GMKtec EVO-X3

Published on August 29, 2026

The GMKtec EVO-X3 stands the company’s Strix Halo mini PC on its edge. Where the EVO-X2 was a squat flat box, the X3 is a 13.9-inch tall metal slab barely an inch and a half thick, sitting upright in a flat metal foot, with the whole narrow top edge given over to a perforated lime-green intake cap. The silicon underneath is unchanged: the same AMD Ryzen AI Max+ 395 with 16 Zen 5 cores, 32 threads, a Radeon 8060S packing 40 RDNA 3.5 compute units, and an XDNA 2 NPU that GMKtec rates at 50 TOPS on its own and 126 TOPS across the whole system. What changed is the box around it, and the port cluster tells the story: the X3 adds an OCuLink connector for a genuine PCIe Gen4 x4 external GPU link, and drops the X2’s DisplayPort output, its SD card reader, and its second 2.5GbE jack to make room. Memory is fixed at 128GB of LPDDR5X-8000 soldered to the board, of which up to 96GB can be assigned to the integrated GPU.

Pros and Cons of the GMKtec EVO-X3

Related Videos

GMKtec EVO-X3 Comparison Chart

	GMKtec EVO-X3	GMKtec EVO-X3
Price	List Price: $3,799.99 Amazon Prices: Loading prices...	List Price: $3,999.99 Amazon Prices: Check Price on Amazon
Version	128GB / 2TB	128GB / 4TB
Performance Rating	8.5	8.5
Operating System	Windows 11 Pro	Windows 11 Pro
Processor	Sixteen-core 3.00 Ghz (max 5.10 Ghz) AMD Ryzen AI Max+ 395	Sixteen-core 3.00 Ghz (max 5.10 Ghz) AMD Ryzen AI Max+ 395
GPU	Integrated AMD Radeon 8060S	Integrated AMD Radeon 8060S
RAM	128 GB LPDDR5X (128GB LPDDR5X-8000 unified, soldered)	128 GB LPDDR5X (128GB LPDDR5X-8000 unified, soldered)
Internal Storage	2 TB NVMe SSD	2 TB NVMe SSD
Dimensions width x length x thickness	13.9 x 7.32 x 1.61 inches (353.06 x 185.93 x 40.89 mm)	13.9 x 7.32 x 1.61 inches (353.06 x 185.93 x 40.89 mm)
Weight	5.07 lbs (2.3 kg)	5.07 lbs (2.3 kg)
WiFi	Wi-Fi 7 (802.11be)	Wi-Fi 7 (802.11be)
Bluetooth	Bluetooth 5.4	Bluetooth 5.4
Ethernet	1 Ethernet port at 2.5 Gbps	1 Ethernet port at 2.5 Gbps
HDMI	1 Full-Size HDMI Port	1 Full-Size HDMI Port
DisplayPort	No native DisplayPort; video out via HDMI 2.1 and USB4	No native DisplayPort; video out via HDMI 2.1 and USB4
VGA	No VGA Ports	No VGA Ports
USB Ports	1 USB 2.0, 2 USB 3, 1 USB 4 Rear: 1x USB4 Type-C (40Gbps, video out), 1x USB-A 3.2 Gen2. Front: 1x USB-A 3.2 Gen1, 1x USB-A 2.0.	1 USB 2.0, 2 USB 3, 1 USB 4 Rear: 1x USB4 Type-C (40Gbps, video out), 1x USB-A 3.2 Gen2. Front: 1x USB-A 3.2 Gen1, 1x USB-A 2.0.
Thunderbolt Ports	No	No
OCuLink	PCIe Gen4 x4 for an eGPU; no hot-plug	PCIe Gen4 x4 for an eGPU; no hot-plug
Internal SATA Ports	No SATA ports	No SATA ports
Card Reader	No Card Reader	No Card Reader
Headphone Jack	combo	combo
Fanless	No	No
VESA Mount	No	No
In the Box	Mini PC, vertical stand and screws, rubber feet, HDMI cable, power adapter, manual	Mini PC, vertical stand and screws, rubber feet, HDMI cable, power adapter, manual
Expandability	128GB LPDDR5X soldered (not upgradeable). Dual M.2 2280 PCIe 4.0 x4 slots, 16TB max. OCuLink PCIe Gen4 x4 for an external GPU.	128GB LPDDR5X soldered (not upgradeable). Dual M.2 2280 PCIe 4.0 x4 slots, 16TB max. OCuLink PCIe Gen4 x4 for an external GPU.

