Ryzen AI Max Processors in 2026: Strix Halo Explained

Ryzen AI Max processors, codenamed Strix Halo, are AMD’s largest mobile chips: the flagship Ryzen AI Max+ 395 combines 16 Zen 5 cores, a 40-compute-unit Radeon 8060S graphics engine, an XDNA 2 NPU rated at up to 50 TOPS and a 256-bit LPDDR5X memory interface that can feed up to 128 GB of unified memory. They are built for thin workstations, compact gaming machines and mini PCs that need integrated graphics far beyond ordinary laptop chips, and for running large local AI models.

This guide covers the specifications, how the unified memory design works, where the chips make sense, and what to check before buying a system built around one. Details come from AMD’s published specifications and from reputable references; individual systems differ in power limits and cooling, so system reviews matter as much as the processor name.

Our top picks

Top pickHP OmniBook 16 Touchscreen Laptop AMD Ryzen

Check price on Amazon

Also greatAcer Nitro V 16S AI Gaming Laptop

Check price on Amazon

Also greatAcer Aspire Go 15 AI Ready Laptop

Check price on Amazon

Also greatHP OmniBook Touchscreen Laptop

Check price on Amazon

As an Amazon Associate we earn from qualifying purchases at no extra cost to you.

Ryzen AI Max at a glance

Compared with the standard Ryzen AI 9 HX 370, which has 12 cores, 16 graphics CUs and a 128-bit memory bus, the Max+ 395 more than doubles the graphics hardware and doubles the memory bus width. That is why it lands in a different product category, sitting between ordinary thin laptops and machines with a separate graphics card.

The Ryzen AI Max family

The business-oriented PRO 495, introduced in 2026 according to published reports, keeps 16 cores and moves to LPDDR5X-8533 with support for up to 192 GB of memory. Lower models cut cores and compute units, which reduces cost but keeps the wide memory interface.

Unified memory: the defining feature

A traditional laptop with a discrete graphics card has two separate memory pools: system memory for the processor and a fixed amount of video memory on the graphics card. Strix Halo uses one shared pool of LPDDR5X that both the CPU and GPU can address. AMD’s Variable Graphics Memory lets a 128 GB system assign up to 96 GB to the GPU, leaving about 32 GB for the operating system and applications.

Why this matters for AI

Large language models need to fit entirely in graphics-accessible memory to run quickly. Consumer graphics cards typically offer 8 to 24 GB, which limits models to a modest size. A system with 96 GB available to the GPU can load far larger models, including quantized models in the range of tens of billions of parameters, that would not fit on a mainstream discrete card at all. Speed is lower than a high-end graphics card because the 256 GB/s of bandwidth is a fraction of what the memory on a flagship card delivers, but the ability to run the model at all is the selling point.

Worked example: will a model fit?

A model with 70 billion parameters stored at 4 bits per weight needs about 35 GB for the weights alone (70 billion multiplied by 0.5 bytes), plus extra memory for the context cache and working buffers, so plan roughly 40 to 45 GB in total. That is comfortably inside a 96 GB graphics allocation but beyond any single consumer graphics card. By contrast, a 7 billion parameter model at the same precision needs only about 4 GB and runs on far cheaper hardware, so the Max chips are overkill for small models.

The memory trade-off

Because the memory is soldered, you cannot upgrade it later. Buy the capacity you will need on day one: 32 GB suits gaming and creative work, 64 GB suits heavier workloads, and 128 GB is for large local AI models.

Graphics performance in practice

The Radeon 8060S has 40 compute units, which is a larger count than the Radeon RX 7600-class discrete GPUs in some laptops. Reviews of Strix Halo systems report gaming results that sit between mainstream integrated graphics and mid-range laptop discrete cards, with many modern games playable at 1080p high settings and some at 1440p with upscaling. Results depend on the power limit: at 120 W in a mini PC the chip performs noticeably better than at 55 W in a thin laptop.

FSR upscaling and frame generation help bridge the gap in demanding games, and AMD’s driver supports the features on these chips.

