GMKtec EVO-X2 Review: A 120B Model on Your Desk

The high-end mini-PC market has changed shape in a matter of months. Where the machines being tested not long ago ran 32 GB of RAM and a modest integrated graphics chip, we now see systems built on AMD's Strix Halo platform, originally designed for professional laptops and compact workstations. The GMKtec EVO-X2 is one of the machines pushing the boundaries of what a mini-PC can do.

Inside sits an AMD Ryzen AI Max+ 395 with 16 Zen 5 cores and 32 threads, alongside a Radeon 8060S integrated GPU with 40 RDNA 3.5 compute units — the equivalent of an entry-to-mid-range graphics card built into the processor. That is backed by 128 GB of unified LPDDR5X-8533 memory, of which 64 GB can be allocated as dedicated VRAM. It is that memory which changes everything for running AI locally. The unit assessed here was the 128 GB / 2 TB configuration.

The EVO-X2 stays compact despite what it carries. The aluminium chassis, with a touch of bronze-gold on the sides, evokes a laptop casing but denser. The finish is careful, without excessive RGB — just discreet lighting on the fans, with a dedicated front button to change the lighting mode and reach the fan profiles. It is understated and well judged for a professional desk.

Connectivity is well handled. On the front are USB-A 3.2 Gen 2 ports and a USB4 port at 40 Gbps with DisplayPort Alt Mode and Power Delivery. At the back sit two more USB4 ports, HDMI 2.1, DisplayPort 1.4, 2.5 GbE Ethernet and the power input — enough to drive up to four screens at once. The one regret is the absence of OCuLink, so anyone wanting an external GPU has to go through USB4. The external power brick is a substantial 230 watts, consistent with a machine that can pull up to 140 W at peak, and a VESA mounting kit is supplied.

Getting inside is not as simple as on some siblings. Several screws have to come out, some hidden under the rubber feet, which are glued and can deform when prised off. Once the shell is away, though, the design is careful and the cooling ambitious: a system GMKtec calls Max3.0, with three copper heat pipes, two decently sized CPU fans and a third dedicated to cooling the M.2 SSDs. The main fan is noticeably larger than on most mini-PCs, and that matters — a large fan turning slowly moves more air with less noise than a small one spinning hard.

For expansion there are two M.2 2280 PCIe 4.0 slots taking up to 16 TB of storage. The memory, however, is soldered — inherent to the Strix Halo platform, with no way to change it. At 128 GB that is hardly limiting, even for AI work.

The Ryzen AI Max+ 395 offers 16 Zen 5 cores and 32 threads through simultaneous multithreading, where Intel's Panther Lake chips have one thread per core. That difference weighs heavily in multithreaded work.

In Geekbench 6 it scored 22,009 points multi-core, against 15,665 for the Core Ultra X7 358H and 11,282 for the Minisforum M2 — a massive gain over the Intel Panther Lake machines. In Cinebench 2026 it managed 7,042 points multi-core, nearly 1.6 times the X7. For compilation, 3D rendering, heavy transcoding or containers, the EVO-X2 takes a clear advantage.

Single-core is tighter. Geekbench measured 2,939 points, essentially level with the X7's 2,789, and in Cinebench single-core the AMD is actually slightly behind at 451 against 487. Panther Lake keeps a small edge on single-threaded work, which affects desktop responsiveness and some games. On parallel workloads, though, the Ryzen is in another category.

Thermal behaviour is healthy. Under sustained 32-thread load for 90 seconds the frequency settled on a perfectly flat plateau around 123 per cent of base clock with no thermal drop-off, and the sensor read around 39 °C. The Max3.0 cooling absorbs the chip's 124 W without complaint. Note that these benchmarks were run in the BIOS Balanced mode, not Performance — in maximum Performance mode the multi-core figures should rise another 5 to 10 per cent, so treat them as a floor rather than a ceiling.

Strix Halo's other major distinction is an integrated GPU with 40 RDNA 3.5 compute units, where most rival integrated graphics stop at 12 or 16.

