AMD MI300X
CDNA 3 architecture · 192GB memory · 1300 FP16 TFLOPS · 750W TDP
Cloud Pricing Today
About the AMD MI300X
The AMD Instinct MI300X is AMD's most competitive entry in the data centre AI accelerator market. Based on the CDNA 3 architecture, it offers 1,300 FP16 TFLOPS and 192GB of HBM3 memory — significantly more memory than the NVIDIA H100's 80GB.
The MI300X's memory advantage is its strongest differentiator. With 192GB of HBM3 providing 5.3 TB/s of bandwidth, it can serve larger language models on fewer GPUs than the H100, potentially reducing total cost of ownership for inference workloads. This memory capacity puts it on par with NVIDIA's newer B200 and GB200.
The MI300X uses AMD's ROCm software stack, which has made significant progress in supporting major AI frameworks including PyTorch and JAX. However, the CUDA ecosystem remains more mature, and some workloads may require additional optimisation effort to run efficiently on AMD hardware.
In the cloud market, the MI300X is increasingly available as providers diversify their GPU offerings. It is particularly attractive for organisations looking to reduce dependency on a single vendor or seeking better price/performance for memory-intensive workloads.
Common Use Cases
Key Facts
- Manufacturer
- AMD
- Architecture
- CDNA 3
- Accelerator Type
- GPU
- Primary Use
- training
- Memory (VRAM)
- 192 GB
- FP16 Performance
- 1300 TFLOPS
- Thermal Design Power
- 750W
Frequently Asked Questions
How much does an MI300X cost per hour?
The AMD MI300X blended cloud pricing typically ranges from $2.50–$5.00 per hour, often competitive with or lower than NVIDIA H100 pricing for comparable performance.
Can I run PyTorch on the MI300X?
Yes. The MI300X supports PyTorch through AMD's ROCm software stack. Most PyTorch models can run on MI300X, though some CUDA-specific operations may require adaptation. Major frameworks have increasingly strong ROCm support.
MI300X vs H100 — which is better?
The MI300X offers 31% more FP16 TFLOPS (1,300 vs 990) and 140% more memory (192GB vs 80GB). The H100 has a more mature CUDA ecosystem and wider availability. For memory-bound workloads, the MI300X can offer better total cost of ownership.
Related Accelerators
Compare AMD MI300X
MI300X offers 31% more TFLOPS (1,300 vs 990) and 140% more memory (192GB vs 80GB HBM3). H100 has the more mature CUDA ecosystem. MI300X can be more cost-effective for memory-intensive workloads.
MI300X delivers 4.2x the performance (1,300 vs 312 TFLOPS) with 2.4x the memory (192GB vs 80GB). MI300X is a generational leap but costs significantly more per hour.
Calculate AMD MI300X ROI
Estimate payback period, annual returns, and 3-year ROI with live Signwl pricing data.
Track AMD MI300X pricing over time
Get access to historical pricing data, regional analysis, and custom alerts.