GMKtec EVO-X2 Review: AMD's Ryzen AI Max+ 395 Is the £2,099 Mini PC That Runs 70B Models Locally

The Verdict

The GMKtec EVO-X2 is the first Windows mini PC that puts 96GB of on-package unified memory in front of an AMD Ryzen AI Max+ 395, and at £2,599.96 on Amazon UK it is category-creating rather than category-competing. It is simultaneously a compact gaming rig that holds its own at 1440p Medium-High, a silent-ish quad-monitor workstation, and — crucially — a local-LLM box that hosts Llama 3.3 70B at Q4 in addressable memory without spilling to disk. Nothing else in the Windows ecosystem at this price does all three at once.

Check Current Price →

---

What Ryzen AI Max+ 395 Actually Is

Strix Halo is AMD’s answer to the question Apple’s M-series has been asking for four years: what happens when you stop treating CPU, GPU and memory as three separate things connected by a bus, and instead fuse them onto a single package with wide, on-die memory channels? The Ryzen AI Max+ 395 is a laptop-class APU that has been extracted from its thin-and-light chassis and dropped into a mini PC form factor where it can actually be cooled properly. Sixteen Zen 5 cores. Thirty-two threads. A 40-CU RDNA 3.5 iGPU running at up to 2.9 GHz. An XDNA 2 NPU for AI acceleration. And 96GB of LPDDR5X-8000 soldered to the package in an eight-channel configuration that no socketed DDR5 system can match on latency or effective bandwidth.

The maths matter here. Apple’s M4 Max with its unified memory architecture clocks in around 400 GB/s of memory bandwidth, which is the number Apple has been weaponising in keynotes since the M1 Ultra launched. Strix Halo is less — roughly 256-270 GB/s depending on workload — but it is dramatically more than any conventional x86 desktop with two DDR5 DIMM slots, and it sits inside the Windows ecosystem with full x86-64 software compatibility. You do not need to recompile anything. You do not need to wait for ARM ports. Your Steam library runs. Your CUDA-free PyTorch scripts run. Your llama.cpp build runs. The 40-CU iGPU is, by Amazon’s own listing text and by independent benchmark consensus, positioned between an RTX 4060 Laptop and an RTX 4070 Laptop in real-world gaming performance — which is to say it is a proper GPU, not a token integrated afterthought.

---

Core Specifications

Spec	GMKtec EVO-X2
APU	AMD Ryzen AI Max+ 395 (Strix Halo)
CPU cores / threads	16 Zen 5 cores / 32 threads
Boost clock	Up to 5.1 GHz
L3 cache	64MB
iGPU	AMD Radeon 8060S (RDNA 3.5), 40 CUs
iGPU clock	Up to 2.9 GHz
NPU	XDNA 2 architecture (on-die)
Memory	96GB LPDDR5X-8000, eight-channel, on-package (8x 12GB, non-upgradable)
Storage	2TB PCIe 4.0 NVMe SSD
Display outputs	1x HDMI 2.1 (8K@60Hz), 1x DisplayPort 1.4 (4K@60Hz), 2x USB4 40 Gbps (DP Alt Mode 4K@60Hz each)
Max simultaneous displays	Four
Networking	2.5GbE Ethernet, Wi-Fi 7 (802.11be, up to 46 Gbps theoretical), Bluetooth 5.4
Cooling	Triple-fan (dual CPU turbo fans + dedicated DDR5/SSD fan), three heatpipes, 360° airflow
Claimed noise (Quiet Mode)	35 dB(A)
Lighting	13 RGB modes
Operating system	Windows 11 Pro (licensed)
Form factor	Mini PC
Warranty	GMKtec 3-year limited (1-year replacement, 2-year repair)
Amazon UK price	£2,599.96

Two things stand out in that table once you read it properly. The eight-channel LPDDR5X-8000 configuration is what makes Strix Halo tick — it is the reason the iGPU performs like a mid-range discrete mobile GPU despite having no dedicated VRAM. And the memory is on-package, soldered, permanent. What ships in the box is what you have for the lifetime of the machine. There is no future upgrade path to 128GB or 192GB. You commit to 96GB now, forever.

---

Gaming Performance: Better Than You Expect

Nobody is buying an integrated graphics mini PC to play Cyberpunk at 4K Ultra with path tracing. That is not the conversation. The conversation is whether the Radeon 8060S can deliver playable, visually acceptable performance on modern AAA titles at resolutions and settings most players actually use — and the answer is yes, emphatically, with caveats that favour it over equivalent discrete hardware in a meaningful subset of cases.

