?
Businesses can join the EverXCloud Network and sell hours of their GPU servers on the spot exchange.