APU (Accelerated Processing Unit) Architecture
What is an APU
APU (Accelerated Processing Unit) is a processor that integrates CPU + GPU + NPU into a single package / single chip, sharing unified memory (UMA, Unified Memory Architecture). It eliminates data copying between CPU and GPU, making it ideal for heterogeneous computing scenarios (HPC, AI inference, local LLM).
Representative products:
- Apple M-Series (M2/M3/M4 Max/Ultra)
- AMD MI300A (CDNA 3 + Zen 4)
- Intel Core Ultra (Meteor Lake / Lunar Lake, integrated NPU)
APU Core Advantages
Unified Memory Architecture (UMA)
- CPU and GPU share the same LPDDR5/HBM pool
- No data copying (vs discrete GPU requiring PCIe transfer)
- 192GB Mac Studio can load a full 70B LLM
Memory Bandwidth
- Apple M2 Ultra: 800 GB/s
- AMD MI300A: 5.3 TB/s (HBM3)
- Bandwidth advantage is significant for data-intensive LLM inference
Heterogeneous Computing
- Numerical simulation (CPU) + AI inference (GPU) + signal processing (NPU) on the same chip
- Ideal for HPC + AI joint workflows
Mainstream APU Comparison
| APU | Process | CPU | GPU | Memory | Memory Bandwidth | Deployment |
|---|---|---|---|---|---|---|
| Apple M3 Ultra | 3nm | 32-core | 80-core | 192GB LPDDR5 | 800 GB/s | Workstation |
| Apple M4 Max | 3nm | 16-core | 40-core | 128GB LPDDR5X | 546 GB/s | Workstation |
| AMD MI300A | 5nm + 6nm | 24-core Zen 4 | 14,592 SP | 128GB HBM3 | 5.3 TB/s | Data center |
| Intel Core Ultra 9 285H | 3nm | 16-core | 8-core Arc | 96GB DDR5 | 89 GB/s | Laptop |
APU vs Discrete CPU+GPU
| Dimension | APU | Discrete CPU + GPU |
|---|---|---|
| Memory access | Shared (no copy) | PCIe transfer |
| Memory bandwidth | 800 GB/s - 5.3 TB/s | GPU HBM + CPU DDR |
| Compute | Medium (TDP limited) | High (independent cooling) |
| Flexibility | Unified software stack | CPU/GPU separated |
| Price | Medium | High |
| Best for | Local AI, laptops | Data center, training |
APU Use Cases
- ✅ Local LLM inference (UMA advantage, 70B+ models loadable)
- ✅ HPC + AI joint (El Capitan supercomputer)
- ✅ Workstation creative work (Final Cut Pro, DaVinci Resolve)
- ✅ On-device GenAI (Apple Intelligence, Copilot+ PC)
- ✅ Laptop / all-in-one AI (low power)
- ❌ Ultra-large model training (use discrete GPU clusters)
- ❌ Data center high density (use H100/MI300X)
Detailed Product Pages
Apple Silicon
- Apple M-Series Overview - M1/M2/M3/M4/M5 full comparison
- Apple M1 Pro - 2021-10, 10-core CPU 200 GB/s MacBook Pro 14"/16"
- Apple M1 Max - 2021-10, 32-core GPU 400 GB/s 64GB UMA dual ProRes
- Apple M2 Max - 2023-01, 96GB UMA 400 GB/s
- Apple M2 Ultra - 2023-06, 192GB UMA 800 GB/s UltraFusion
- Apple M3 Max - 2023-10, 128GB LPDDR5 400 GB/s
- Apple M3 Ultra - 2023-12, 256GB UMA 800 GB/s Apple Silicon flagship
- Apple M4 Pro - 2024-10, 14-core CPU 64GB UMA 273 GB/s
- Apple M4 Max - 2024-10, 128GB LPDDR5X 546 GB/s 38 TOPS ANE
- Apple M4 Ultra - 2025-Q4 estimated, 256GB fits 200B FP16
- Apple M5 Ultra - 2026 H2 estimated, 384GB LPDDR6 ~1 TB/s fits 400B FP8
AMD
- AMD MI300A - CDNA 3 + Zen 4, 128GB HBM3 5.3 TB/s data center APU
- AMD Ryzen AI Max (Strix Halo) - 128GB UMA 96GB VRAM, 70B LLM on-device, 3.9× M4 Pro SD
Intel
- Intel Core Ultra 2 (Lunar Lake) - Foveros 3D NPU 4.0 48 TOPS Xe2 GPU 60+ TOPS LPDDR5X on-package Copilot+ PC
Related Architectures
- GPU - Discrete GPU
- NPU - Edge NPU
- WSE - Wafer-scale
- Complete Comparison Table