Detailed Insights into the GMKtec EVO-X3

The chassis measures roughly 13.9 by 7.3 by 1.6 inches without the stand and weighs about five pounds. That is a strange shape for a mini PC and a deliberate one: standing upright, the chassis presents only its 1.6-inch edge to the desk, while the tall interior gives GMKtec room for a cooler it could not fit in the older flat-box design. GMKtec’s spec table lists triple heat pipes and three fans, though ETA PRIME described the cooler as a dual-fan design. Level1Techs, handling the machine, described air being drawn in through the top and exhausted out the back, with a further fan on the side panel. The body is CNC-machined metal in gray with the green accents, and a fine dust filter mesh sits behind the vent slots.

Port selection is the sharpest departure from the EVO-X2, and it cuts both ways. The front face carries a USB-A 3.2 Gen1 port, a USB-A 2.0 port, a 3.5mm combo audio jack, and the power button. The back holds HDMI 2.1 rated for 8K at 60Hz, a full-function USB4 Type-C port running 40Gbps with video output, one more USB-A 3.2 Gen2, a single 2.5GbE jack on a Realtek RTL8125BG controller, the barrel DC input, and the OCuLink connector. There is no DisplayPort, no SD card reader, and no second Ethernet port, all three of which the EVO-X2 had. Wireless is Wi-Fi 7 and Bluetooth 5.4. That leaves two video outputs total, so multi-monitor setups past a pair of displays need a USB4 dock rather than the machine’s own ports.

The OCuLink connector is the addition that most changes what the machine can do. It carries PCIe electrically rather than tunneling it over USB, so a four-lane Gen4 link runs at the full 8 GB/s in each direction with none of the protocol overhead a Thunderbolt or USB4 eGPU enclosure adds. On a chip like Strix Halo, whose integrated graphics are strong but whose memory bandwidth is the ceiling on local inference, an external card with its own fast VRAM is a real upgrade path rather than a curiosity. The cost is usability: OCuLink is a server-side connector, GMKtec’s own documentation says the machine must be powered off before connecting or disconnecting, and there is no USB fallback if the dock misbehaves.

Both shipping configurations use the same 128GB of LPDDR5X-8000, differing only in whether the pre-installed drive is 2TB or 4TB. Storage sits in two M.2 2280 PCIe 4.0 x4 slots that accept up to 8TB each, though hitting the 16TB ceiling means pulling the pre-installed drive rather than just filling the empty slot. Power modes are exposed in firmware as silent, balanced, and performance, rated by GMKtec at 54W, 85W, and a 140W peak, and the BIOS also carries a manual fan curve and a VRAM allocation setting that goes up to 96GB, with AMD’s Adrenalin software offering finer steps down to half a gigabyte.

The chip itself deserves context, because Strix Halo parts have been shipping since early 2025 and the EVO-X3 launched in mid-2026 on the same silicon. AMD’s Ryzen AI Max 400 series, codenamed Gorgon Halo, arrived in the interim, and Starry Hope’s read is that Gorgon Halo is a rename rather than a generational jump: same Zen 5 cores, same RDNA 3.5 graphics, same XDNA 2 NPU, with a 192GB memory ceiling as the one genuine addition. Tom’s Hardware, covering the EVO-X3’s announcement in June 2026, noted GMKtec stayed on the 395 rather than moving to the 495 and reported the entry price as far steeper than the EVO-X2’s. That is the tension in this machine: the redesign is substantial, the silicon is a year and a half old, and the asking price went up rather than down.

