AMD Instinct MI300A (APU)
Overview
AMD Instinct MI300A is an APU-architecture AI training card featuring GPU + CPU integrated packaging and a unified memory architecture similar to Apple's M-Series. Building on the MI300X (pure GPU) foundation, it adds 24 Zen 4 CPU cores sharing a 128GB HBM3 memory pool.
HPC performance monster: 1.5 PFLOPS FP8 / 2.5 PFLOPS FP16. The world's first exascale supercomputer, El Capitan (Lawrence Livermore National Laboratory), uses 44,000+ MI300A units.
Core Specifications
| Item | Spec |
|---|
| Architecture | CDNA 3 + Zen 4 (APU) |
| Process | TSMC 5nm + 6nm Chiplet |
| GPU Stream Processors | 14,592 (228 CUs) |
| CPU Cores | 24 Zen 4 cores (×4 CCD) |
| Unified Memory | 128 GB HBM3 (CPU+GPU shared) |
| Memory Bandwidth | 5.3 TB/s |
| FP16 Compute | 1.5 PFLOPS (dense) / 2.5 PFLOPS (sparse) |
| FP8 Compute | 1.5 PFLOPS (dense) / 2.5 PFLOPS (sparse) |
| INT8 | 1.5 POPS |
| TDP | 600 W |
| Interface | PCIe Gen5 ×16 + Infinity Fabric |
| Interconnect | Infinity Fabric 4 (896 GB/s) |
| Launch | 2024-01 (El Capitan deployment) |
| Price | $15,000-$20,000 (OEM) |
APU Architecture Explained
Unified Memory Advantage
- CPU + GPU share 128GB HBM3 (no data copies needed).
- 5.3 TB/s bandwidth (HBM3e rated 5.3 TB/s).
- Ideal for HPC numerical simulation (CPU handles logic, GPU handles parallel computation).
Chiplet Design
- 3× 5nm SoC chiplets (GPU + I/O)
- 6× 6nm IOD chiplets (memory controller + Infinity Fabric)
- 24 Zen 4 cores distributed across SoC die
- Active interposer interconnect
Comparison with MI300X
| Metric | MI300A | MI300X |
|---|
| CPU | 24 Zen 4 cores | None |
| Memory | 128GB HBM3 | 192GB HBM3 |
| Bandwidth | 5.3 TB/s | 5.3 TB/s |
| FP16 | 1.5 PFLOPS | 1.5 PFLOPS |
| TDP | 600W | 750W |
| Use | HPC + AI | Pure AI |
El Capitan Supercomputer
- 2024 TOP500 #1 (2024-11)
- 1.742 ExaFLOPS FP64 (double precision)
- 44,544 MI300A units
- Power consumption ~30 MW (vs 50+ MW for top x86 supercomputers)
- HPC tasks: Nuclear weapon simulation, climate change, materials science
Use Cases
- ✅ HPC + AI convergence (El Capitan-class supercomputers)
- ✅ Numerical simulation + ML hybrid (climate, materials, life sciences)
- ✅ Large model training (replaces 192GB MI300X)
- ✅ Graph neural networks requiring CPU acceleration
- ❌ Pure LLM inference (use MI300X or H100)
- ❌ Edge deployment (600W TDP)