Question 1

What is the AMD MI300X price per GPU per hour?

Accepted Answer

Runcrate AMD MI300X rental starts at $2.30/hr on-demand (avg $2.50/hr). RunPod lists $3.49, AWS $4.95, Azure $5.20, Oracle $4.30. Runcrate is the lowest published MI300X cloud rental rate.

Question 2

Is the AMD MI300X available on demand?

Accepted Answer

Yes. MI300X is available across 4 regions on Runcrate with reservation-first capacity for predictable supply. Per-second billing, no minimums.

Question 3

AMD MI300X vs NVIDIA H200 — which to pick?

Accepted Answer

MI300X has 36% more VRAM (192GB vs 141GB) and uses ROCm instead of CUDA. For models that need >141GB on a single GPU (Llama 405B, DeepSeek-V3 full precision), MI300X is the only single-GPU option under $3/hr. For everything else and any CUDA-only workload, H200 is more compatible.

Question 4

Does MI300X work with PyTorch and Hugging Face?

Accepted Answer

Yes — ROCm 6 has full PyTorch support, and Hugging Face Transformers, vLLM, TGI, llama.cpp, and most major frameworks have first-class ROCm builds. Runcrate ships pre-baked ROCm images so you can skip the setup.

Question 5

How does MI300X compare to H100?

Accepted Answer

MI300X has 2.4x the VRAM (192 vs 80GB), ~3x the FP32 throughput (163 vs 60 TFLOPS), and 58% more memory bandwidth. The trade-off is ROCm vs CUDA — if your stack runs on PyTorch with stock kernels, MI300X is dramatically more cost-efficient. For custom CUDA kernels or TensorRT, stick with H100.

Question 6

Can I rent multiple MI300X GPUs together?

Accepted Answer

Yes — up to 8 MI300X GPUs in a single OAM node with Infinity Fabric interconnect. Cluster pricing available on a reserved-term contract.

AMD MI300X.

What you'll actually use this for.

Llama 405B Inference

DeepSeek-V3 Serving

Train Mixtral 8x22B

Long-context Inference

Full specifications.

Other AMD Instinct GPUs.

Frequently asked.

Rent the AMD MI300X on Runcrate.