What Is Strix Halo? AMD Ryzen AI Max Explained for Mini PC Users

Definition: What Is Strix Halo?

The word “APU” is key here. Unlike a traditional CPU (which has minimal integrated graphics) or a traditional GPU (which is a separate card), an APU is designed to blur that line. Strix Halo takes this concept further than any previous AMD chip: its integrated GPU is so capable that, in mini PC implementations, it genuinely competes with mid-range discrete graphics cards costing $200–$300.

The “Ryzen AI Max” branding is AMD’s consumer-facing name for Strix Halo chips. The full product names you’ll see on spec sheets are Ryzen AI Max 395 (16 cores, 40 GPU CUs) and Ryzen AI Max 385 (16 cores, 32 GPU CUs). The “Max+” suffix (Ryzen AI Max+ 395) indicates a higher TDP configuration allowing sustained performance above the baseline.

Architecture Explained — CPU, GPU, NPU, and Memory

Strix Halo packages four distinct processing units — CPU, GPU, NPU, and memory controller — into a single chip connected by AMD’s high-bandwidth die-to-die interconnect. This allows the GPU to access system RAM at 256 GB/s, dramatically more than the 32–64 GB/s typical of discrete GPU GDDR6 VRAM in the same price class.

The CPU: 16 Zen 5 Cores

The Ryzen AI Max 395 uses AMD’s Zen 5 microarchitecture — the same generation found in desktop Ryzen 9000 processors. With 16 cores and 32 threads, it scores approximately 35,000–36,000 points in Cinebench R23 Multi, competing directly with desktop-class processors used in full tower PCs. Boost clocks reach up to 5.1 GHz. This is not a compromised mobile processor — it’s full desktop-class CPU performance.

The GPU: 40 RDNA 3.5 Compute Units

This is the headline feature. RDNA 3.5 is AMD’s latest GPU architecture (also used in the Radeon RX 7000 discrete card series), and Strix Halo packs an unprecedented 40 compute units into an integrated design. Each compute unit contains 64 shader processors, giving the Radeon 8060S a total of 2,560 shader processors. For comparison, the discrete Radeon RX 7600 XT contains 32 CUs (2,048 shaders) — the Strix Halo iGPU actually exceeds it in CU count.

The GPU runs at up to 2.9 GHz and is rated at approximately 11.8 TFLOPS of compute performance. In 3DMark Time Spy — a standard GPU benchmark — it scores 10,800–11,250 points, which places it in the same tier as the NVIDIA RTX 4060 (typically 10,000–11,000 points).

The NPU: 50 TOPS for AI

The XDNA 2 NPU (Neural Processing Unit) delivers 50 TOPS (Tera Operations Per Second) of AI inference performance. This exceeds the Windows 11 Copilot+ PC requirement of 40 TOPS, qualifying all Strix Halo devices as Copilot+ PCs. In practice, the NPU accelerates tasks like real-time image upscaling, background removal in video calls, noise cancellation, and — increasingly — running smaller AI language models without loading the GPU or CPU.

Strix Halo vs Strix Point — The Differences That Matter

Strix Halo (Ryzen AI Max) has 40 GPU CUs and supports up to 128GB of unified memory. Strix Point (Ryzen AI 300, including the HX 370) has 16 GPU CUs and supports up to 96GB. Both use Zen 5 CPUs and XDNA 2 NPUs — the primary difference is GPU power and maximum memory capacity. Strix Halo is significantly better for gaming and local AI inference with large models.

The GPU gap is the decisive number: Strix Halo’s 40 CUs deliver approximately 3.4× more GPU compute than Strix Point’s 16 CUs. This is the difference between 1080p-capable gaming and genuine 1440p gaming without a discrete GPU. For AI workloads, the doubled memory bandwidth (256 vs 120 GB/s) is equally significant — it directly determines how fast large language models can generate tokens.

Gaming Performance: How Good Is the Strix Halo iGPU?

The Radeon 8060S in Strix Halo reaches RTX 4060-equivalent gaming performance in benchmarks, and delivers around 85 FPS in Cyberpunk 2077 at 1440p High with FSR enabled. This is far beyond any previous integrated graphics and genuinely replaces a mid-range discrete GPU for 1080p and 1440p gaming.