Real-world benchmarks from the early review cohort land roughly where Amazon’s own positioning suggests — between an RTX 4060 Laptop and an RTX 4070 Laptop:

• Cyberpunk 2077, 1440p High, FSR Quality: ~52 fps avg
• Cyberpunk 2077, 1080p High, FSR Quality: ~78 fps avg
• Call of Duty: Warzone, 1440p High, FSR Quality: ~118 fps avg
• Baldur’s Gate 3, 1440p High (native): ~74 fps avg
• Hogwarts Legacy, 1440p Medium: ~65 fps avg
• Elden Ring, 1440p High: ~71 fps avg (capped at 60 by engine but averages with uncapped mods)
• Red Dead Redemption 2, 1440p High: ~68 fps avg

Frame that honestly. This is 1080p High to 1440p Medium-High territory for modern AAA. Older titles and competitive esports games — CS2, Valorant, Apex, Rocket League — fly. If your expectation is 4K Ultra with ray tracing maxed, you are buying the wrong machine; a desktop with an RTX 5080 is your answer. If your expectation is “I want this to play the games I actually play at a resolution and framerate that does not embarrass itself,” the 8060S clears that bar without strain.

There is a quiet advantage hiding in the architecture that modern gaming discussions keep missing. The 40-CU iGPU shares the 96GB LPDDR5X pool with the CPU. In 2026, discrete GPUs with 8GB of VRAM are starting to choke on modern AAA textures — Hogwarts Legacy, Alan Wake 2, The Last of Us Part I all famously punish 8GB cards at high texture settings with stutter, pop-in and outright crashes. The EVO-X2 does not care. Its texture budget is measured against the entire 96GB pool minus what the OS and applications are using, which typically leaves well over 32GB of de facto graphics memory even under normal desktop load. VRAM-bound titles that kneecap a 4060 Laptop’s 8GB simply do not kneecap this machine.

---

Local LLM Performance: The Real Story

This is where the EVO-X2 goes from “interesting mini PC” to “the reason this review exists.” For the first time in the Windows ecosystem, at a price point south of £2,100, you can host frontier-adjacent large language models entirely in addressable memory without the performance collapse that happens on discrete-GPU machines when the model exceeds VRAM.

Measured inference performance on shipping hardware, using llama.cpp with the Vulkan backend (out-of-the-box install) and with tuned ROCm where the brackets diverge:

• Llama 3.3 70B Q4_K_M (42GB on disk): fits comfortably with generous room for OS, context window, and KV cache. ~11 tokens/sec on Vulkan. ~18 tok/sec after ROCm tuning and batch optimisation.
• Mixtral 8x7B Q5_K_M (33GB): ~22 tok/sec. This is the practical sweet spot — fast enough for conversational agent work.
• Mixtral 8x22B Q3_K_S (63GB): fits. ~6 tok/sec. Slow, but it runs, which on a 24GB RTX 5090 laptop it simply does not.
• Qwen 2.5 72B Q4_K_M (44GB): ~10 tok/sec.
• Llama 3.3 405B Q2: does not fit (200GB+ footprint). Wait for Strix Halo’s successor or buy a Pro 6000.

The XDNA 2 NPU accelerates a subset of quantised inference paths when software targets it — ONNX Runtime with the AMD execution provider, DirectML workloads, and a growing list of AMD-optimised pipelines. Treat it as bonus upside rather than primary compute for LLMs in 2026. The ecosystem is maturing fast but it is not yet where CUDA was in 2020. By 2027 this picture is likely to look materially better.

Here is the comparison that matters. An RTX 5090 laptop has 24GB of VRAM. Llama 3.3 70B at Q4 is 42GB on disk and materially more once you account for context. It does not fit. When it does not fit, llama.cpp spills the overflow into system RAM across the PCIe bus, and inference speed collapses by an order of magnitude — from tens of tokens per second to low single digits. On the EVO-X2, the entire 42GB model lives in the unified memory pool with the iGPU able to address it directly. No PCIe transit, no offload heuristics, no thrashing.

This is the Apple M-series pitch — unified memory as the architecture that unlocks local AI at the desktop — brought over to Windows at a dramatically lower price. A Mac Studio with comparable memory capacity starts north of £3,500 once you spec it properly. The EVO-X2 at £2,599.96 is offering the same category of workload for roughly £900 less, inside the x86 ecosystem where your existing Windows-native tooling already runs.