Systems built around Ryzen AI Max

Several manufacturers have released systems using the Max+ 395. Examples include the ASUS ROG Flow Z13 two-in-one gaming tablet, the HP ZBook Ultra G1a mobile workstation, the Framework Desktop compact desktop, and mini PCs such as the Beelink GTR9 Pro and the GMKtec EVO-X2. Each handles power and cooling differently: a tablet may sustain around 55 to 70 W, while a mini PC with a large fan can hold close to 120 W.

Memory bandwidth explained

Bandwidth is the speed at which the processor and graphics engine can read and write memory, and it is the main limit on both gaming frame rates and AI token generation. For Strix Halo the arithmetic is simple. A 256-bit bus is 32 bytes wide, and LPDDR5X-8000 performs 8,000 million transfers per second, so the peak is 8,000 million multiplied by 32 bytes, or 256 GB/s. A standard laptop chip with a 128-bit bus at the same speed reaches only 128 GB/s, which is why the Max line is far better at graphics and large models.

Dedicated graphics cards still have an edge. Memory on a mainstream discrete card commonly delivers several hundred GB/s, and flagship cards exceed one terabyte per second. That gap explains why token generation on a Max system, while workable, is slower than on a high-end discrete card when the model fits in both. A rough rule: the number of tokens produced per second cannot exceed the bandwidth divided by the size of the model in memory. A 35 GB model on 256 GB/s therefore tops out near 7 tokens per second in theory, and real systems reach somewhat less. Smaller or more heavily quantized models run proportionally faster.

CPU performance

The 16 Zen 5 cores are a complete desktop-class complement, unlike the hybrid Zen 5 and Zen 5c mix in the Ryzen AI 9 chips. All cores are full Zen 5 and share 64 MB of L3 cache, with boost up to 5.1 GHz. At the higher power limits the Max+ 395 competes with desktop 12-core and 16-core processors in rendering and compilation, according to reviewers, while at lower limits it still outperforms most thin-and-light laptop chips. The advantage is greatest in long multithreaded jobs where the larger cooling in a mini PC allows sustained clocks.

Strix Halo versus a laptop with a discrete graphics card

The comparison shows why the right choice depends on the job. For mixed work that includes large models, big datasets and moderate gaming, the unified design is compelling. For games alone, a discrete card is faster.

The NPU on a Max system

The XDNA 2 NPU in the Max chips matches the one in the Ryzen AI 300 series at up to 50 TOPS. It handles Windows AI features and small efficient models with low power use, which keeps battery drain down during video calls. The GPU, not the NPU, does the heavy lifting for large language models and image generation, so the headline platform figure of about 126 TOPS reflects the combined theoretical output of CPU, GPU and NPU rather than a single engine. Treat it as a marketing total and rely on application-specific results when planning purchases.

Storage, ports and expandability

Systems built on Strix Halo typically include one or two M.2 slots for NVMe storage, with PCIe 4.0 speeds that are more than adequate for most tasks. USB4 ports support fast external drives and docks, and many mini PCs add 2.5 or 10 gigabit Ethernet, which helps when the machine serves as a local AI server on a home network. There is generally no free slot for a discrete graphics card in laptops and tablets, though some small desktops expose a PCIe slot for add-in cards.

Reliability and ownership realities

• Thermal paste and dust: sustained high loads in compact enclosures collect dust in heat sinks; clean fans every six to twelve months.
• Firmware updates: BIOS and driver releases for new platforms improve stability, memory handling and power management, so check for updates after purchase.
• Warranty and support: smaller mini PC brands vary in service quality, so verify warranty terms and replacement part availability.
• Soldered components: memory and often storage connectors are fixed or limited, so a failure may require replacing the whole board.
• Power adapters: high-wattage designs need the supplied adapter; third-party chargers may throttle performance.

Who should consider a Ryzen AI Max system

• Local AI users: developers and researchers running large models privately without a data center or an expensive multi-GPU rig.
• Compact workstation buyers: designers, video editors and engineers who want a small, quiet machine with big memory.
• Space-limited gamers: people who want solid 1080p gaming in a small enclosure without a discrete graphics card.

