AMD Ryzen AI Max (Strix Halo XDNA 2 128GB UMA)
Overview
AMD Ryzen AI Max / Max+ (codename Strix Halo) is AMD's flagship AI PC APU, announced at CES 2025-01 and mass-produced 2025 H1. Built on TSMC N4P + N6 chiplets, it features a 16-core Zen 5 CPU @ 5.1 GHz, 40 CU RDNA 3.5 GPU (Radeon 8060S integrated graphics, matching discrete GPU performance), XDNA 2 NPU 50 TOPS (2× the Ryzen AI 300), 256-bit LPDDR5X-8000, up to 128 GB UMA (96 GB assignable to GPU as VRAM), and 256 GB/s memory bandwidth. The Ryzen AI Max+ 395 is the flagship model. Stable Diffusion 3.5 Large performance is 3.9× Apple MacBook Pro M4 Pro 48GB (AMD official data), enabling 70B LLM on-device execution. Strix Halo is the true monster of the AI PC era.
Core Specifications
| Item | Spec |
|---|
| Architecture | Chiplet heterogeneous (CPU + GPU + I/O die) |
| Process | TSMC N4P (CPU/GPU die) + TSMC N6 (I/O die) |
| CPU Cores | 16 Zen 5 cores (dual CCD) |
| CPU Frequency | 3.0 GHz base / 5.1 GHz boost |
| L2/L3 Cache | 16 MB L2 / 64 MB L3 |
| GPU Architecture | Radeon 8060S (RDNA 3.5) |
| GPU CUs | 40 CU (comparable to RTX 4060 Laptop performance) |
| NPU Architecture | XDNA 2 (upgraded from XDNA) |
| NPU TOPS | 50 TOPS INT8 (2× Ryzen AI 300 Phoenix / Hawk Point at 25-30 TOPS) |
| Total AI TOPS | 126 TOPS (CPU + GPU + NPU combined) |
| Memory | 256-bit LPDDR5X-8000 |
| Max Memory | 128 GB UMA |
| Assignable VRAM | 96 GB (AMD Variable Graphics Memory) |
| Memory Bandwidth | 256 GB/s (128 GB 8-channel LPDDR5X-8000) |
| PCIe | PCIe Gen 4 (16 lanes) |
| TDP | 55W / 85W / 120W (flexible cTDP) |
| Wi-Fi | Wi-Fi 7 |
| Launch Models | Ryzen AI Max 390, Ryzen AI Max+ 395, Ryzen AI Max+ 388, PRO series |
| Announcement | CES 2025-01, mass production 2025 H1, detailed release 2025-03-17 |
| Ryzen AI Max+ 395 Price | Laptop $1,499-$2,499 |
Comparison with Apple M4 Pro / M4 Max
| Metric | Ryzen AI Max+ 395 | M4 Pro 48GB | M4 Max 128GB |
|---|
| CPU Cores | 16 Zen 5 | 14 (10P+4E) | 16 (12P+4E) |
| CPU Frequency | 5.1 GHz | 4.5 GHz | 4.0 GHz |
| GPU CUs | 40 RDNA 3.5 | 20 Apple GPU | 40 Apple GPU |
| NPU TOPS | 50 (XDNA 2) | 38 (NE) | 38 (NE) |
| Memory | 128 GB LPDDR5X | 48 GB LPDDR5X | 128 GB LPDDR5X |
| Memory Bandwidth | 256 GB/s | 273 GB/s | 546 GB/s |
| Assignable VRAM | 96 GB | Shared UMA | Shared UMA |
| Stable Diffusion 3.5 | 3.9× M4 Pro 48GB | Baseline | 2.5× M4 Pro |
| 70B LLM (Q4) | Fits 96 GB VRAM | Swap performance loss | Fits |
| TDP | 55-120W | 38-46W | 70-100W |
3.9× Stable Diffusion advantage: In AMD's official testing, the Ryzen AI Max+ 395 128GB running Stable Diffusion 3.5 Large image generation is 3.9× faster than the MacBook Pro M4 Pro 48GB. The reason: M4 Pro 48GB requires memory swap (can't fully fit 70B LLM), while the Ryzen AI Max+ 395's 96GB VRAM loads the full model without swapping.