The Radeon 8060S scored 102,798 points in Geekbench Vulkan, roughly 1.6 times the Arc B390 in the EVO-T2S and more than five times the integrated graphics of the Minisforum M2. It delivers performance comparable to a mobile RTX 4060 across many scenarios. In gaming it managed 11,289 points in 3DMark Time Spy, over 60 fps in Spider-Man 2 at 1440p Very High with FSR Quality, over 60 fps in Doom The Dark Ages at 1440p High with FSR Balanced, and 74 fps on average in Cyberpunk 2077 at 1440p Ultra with FSR Quality. With FSR3 frame generation enabled, the most demanding games clear 120 fps without difficulty. For an integrated GPU in a mini-PC, that level is unprecedented.

The EVO-X2 carries 128 GB of LPDDR5X-8533 on a 256-bit, eight-channel bus soldered to the board. CPU and GPU share the same pool with no copying between system RAM and dedicated VRAM, and the GPU can address up to 96 GB for its own calculations.

That is what makes local AI work practical here. Running a 70-billion or even 120-billion-parameter model locally at a usable throughput, without sending anything to an external provider, has until now meant a machine with a professional graphics card costing several times as much.

For context on the numbers: model size is expressed in billions of parameters, and a 3B model fits in roughly 2 to 4 GB at 4-bit quantisation, a 7B in 4 to 6 GB, a 30B in 18 to 22 GB, a 70B in 40 to 50 GB, and a 120B in 55 to 65 GB. Throughput is measured in tokens per second, where a token is roughly a word or word fragment; above 20 tokens per second generation is fluid to read, and below 5 the wait becomes tedious.

Wi-Fi 7 measured a real throughput of 1.26 Gbit/s over iperf3, which is very good and better than wired gigabit. The 2.5 GbE Ethernet port is a step behind the 10 GbE on some siblings but remains ample for most home and office use.

Power is the counterpart to the performance. CPU and GPU share a package envelope of around 130 W rather than adding together; for comparison, the Minisforum M2 with a Core Ultra 7 356H peaks at 70 W, so the EVO-X2 draws double.

On noise it stays controlled given the thermal envelope: 47 dB at one metre under full load, higher than the 37 dB of the EVO-T2S but consistent with an APU pulling twice the current. The noise is also notably muffled — no high-pitched whine, just a steady rush — which makes it acceptable in a working environment.

This is clearly not a mini-PC for everyone, and its price and power profile put it in a specific category.

It makes most sense for AI developers, researchers and professionals who want a local inference workstation, for anyone doing agentic AI, retrieval-augmented generation or local coding assistance, or simply for maximum confidentiality. It also suits content creators who need CPU and GPU power without a tower — the 16 cores devour parallel loads such as video rendering, compilation and encoding, and with 64 GB of VRAM available you can work on 4K HDR video projects or massive compositions without hitting a graphics-memory limit. And it suits mini-PC enthusiasts who want 1080p or 1440p gaming without a discrete card, at a level comparable to a gaming laptop with an RTX 4060 in a package four to five times more compact.

At this level the EVO-X2 competes with the Minisforum MS-S1 Max and the Framework Desktop, which share the same chip. All three deliver comparable performance once the BIOS is configured properly — the gap measured against the MS-S1 Max came from its 2 GB default VRAM carveout forcing models onto the CPU, and vanished once that was set to 64 GB. The choice therefore comes down to physical format, cooling, noise, modularity and who it is aimed at. For anyone wanting an AI workstation in a compact format with excellent cooling, the EVO-X2 is among the best options available.

A UK price note worth reading before you buy: the 128 GB / 2 TB configuration assessed here was £3,666.66 on Amazon UK on 2 August 2026, while the 64 GB / 1 TB version was £2,099.96. GMKtec also sells the machine directly from its own store, where the base configuration is listed considerably below the Amazon figure — so check both before ordering. Check the price on Amazon UK or search on Amazon US