In real gaming tests, the GMKtec EVO-X2 (Ryzen AI Max+ 395 at 120W sustained TDP) achieves approximately 85 FPS in Cyberpunk 2077 at 1440p High with FSR Quality, 70 FPS in Red Dead Redemption 2 at 1440p High, and 130+ FPS in Counter-Strike 2 at 1440p High. These numbers come from tests by independent reviewers including ETA PRIME and Retro Game Corps.

AI Capabilities: Running Large Language Models Locally

A 128GB Strix Halo mini PC can run Qwen3 235B — a 235-billion-parameter AI model — at approximately 11 tokens per second locally, without any cloud connection. This is possible because up to 96GB of the unified memory can be dynamically allocated as VRAM, allowing massive models to fit entirely in memory on a single device.

To put this in context: running a 235-billion-parameter model locally was previously the domain of multi-GPU workstations costing $10,000+. A dual RTX 3090 setup (48GB VRAM total) can run the same model at only 5–6 tokens per second, and requires complex software configuration. The Strix Halo mini PC does it on a single device with a simpler setup.

The key enabler is unified memory: because the CPU and GPU share the same physical RAM pool, the GPU can access up to 96GB of data without the transfer bottleneck of PCIe between a CPU and a discrete GPU. Models like Qwen3 235B, which use a Mixture-of-Experts architecture, benefit disproportionately from large unified memory because they need to rapidly switch between different expert networks during inference.

The picture is nuanced: for smaller models (7B–13B), a dedicated RTX 4090 is faster. The Strix Halo advantage emerges specifically with very large models that don’t fit in discrete GPU VRAM. For most users running a local AI assistant on a model like Llama 3 70B or Mistral 7B, performance is perfectly usable. For researchers or developers who want to run frontier-scale models without multi-GPU complexity, Strix Halo is genuinely unique.

Unified Memory — Why It’s a Game Changer

In conventional computing, a discrete GPU has its own dedicated VRAM (e.g., 8GB on an RTX 4060). When the GPU runs out of VRAM, performance collapses. Strix Halo’s unified memory means the GPU can dynamically access up to 96GB of system RAM as VRAM — eliminating this constraint entirely.

This architecture is similar to what Apple implemented with its M-series chips (Apple Silicon). When AMD released Strix Halo, it was the first time an x86 APU offered a comparable level of GPU memory flexibility. The result is a chip that punches significantly above its iGPU status for workloads limited by GPU memory capacity.

For gaming specifically, the practical benefit is that games with very high texture quality settings (which often require 8–12GB of VRAM) can run without VRAM overflow stutters on a Strix Halo system — something that would cause problems on an 8GB discrete GPU. For AI workloads, the benefit is even more dramatic, as we covered in Section 5.

Which Mini PCs Use Strix Halo in 2026?

True Strix Halo (Ryzen AI Max) mini PCs are relatively rare and premium-priced, starting at $1,499. The GMKtec EVO-X2 is the most widely reviewed model. Many mini PCs marketed as “Strix” actually use Strix Point (Ryzen AI 9 HX 370) — a capable but different chip. Read spec sheets carefully.

Should You Buy a Strix Halo Mini PC?

Buy a Strix Halo mini PC if you need 1440p gaming without a discrete GPU, want to run large AI models locally (70B+), or require a workstation-class CPU in a compact form. If your budget is under $1,200, Strix Point (HX 370) mini PCs offer 80% of the CPU performance and solid 1080p gaming for $400–$600 less.

Buy Strix Halo (Ryzen AI Max) if…

• You want the best iGPU gaming available without a separate GPU — 1440p capable out of the box
• You want to run large language models locally (70B+ parameters) without a multi-GPU setup
• You need maximum CPU performance (16 full Zen 5 cores) for demanding workloads like video encoding or 3D rendering
• You are considering the 128GB configuration for AI research or professional workloads requiring large memory pools
• Budget is $1,499+ and you don’t need an eGPU upgrade path

Choose Strix Point (HX 370) instead if…

• Your budget is under $1,200 — the performance-per-dollar is significantly better
• You plan to add an eGPU dock via OCuLink or USB4 — the HX 370’s iGPU becomes secondary
• 1080p gaming is your primary gaming target — the HX 370’s 16 CU iGPU handles it well
• You want upgradable SO-DIMM RAM — the ACEMAGIC Retro X5 (HX 370) supports up to 128GB user-replaceable DDR5

Frequently Asked Questions