AMD Strix Halo Explained: Ryzen AI Max Mini PCs for Gaming and Local AI (2026)
Strix Halo, Ryzen AI Max, Radeon 8060S: three names for the same chip, and the most powerful processor you can find in a mini PC today. Here is what it is, the full lineup, what it really does for gaming and local AI — with measured numbers — and which mini PCs use it.
In short
Strix Halo is AMD’s codename for the Ryzen AI Max processors: one chip with up to 16 Zen 5 cores, a 40-CU Radeon 8060S GPU, a 50 TOPS NPU and up to 128GB of memory shared by the CPU and GPU. In a mini PC it means 1440p-class gaming without a graphics card and the ability to run 100B-class AI models locally. It is expensive — and memory prices pushed it higher in 2026 — so it only makes sense if you need that GPU or that memory.
What changed in this update (Oct 8, 2026)
We replaced our local AI speed table with measured results (our April figures for dense 70B and 7B models were too optimistic), added the full Ryzen AI Max lineup including the Max+ 392 and 388 announced at CES 2026, listed all the Strix Halo mini PCs we cover, corrected the Strix Point memory limit and removed written prices.
What is Strix Halo?
An APU — CPU, GPU and NPU on one package — designed to give compact devices desktop-class performance without a separate graphics card.
“Strix Halo” is AMD’s internal codename; “Ryzen AI Max” is the name on spec sheets; “Radeon 8060S” is its integrated GPU. What makes it different from every earlier AMD laptop chip is scale: a GPU with 40 compute units (2,560 shaders) and a 256-bit memory bus, twice the width of the usual laptop design. That wide bus is what feeds both the GPU in games and large AI models.
Don’t confuse Strix Halo with Strix Point
Strix Point is the Ryzen AI 300 family, including the popular Ryzen AI 9 HX 370 with a 16-CU Radeon 890M. It is a very good chip, but a much smaller one. Many listings say “Strix” or “AI Max-class” loosely: look for “Ryzen AI Max” and “Radeon 8060S” (or 8050S).
The Ryzen AI Max lineup
The “+” models have the full 40-CU GPU; the others have 32 or fewer. Most mini PCs use the Max+ 395.
| Model | CPU | GPU | Note |
|---|---|---|---|
| Ryzen AI Max+ 395 | 16 cores / 32 threads | 40 CU (Radeon 8060S) | Flagship, used in most mini PCs |
| Ryzen AI Max+ 392 | 12 cores / 24 threads | 40 CU | Announced at CES 2026 |
| Ryzen AI Max+ 388 | 8 cores / 16 threads | 40 CU | Announced at CES 2026: full GPU, fewer cores |
| Ryzen AI Max 390 | 12 cores / 24 threads | 32 CU (Radeon 8050S) | Rare in mini PCs |
| Ryzen AI Max 385 | 8 cores / 16 threads | 32 CU (Radeon 8050S) | Entry Strix Halo |
The Max+ 392 and 388 matter for buyers: they keep the full 40-CU GPU and the same memory bus with fewer CPU cores, so for gaming and AI they should come close to the 395 in machines that cost less.
Inside the chip: CPU, GPU, NPU and memory
CPU: 16 Zen 5 cores
The same Zen 5 core generation as AMD’s desktop Ryzen 9000 processors, with boost clocks up to 5.1GHz. In mini PCs running at high power, published Cinebench R23 multi-core results exceed 30,000 points — tower-PC territory.
GPU: Radeon 8060S, 40 RDNA 3.5 compute units
By far the largest integrated GPU AMD has made. In published 3DMark Time Spy results it lands close to an entry desktop graphics card such as the RTX 4060, roughly three times the Radeon 890M of Strix Point. Performance depends heavily on the power limit the mini PC maker allows.
NPU: 50 TOPS
The XDNA 2 NPU exceeds Microsoft’s 40 TOPS bar for Copilot+ PC features (live captions, Studio Effects, Recall). For running chatbots and other large language models, the GPU does the heavy lifting; see what an NPU actually does.