Who should look elsewhere

• Buyers needing the fastest gaming: a laptop or desktop with a high-end discrete graphics card is faster per dollar of hardware.
• Battery-life seekers: a regular Ryzen AI 300 or similar efficient chip lasts longer on light tasks.
• Users who want upgradeable memory: the soldered design rules that out.

Power, cooling and noise

The 55 W default and configurable range up to 120 W mean cooling design is critical. In slim laptops, fans spin up under load, and a sustained 55 W workload makes the chassis warm. Mini PCs with larger heat sinks run quieter and hold higher clocks. When reading reviews, look for sustained-power graphs rather than single benchmark runs, because short bursts above the sustained limit flatter the numbers.

Estimating running cost

A mini PC averaging 80 W for eight hours of daily work uses 0.64 kWh per day. At 0.15 per kWh, that costs about 0.10 per day or roughly 35 per year, which is low compared with a desktop that pairs a high-end CPU with a power-hungry discrete card at 400 W or more.

Buying checklist

1. Decide on memory first: 32, 64 or 128 GB cannot change later.
2. Check the sustained power limit in reviews, not just the chip name.
3. Confirm the memory is running in its full 256-bit configuration.
4. Look at ports: USB4, a fast Ethernet port and HDMI or DisplayPort outputs matter for mini PCs.
5. Verify driver and software support for your AI frameworks, since the GPU uses ROCm and Vulkan paths that vary by application.
6. Compare against a system with a discrete graphics card if gaming speed is the main goal.

Software notes for AI users

Running models on Strix Halo works through several routes. Popular local-model runners support the GPU via Vulkan, and AMD continues to extend ROCm support for the integrated graphics. The NPU is most useful for small, efficient models and Windows features rather than for large language models, which run on the GPU. Expect software maturity to vary by tool, and check each project’s documentation before purchase if a specific workflow matters to you.

Tips for getting the best results

1. Set the graphics memory allocation deliberately. Giving the GPU too little limits large models, while giving it too much starves the operating system. For a 128 GB machine, 64 to 96 GB for the GPU is a common range depending on workload.
2. Pick the right power mode. Use the highest sustained mode when plugged in for games and renders, and a balanced mode on battery to avoid rapid drain.
3. Use quantized models. Four-bit and eight-bit versions of large models reduce memory use and raise speed with modest quality loss.
4. Keep the system cool. Elevate laptops or leave breathing room around mini PCs so fans do not hit thermal limits.
5. Update graphics drivers regularly. Strix Halo is a newer design and driver releases continue to add performance and compatibility improvements.

Following these steps usually matters more than small differences between brands, because the processor is the same across systems and results are determined by power, cooling and memory configuration.

Frequently Asked Questions

What is Ryzen AI Max?

It is AMD’s high-end mobile processor line, also known by the codename Strix Halo. The flagship Max+ 395 has 16 Zen 5 cores, 40 graphics compute units and a 256-bit LPDDR5X memory interface.

How much memory can the Ryzen AI Max+ 395 use?

It supports up to 128 GB of soldered LPDDR5X-8000. On a 128 GB system, AMD says up to 96 GB can be assigned to the GPU.

Is Ryzen AI Max good for gaming?

It is strong for integrated graphics and handles many games at 1080p with high settings, but a discrete mid-range or high-end graphics card is still faster. Its strength is combining solid graphics, a capable CPU and large memory in one small package.

Can it run large AI models locally?

Yes, that is one of its main strengths. With 96 GB available to the GPU, quantized models far larger than any mainstream graphics card can hold will fit, though token generation speed is limited by memory bandwidth.

Can I upgrade the RAM later?

No. The LPDDR5X memory is soldered to the board, so you must choose the capacity when buying the system.

How is it different from the Ryzen AI 9 HX 370?

The HX 370 has 12 cores, 16 graphics compute units and a 128-bit memory bus. The Max+ 395 has 16 cores, 40 compute units and a 256-bit bus, so it is much stronger in graphics and AI memory capacity but uses more power.