---

Thermals, Noise, Cooling Design

Cramming a 16-core Zen 5 APU with a 40-CU iGPU into a mini PC chassis requires a cooling solution that most small-form-factor boxes simply do not have. GMKtec’s answer is a triple-fan layout — dual turbo CPU fans running the primary heatsink, plus a dedicated third fan pulling air across the DDR5 memory and NVMe SSD area. Three copper heat pipes distribute load from the APU die. The chassis is designed for 360° airflow, meaning intake and exhaust are distributed rather than concentrated on a single face.

Claimed Quiet Mode noise is 35 dB(A). Reviewer measurements from the first wave of units land at roughly 34-36 dB at idle — consistent with GMKtec’s claim — ramping to roughly 46 dB under sustained full-load gaming or extended LLM inference. That is not whisper-silent the way a fanless Mac mini is, but it is quieter than a typical tower gaming PC under load and materially quieter than a mid-range gaming laptop at maximum fan. For context, 46 dB is somewhere between a quiet conversation and a domestic fridge humming — noticeable if your desk is in a bedroom, easily ignored in a home office.

The DDR5/SSD fan is the smart bit. Strix Halo’s on-package memory runs hot under sustained inference workloads because it is being hammered at near-peak bandwidth for extended periods, which is not a typical consumer PC thermal profile. Most mini PCs cool the CPU and let the memory thermal-throttle when it becomes a problem. GMKtec has designed for the workload this machine will actually run.

---

Display Setup: Quad 8K Is Real

The display output configuration reads like marketing fiction until you count the ports. One HDMI 2.1 at 8K@60Hz, one DisplayPort 1.4 at 4K@60Hz, and two USB4 ports at 40 Gbps with DisplayPort Alt Mode capable of 4K@60Hz each. That is four independent display outputs on a machine the size of a hardback book.

In practice the “quad 8K” pitch is theoretical — you can drive four 8K-capable displays simultaneously but you will never feed them with native 8K content across the board. What matters is the practical configuration the port layout unlocks: two 4K monitors at 120Hz plus a third vertical 1440p reference panel, say, or one 4K primary plus two 1440p secondaries plus a small portrait strip for chat/logs/vitals. For a developer, researcher, or creator with a multi-monitor setup, this is a proper workstation output profile in a chassis that will sit behind a single 27-inch screen with room to spare.

The dual USB4 ports also double as high-bandwidth data ports for external Thunderbolt-compatible storage and docks, so you are not giving up I/O to get the display count. Wi-Fi 7 and 2.5GbE round out the connectivity story for a machine that increasingly reads like a proper workstation rather than a toy.

---

Where the Compromises Live

The honest section, because every mini PC review that skips this is lying to you.

Memory is not user-upgradable. Ever. The 96GB LPDDR5X is soldered onto the APU package because that is how Strix Halo achieves its bandwidth numbers. What ships in the box is what you have for the life of the machine. If you want 128GB or 192GB in two years, you buy the next generation — Strix Halo’s successor, or whatever AMD brings to market in 2027. Commit to 96GB as a permanent ceiling or walk away.

There is no discrete GPU upgrade path. This is a sealed mini PC, not a tower. You cannot drop in an RTX 5080 later to boost gaming performance. The iGPU is the GPU, full stop. If your expectation is “buy the mini PC now, add a real GPU later when I have more budget,” adjust that expectation — this machine is its final form on day one.

The SSD is user-upgradable — the 2TB NVMe ships in a standard M.2 slot. The listing and teardowns suggest a second M.2 slot exists on the board for expansion, but confirm via the physical unit or a community teardown video before counting on it for your particular configuration.

Windows 11 Pro licence is included. This is not a complaint, it is the opposite — a surprising number of mini PC brands in this price bracket ship Home and force you to upgrade to Pro at your own expense. GMKtec has included Pro, and that is worth roughly £80-100 of licence value baked into the sticker price.

AMD’s ROCm stack on Windows is less mature than NVIDIA CUDA. This is the elephant in the AI-workstation conversation. For inference via llama.cpp with the Vulkan backend, things work out of the box and work well. For PyTorch with full ROCm support, for fine-tuning workflows, for the long tail of AI tooling that assumes CUDA, you will want to use WSL2 where ROCm support is materially better than native Windows. If your workflow is CUDA-dependent for anything critical, this is not your machine.

GMKtec is a Shenzhen-based brand. The 3-year warranty is real and reviewers have had positive RMA experiences, but logistics mean shipping to China for repair if your claim falls outside the consumer-law 2-year period your UK retailer covers. Amazon’s 30-day returns window is the practical safety net for early-life failures.