Unified memory: up to 128GB, about 256 GB/s
CPU and GPU share one pool of LPDDR5X-8000. In Windows, up to 96GB can be assigned to the GPU; Linux users report allocating more. Theoretical bandwidth is about 256 GB/s, around 215 GB/s in practice. That is far more than ordinary laptop memory, but well below an Apple M4 Max (546 GB/s) or an RTX 4090 (1,008 GB/s) — which is exactly why some AI workloads fly and others crawl (next section). The memory is soldered: choose your capacity at purchase.
Local AI: what Strix Halo really does
It can load models no consumer graphics card can hold, but speed depends on how much of the model is read per word: mixture-of-experts models are fast, large dense models are slow.
Generating each token means reading the active part of the model from memory. With about 215 GB/s of real bandwidth, a 4-bit dense 70B model (around 40GB) tops out near 5 tokens per second. Mixture-of-experts (MoE) models only activate a few billion parameters per token, so they run many times faster while still benefiting from the 128GB capacity.
| Model | Type | Tokens per second | In practice |
|---|---|---|---|
| Qwen3-30B-A3B | MoE, ~3B active | ~72–75 | Very fast, great everyday assistant |
| Llama-class 7–8B | Dense | ~42–53 | Fast |
| GPT-OSS 120B | MoE, ~5B active | ~31 | Comfortable reading speed for a 120B model |
| Llama 3.3 70B | Dense | ~5 | Usable but slow |
Speeds vary with the software (llama.cpp, LM Studio, Ollama), backend (Vulkan or ROCm), context length and quantization. Our guides to running Ollama and LM Studio on a mini PC walk through setup, and our best mini PCs for local AI compares every price tier.
Strix Halo vs NVIDIA DGX Spark vs Mac
NVIDIA’s DGX Spark also offers 128GB of unified memory (about 273 GB/s) with CUDA support, at a higher price; a Mac Studio with M4 Max has roughly twice the bandwidth. Strix Halo’s case is Windows and Linux flexibility, gaming ability and x86 software. Full comparison: DGX Spark 64GB vs 128GB vs Strix Halo.
Gaming on Strix Halo
The best integrated graphics in any mini PC: 1440p in many games with upscaling, 1080p high settings comfortably — but a dedicated GPU still gives more frames for the money.
With 16GB or more assigned to the GPU, high-resolution textures never run out of video memory, a real advantage over 8GB graphics cards. The limits are bandwidth and power: heavy ray tracing is out of reach, and results vary between mini PCs depending on their power limit and cooling. If gaming is the main goal and AI is not, a mini PC with a dedicated GPU or an OCuLink eGPU setup is often better value — see best gaming mini PCs and what is OCuLink.
Power settings (TDP) matter
AMD lets manufacturers configure Ryzen AI Max chips roughly between 45W and 120W. Mini PCs usually run near the top of that range, which is why they outperform most Strix Halo laptops with the same chip. The flip side is heat and fan noise under load: two mini PCs with the same Max+ 395 can differ noticeably. Check each model’s sustained power mode and reviews of its noise.
Prices in 2026
Strix Halo is a premium platform, and the 2026 memory shortage hit it hard because every unit carries 64GB or 128GB of soldered LPDDR5X.
AMD’s own Ryzen AI Halo mini PC launched in the US with 128GB at a $3,999 list price, and MINIX listed a Strix Halo model at $4,600. Even Intel’s newer Panther Lake mini PCs did not undercut it: GMKtec announced its Panther Lake model at a higher list price than its own Strix Halo machine. Our RAM crisis guide explains why memory-heavy devices rose the most. Strix Point (HX 370) mini PCs remain the much cheaper route to Zen 5.
