AMD Instinct MI300X (CDNA 3)
Overview
The AMD Instinct MI300X, launched in early 2024, is based on the CDNA 3 architecture with 3D chiplet packaging. 192GB HBM3 memory (~2.4× that of the H100), delivering 1,307 TFLOPS FP16 compute. It is a direct competitor to the NVIDIA H100 / H200.
Core Specifications
| Item | Specification |
|---|
| Architecture | CDNA 3 |
| Process Node | TSMC 5nm + 6nm (hybrid) |
| Transistor Count | 146 billion |
| Memory | 192 GB HBM3 |
| Memory Bandwidth | 5,300 GB/s |
| FP32 | 65.3 TFLOPS |
| FP64 | 32.6 TFLOPS |
| FP16/BF16 Matrix | 1,307 TFLOPS |
| FP8 Matrix | 2,614 TFLOPS |
| INT8 Matrix | 2,614 TOPS |
| TDP | 750 W |
| Interconnect | Infinity Fabric 4.1 (448 GB/s), PCIe 5.0 |
| Form Factor | OAM (UBB 8-card) |
MI300X vs H100 / H200
| Metric | MI300X | H100 | H200 |
|---|
| Memory | 192 GB | 80 GB | 141 GB |
| Memory Type | HBM3 | HBM3 | HBM3e |
| Bandwidth | 5.3 TB/s | 3.35 TB/s | 4.8 TB/s |
| FP16 Compute | 1,307 TFLOPS | 989 TFLOPS | 989 TFLOPS |
| FP8 Compute | 2,614 TFLOPS | 3,958 TFLOPS | 3,958 TFLOPS |
| TDP | 750 W | 700 W | 700 W |
| Price | ~$15,000 | ~$30,000 | ~$33,000 |
Key advantage: Large memory, lower price, significant ROCm 7 performance improvements.
Use Cases
- LLM training and inference (192GB memory advantage)
- MoE models (70B+ fits on a single card)
- HPC and AI mixed workloads
- Cost-sensitive customers