| Model / Task | Ryzen AI Max+ 395 128GB | Note |
|---|
| Llama 3 70B (Q4_K_M) | Full model fits 96GB VRAM | On-device 70B |
| Llama 3 70B (Q4) | ~5 tok/s | M4 Pro 48GB with swap ~2 tok/s |
| Stable Diffusion 3.5 Large | 3.9× M4 Pro 48GB | 1024×1024 image generation |
| Stable Diffusion 3.5 Medium | 3.3× (concurrent) | Parallel multi-image generation |
| Phi-4 14B | 2.6× (concurrent) | Running simultaneously with SD 3.5 |
| Whisper Large V3 | Real-time | ONNX accelerated |
| Video conferencing AI (background blur) | Real-time | NPU 4.0 equivalent |
| Image segmentation | 60+ FPS | GPU accelerated |
Competitor Comparison
| Metric | Ryzen AI Max+ 395 | Intel Core Ultra 9 288V | Qualcomm Snapdragon X Elite |
|---|
| CPU | 16 Zen 5 | 4P+4E (8 cores) | 12 Oryon |
| GPU | 40 RDNA 3.5 CUs | 8 Xe2 (Arc 140V) | Adreno (unreleased) |
| NPU | 50 TOPS (XDNA 2) | 48 TOPS (NPU 4.0) | 45 TOPS (Hexagon) |
| Total AI TOPS | 126 | 120 | 75 |
| Memory | 128 GB LPDDR5X | 32 GB on-package | 32-64 GB LPDDR5X |
| Assignable VRAM | 96 GB | Shared UMA | Shared UMA |
| 70B LLM On-Device | Fits | Doesn't fit | Doesn't fit |
| Copilot+ PC | Certified | Certified | Certified |
| Battery Life | ~8-12 hours | ~20+ hours | ~20+ hours |
Use Cases
- AI PC laptops (Framework Desktop, Asus ROG Flow Z13, HP ZBook Ultra G1a, Lenovo ThinkPad P16)
- Local 70B LLM inference (128GB / 96GB VRAM fits full 70B Q4)
- Stable Diffusion 3.5 image generation (3.9× M4 Pro)
- Professional 3D rendering / CAD (RDNA 3.5 40 CU near RTX 4060 Laptop)
- 4K/8K video editing (DaVinci Resolve, Premiere Pro)
- Large model fine-tuning (LoRA QLoRA 7B-13B local training)
- Scientific computing (16-core CPU + 40 CU GPU hybrid)
| Item | Detail |
|---|
| Vendor | AMD Inc. (Santa Clara, USA) |
| Foundry | TSMC Taiwan (N4P chiplet + N6 I/O die) |
| Software Stack | Windows 11 24H2 (Copilot+ PC), Linux (Ubuntu 24.04, Fedora 41) |
| AI Frameworks | ROCm (AMD GPU), ONNX Runtime, DirectML, Lemonade SDK (AMD on-device LLM), VLLM, llama.cpp (Vulkan backend) |
| OEM Customers | Framework Desktop, HP ZBook Ultra G1a, Asus ROG Flow Z13, Lenovo ThinkPad P16 |
| Launch Price | Framework Desktop 16" from $1,999 / HP ZBook Ultra G1a $2,499 |
| PRO Series | Ryzen AI Max PRO 390 / PRO 395 (workstation-class, vPro, ECC RAM) |
| Competitors | Intel Core Ultra 200V (Lunar Lake), Apple M4 Pro/Max, Qualcomm Snapdragon X Elite |
Key Features
- Chiplet heterogeneous (CPU/GPU/NPU integrated)
- TSMC N4P process (ahead of Intel 4)
- 16 Zen 5 cores @ 5.1 GHz (top-tier consumer PC)
- 40 CU RDNA 3.5 GPU (Radeon 8060S, near RTX 4060 Laptop discrete GPU)
- XDNA 2 NPU 50 TOPS (2× Ryzen AI 300)
- Total AI compute 126 TOPS (CPU + GPU + NPU)
- 128 GB LPDDR5X unified memory (largest in consumer PC history)
- 96 GB assignable VRAM (runs 70B LLM on-device)
- 256 GB/s memory bandwidth (256-bit bus)
- AMD Variable Graphics Memory (dynamically adjustable VRAM)
- PCIe Gen 4 + Wi-Fi 7
- 55W / 85W / 120W flexible cTDP
- Copilot+ PC certified (Recall, Live Captions, Cocreator)
- Stable dual-platform Windows / Linux