Strix Halo mini PCs we cover
| Model | Chip / memory | Stands out for | Links |
|---|---|---|---|
| GMKtec EVO-X2 | Max+ 395 · 64 or 128GB | The most widely reviewed Strix Halo mini PC | Review · Price |
| GMKtec EVO-X3 | Ryzen AI Max | OCuLink port for an external GPU | Review |
| Bosgame M5 | Max+ 395 · 128GB | One of the most affordable 128GB options | Review |
| Minisforum MS-S1 Max | Max+ 395 · 128GB | Workstation design, high-speed networking | Review · Price |
| AMD Ryzen AI Halo | Ryzen AI Max · 128GB | AMD’s own AI mini PC, US launch at $3,999 list | News |
| HP Z2 Mini G1a | Ryzen AI Max PRO | Business workstation with HP support | — |
Not Strix Halo, but often confused with it: Ryzen AI 9 HX 370 mini PCs such as the PELADN HO5, ACEMAGIC Retro X5 and GEEKOM A9 Max — excellent value for 1080p gaming and productivity.
Should you buy a Strix Halo mini PC?
- want to run 70B–120B-class AI models locally (take 128GB)
- want strong 1440p-class gaming with no graphics card
- need 16 Zen 5 cores for encoding, rendering or compiling in a small box
- mostly game: a dedicated-GPU or OCuLink setup gives more frames per dollar
- need speed on small models only: a cheaper HX 370 mini PC is enough
- want upgradeable memory: Strix Halo’s is soldered
FAQ
What is AMD Strix Halo?
Strix Halo is AMD’s codename for the Ryzen AI Max processors: a single chip with up to 16 Zen 5 CPU cores, a Radeon 8060S GPU with up to 40 RDNA 3.5 compute units, a 50 TOPS NPU and up to 128GB of LPDDR5X memory shared by the CPU and GPU.
What is the difference between Strix Halo and Strix Point?
Strix Halo (Ryzen AI Max) has up to 40 GPU compute units and a 256-bit memory bus (about 256 GB/s). Strix Point (Ryzen AI 300, such as the Ryzen AI 9 HX 370) has up to 16 compute units and a 128-bit bus. Both use Zen 5 cores and a 50 TOPS NPU; Strix Halo is far stronger for gaming and large AI models, and costs much more.
Is Strix Halo good for gaming?
Yes. Its Radeon 8060S is the most powerful integrated GPU available in mini PCs, close to an entry desktop graphics card such as the RTX 4060 in published 3DMark results. It handles 1440p gaming in many titles with upscaling, though a mini PC with a dedicated GPU is still faster for the price.
Can a Strix Halo mini PC run a 70B AI model?
Yes, but slowly: a 4-bit 70B dense model fits easily in 128GB, yet generates only about 5 tokens per second because every token reads the whole model through a roughly 256 GB/s memory bus. Mixture-of-experts models run much faster, for example about 31 tokens per second for GPT-OSS 120B and over 70 for Qwen3-30B-A3B in published tests.
Should I buy 64GB or 128GB?
128GB if local AI is a main reason to buy: it fits 100B-class models with room for long contexts. 64GB is enough for gaming, creative work and models up to about 30–40B. Memory is soldered, so you cannot upgrade later.
What TDP does Strix Halo run at in mini PCs?
AMD allows Ryzen AI Max chips to be configured roughly between 45W and 120W. Mini PCs usually run near the top of that range, which is why they outperform most Strix Halo laptops, but the exact limit and fan noise vary by model.
Is the Ryzen AI Max+ 395 the same as an “HX 395”?
There is no “HX 395”: the chip is the Ryzen AI Max+ 395. The HX name belongs to Strix Point chips such as the Ryzen AI 9 HX 370, which is a different, smaller processor. Check for “Ryzen AI Max” and “Radeon 8060S” in a listing.
Related
- Best mini PC for local AI (2026)
- DGX Spark vs Strix Halo
- What is an NPU?
- Best gaming mini PCs over $1,000
- Mini PC vs gaming laptop
Sources
- Notebookcheck, Ryzen AI Max+ 392 and 388 announced (Jan 6, 2026)
- DataHardware.ai, Strix Halo tokens per second: dense vs MoE (updated Oct 4, 2026) — compiling ServeTheHome, Level1Techs and llm-tracker results
- AMD, Ryzen AI Max+ 395 generative AI performance
- TweakTown, AMD Ryzen AI Halo launches in the US at $3,999
- VideoCardz, MINIX lists Strix Halo mini PC for $4,600
- VideoCardz, GMKtec Panther Lake mini PC priced above its Strix Halo system
