Ryzen AI Max Mini-ITX: Framework Desktop Specs Explained

AMD's Ryzen AI Max 300 series — codename Strix Halo — is the chip behind the current wave of "Ryzen AI Max Mini-ITX" builds, led by Framework's Mini-ITX Framework Desktop board. It isn't a conventional socketed desktop CPU; it's a single soldered package that bundles a Zen 5 CPU, an RDNA 3.5 integrated GPU, and a dedicated NPU behind one large pool of unified memory. That distinction matters for anyone comparing it to a traditional Mini-ITX build, because several specs that apply to normal AM5 boards — swappable coolers, a-la-carte CPU upgrades, DIMM-slot memory — don't carry over.

This piece is a synthesis of AMD's published platform specs and publicly available system data, cross-referenced against SpecPicks' catalog for real-world alternative build components. Where a claim couldn't be grounded in a verifiable source, it's flagged as such rather than stated as fact.

What is Ryzen AI Max Mini-ITX, exactly?

"Ryzen AI Max Mini-ITX" refers to systems built around AMD's Ryzen AI Max 300 series SoC on a Mini-ITX form-factor board — most visibly Framework Desktop, announced by Framework in 2025. The lineup spans four SKUs differentiated mainly by CPU core count and iGPU compute-unit count:

The common thread across all four is the architecture, not a specific core count: Zen 5 CPU cores (not Zen 4 — an easy mix-up given the platform launched alongside several Zen 4-based Ryzen AI 300 mobile parts), an RDNA 3.5 iGPU, and AMD's XDNA2 NPU, which AMD's own marketing rates at up to 50 TOPS. Because the memory controller, CPU, GPU, and NPU sit on one package talking to soldered LPDDR5X, there's no separate GPU slot and no user-swappable DIMMs on the flagship configurations — a meaningful departure from a normal Mini-ITX build.

Memory: the platform's actual headline feature

The reason Ryzen AI Max keeps coming up in AI-workload conversations isn't raw compute — it's memory capacity. Top configurations are specified with up to 128GB of unified LPDDR5X, addressable by the CPU, iGPU, and NPU without copying data across a PCIe bus the way a discrete GPU setup would. That's a meaningfully larger pool than the 8-24GB of VRAM found on most consumer discrete GPUs, and it's the core argument in favor of the platform for local LLM work — see Ryzen AI Max for Local LLMs: What the Hardware Enables for how that translates into which model sizes actually fit.

The tradeoff is that this memory is fixed at purchase. A traditional Mini-ITX build lets you buy RAM separately and upgrade later — for example, pairing a Ryzen 5 5600X (/product/B08166SLDF) with a Crucial 16GB DDR4 kit (/product/B08C4WV6FT) or a smaller 8GB kit (/product/B08C4Z69LN) and adding more down the line. On Ryzen AI Max Mini-ITX boards, whatever memory configuration you order is permanent.

AI and LLM workload performance: what's actually verifiable

Several figures circulating around Ryzen AI Max — claims of matching datacenter accelerators like the MI210 or MI300X, or fixed "tokens per second" numbers for specific models — don't hold up as stated; those are different classes of hardware (multi-hundred-watt datacenter GPUs with dedicated HBM) and comparing a sub-150W integrated platform directly to them overstates the picture. The credible, checkable claim is narrower: the large unified memory pool lets Ryzen AI Max run larger quantized models locally than most consumer discrete-GPU rigs can fit in VRAM, at throughput that depends heavily on model size, quantization, and the inference engine used.

For model-specific throughput data rather than a single blanket number, SpecPicks' companion breakdowns track what's publicly reported:

• Ryzen AI Max LLM Test: What the Public Data Shows
• Ryzen AI Max+ 395 LLM Performance: What to Expect
• AMD Ryzen AI Max for Local LLMs: What the Specs Mean
• Ryzen AI Max+ 395 LLM Inference: Specs, Bandwidth, Reality

The throughline across those pieces: treat any single "X tokens/second" figure with skepticism unless it names the exact model, quantization level, and inference backend, because all three shift the number substantially.

Cooling: there's no aftermarket cooler swap

Because the CPU, iGPU, and NPU are soldered to the board rather than socketed, Ryzen AI Max Mini-ITX systems don't accept standard AM5 tower or AIO coolers. Framework Desktop ships with its own integrated cooling solution sized to the board's power envelope, and AMD's configurable-TDP (cTDP) design lets OEMs tune the power ceiling per device rather than exposing one fixed spec across every implementation.

If a socketed cooler comparison is what you're after, that's a question for a traditional Mini-ITX AM5 or AM4 build instead — see how a tower air cooler stacks up against an AIO on a socketed Ryzen chip in Noctua NH-U12S vs MasterLiquid ML240L for a Ryzen 7 5800X.

Motherboards: it's not a drop-in AM5 upgrade

A claim worth correcting directly: Ryzen AI Max chips are not available as standalone parts that fit standard AM5 boards like an ASRock B650M. The B650 chipset family is built for socketed Ryzen 7000/8000/9000 desktop CPUs — a different product line entirely. Ryzen AI Max ships only as a complete board-plus-chip unit in specific products: Framework Desktop (Mini-ITX), the GMKtec EVO-X2 mini PC, and 2-in-1s like the ASUS ROG Flow Z13. There's currently no path to buying a bare Ryzen AI Max chip and installing it in a motherboard of your choice.

Alternative: a traditional discrete Mini-ITX build

For buyers who don't specifically need the unified-memory AI use case, a conventional Mini-ITX build with a socketed CPU and a discrete GPU remains the stronger choice for gaming and general performance-per-dollar, and it keeps the upgrade path open. A Ryzen 5 3600 (/product/B0BL92CBWG, $81.00) or the retail-boxed Wraith Stealth variant (/product/B07STGGQ18, $209.99) paired with a discrete card delivers real, benchmarked gaming performance:

Stepping up, a Ryzen 5 5600X (/product/B08166SLDF, $179.89) or Ryzen 7 3700X (/product/B07SXMZLPK, $314.37) raises the CPU-bound ceiling further, while a budget Ryzen 5 2600 (/product/B07B41WS48, $245.00) covers a lower-cost entry point. Any of these pair with standard, upgradable DDR4 kits — the Crucial 16GB kit (/product/B08C4WV6FT) or 8GB kit (/product/B08C4Z69LN) — something the soldered-memory Ryzen AI Max platform doesn't offer.

For a always-on server angle specifically (rather than gaming), the tradeoffs between a socketed APU and a full discrete CPU are covered in Ryzen 5 5600G vs Ryzen 7 5700X for a 24/7 LLM Server.

Who Ryzen AI Max Mini-ITX actually fits

• Fits well: buyers who want a small, quiet, single-board system to run larger quantized local LLMs without assembling a discrete-GPU rig, and who are comfortable buying memory capacity once at purchase time.
• Fits poorly: buyers prioritizing gaming FPS per dollar, or anyone who wants to upgrade CPU, GPU, or RAM independently later — a socketed Mini-ITX build serves both of those needs better and at a lower entry price.

Citations and sources

• AMD Ryzen AI Max 300 Series product page
• Framework Desktop official product page

This piece is editorial synthesis based on publicly available information. No independent first-party benchmarking is reported.