---

What’s Great / What’s Not

What’s great:

• 96GB of unified LPDDR5X-8000 memory for sub-£2,100 is simply unmatched in the Windows ecosystem — no direct competitor at the price point
• iGPU gaming performance consistent with RTX 4060 to RTX 4070 Laptop tier — legitimately playable modern AAA at 1080p High / 1440p Medium-High
• 2TB Gen4 NVMe SSD shipped by default, not as a £200 upgrade option
• Quad-display support including 8K@60 via HDMI 2.1 — a real workstation I/O profile
• Wi-Fi 7, 2.5GbE Ethernet, Bluetooth 5.4 — fully current networking stack
• Windows 11 Pro licence included, not a Home-then-upgrade trap
• 3-year warranty with 1-year replacement plus 2-year repair coverage
• Compact, quiet under light load, silent-ish under sustained load (relative to tower PCs)
• XDNA 2 NPU on-die — not critical today, will matter more as Windows AI tooling matures

What’s not:

• Memory is soldered and non-upgradable — 96GB is your permanent ceiling on this chassis
• No discrete GPU upgrade path — sealed mini PC form factor
• AMD ROCm / DirectML AI stack on Windows is less mature than NVIDIA CUDA
• GMKtec brand RMA logistics involve shipping to China beyond the 2-year UK consumer protection window
• Fan noise at sustained full load (~46 dB) is noticeable in a quiet room
• Keyboard, mouse and monitor not included — budget for peripherals on top
• First-generation Strix Halo product means early BIOS and driver quirks are real; expect updates in the first six months

---

Who Should Buy This

No equivocation. Direct lists.

Yes, buy this if you are:

• A local-LLM enthusiast or researcher who wants 70B-class models running at home, in addressable memory, without paying API bills or cloud GPU rental fees
• An AI/ML developer prototyping on frontier-adjacent model sizes where VRAM ceilings on consumer GPUs are the actual bottleneck
• A compact desktop gamer who is happy at 1080p High to 1440p Medium-High and not chasing 4K Ultra with ray tracing
• A developer who wants a quiet-ish Windows workstation in a shoebox with a proper Zen 5 CPU core count for compilation, containers and local development work
• Running a small studio with a multi-monitor 4K setup where a sealed workstation with a permanent memory budget is a feature, not a limitation
• Tired of paying £3-5/hour for cloud GPU instances to experiment with local LLM inference — the EVO-X2 pays itself off in months of saved cloud spend for anyone iterating seriously
• Comparing against an Apple Mac Studio at similar memory specs and realising you want to stay inside the Windows x86 ecosystem

No, do not buy this if you are:

• A 4K Ultra AAA gamer chasing max settings with ray tracing — a desktop with an RTX 5080 or 5090 is the honest answer
• Locked into NVIDIA CUDA-only software stacks, whether that is DaVinci Resolve’s NPU acceleration paths, Adobe-specific CUDA workloads, or any research codebase that assumes a CUDA runtime
• A small-form-factor enthusiast who wants to build a custom RGB rig with bespoke cooling — this is a sealed OEM mini PC, not a DIY platform
• A competitive esports player targeting 240Hz+ at 1440p or 1080p Ultra — a cheaper machine with a discrete mid-range GPU will serve you better
• Someone whose workload genuinely benefits from 128GB+ of memory — wait for the next Strix Halo generation or buy a threadripper workstation

If you fit the first list, the EVO-X2 at £2,599.96 is an obvious buy that solves three problems at once — local LLM hosting, compact gaming, quiet workstation. If you fit the second list, your money goes further elsewhere and the unique proposition of Strix Halo is not the one you need.

---

Where to Buy

GMKtec’s direct sales channel occasionally runs bundle discounts with peripherals — keyboards, mice, monitor offers — which can sweeten the total package if you need the kit anyway. Amazon UK’s advantage is the 30-day returns window, which matters meaningfully on a first-generation Strix Halo product where early BIOS-level quirks on sleep, wake and ROCm compatibility are genuinely being ironed out in the wild. Expect several firmware updates in the first six months of ownership; GMKtec has a track record of shipping them regularly and the early community reports suggest the cadence is holding.

---

Benchmarks and specifications cross-referenced against the Amazon UK product listing (ASIN B0FR17BVSL), the early reviewer cohort covering Strix Halo mini PCs, and community LLM inference reports from the llama.cpp and LocalLLaMA communities running on shipping Ryzen AI Max+ 395 hardware. Affiliate links support independent testing — we buy the machines we cover.