Reviewer Insights on the GMKtec EVO-X3

ETA PRIME

ETA PRIME ran the EVO-X3 through a full first-look in Windows and returned to it later with SteamOS installed. His opening read on the silicon was blunt: “The new X3 is powered by the AMD Ryzen AI Max Plus 395. We’ve seen this before. It’s an absolute monster.” He set the machine to performance mode in the BIOS, split the 128GB pool evenly between system memory and VRAM, and watched the package power hit GMKtec’s rated peak before settling well below it. “It’s got that 140 W boost, but it kind of sustains at around 115, 116,” he said. He also noted a gap in the firmware: the AMD CBS menu is absent, so the TDP cannot be raised manually from the BIOS as it can on some competing Strix Halo boxes.

His measured numbers land where the platform usually does. Cinebench R24 returned 114 single-core and 1,893 multi-core; Geekbench 6 came in at 3,081 single and 22,314 multi; 3DMark Time Spy scored 11,388, against a desktop RTX 4060 he placed around 10,300 to 10,600 and whose GPU alone draws up to 115W. A 70-billion-parameter DeepSeek R1 model ran at three tokens per second on his setup, slower than an M4 Max, but the other machines in his comparison could not finish the test at all because the model would not fit in their memory. Gaming at 1440p held up: Forza Horizon 6 averaged in the low 90s at ultra with no upscaling, Cyberpunk 2077 landed around 74 with FSR 2.1 at quality, and Black Myth Wukong sat near 83.

Thermals and acoustics were the surprise. He recorded a 66C average APU temperature during 1440p gaming and an 83C maximum across all of his testing, and concluded that “This little cooling system here is doing a really good job, and it never sounds like a jet engine.” Total system draw from the wall measured 11W at idle, 18W during 4K video playback, and around 147W in AAA gaming, all in performance mode. He liked the form factor for what it is, saying “I’m a sucker for these mini PCs with vertical stands,” but landed on the same complaint the port list invites: “I think they could have added a little more IO around back here.”

The SteamOS follow-up is where the gaming case gets stronger. Running official SteamOS 3.9 rather than the shipped Windows install, he reported better frame rates than he had previously measured on similar Strix Halo hardware and credited recent SteamOS and Mesa driver work. “In the past, we’ve tested with a very similar system, but I never saw the kind of frame rates that we’re getting out of this thing,” he said, adding that “this thing is ripping through games at 1440p. It’s actually really impressive what this thing’s doing.” Spider-Man 2 averaged in the high 80s at 1440p very high with FSR at quality, Elden Ring held a steady 60 maxed out with no upscaling, and Forza Horizon 6 ran near 78 at 1440p ultra. He did not oversell it: the iGPU still runs into a wall on VRAM-heavy titles, since the shared LPDDR5X cannot match dedicated GDDR6 bandwidth, and he was blunt about the price: “We’ve seen absolutely ridiculous prices on these PCs powered by the CPU, and this one’s really no different.”

Level1Techs

Wendell at Level1Techs approached the EVO-X3 as an inference platform rather than a gaming box, and spent most of his time on the parts that decide how it performs on language models. His AIDA64 run measured 118 GB/s of memory bandwidth with latency near 144 nanoseconds, and he put that figure in context: “That is about as much as a quad-channel Threadripper system.” The unit he received used Micron memory running at 8000 MT/s and shipped with a Crucial P310 2TB PCIe Gen4 drive, a component choice he approved of by name.

He was careful about what the 140W ceiling does and does not buy. Token generation on Strix Halo is bound by memory speed rather than compute, so, as he put it, “the 140 W is not going to help you there. It’ll help you with prompt processing speed.” On the CPU side he rated the 16 Zen 5 cores above the 20 cores in NVIDIA’s DGX Spark for general compute work. In LM Studio on Windows he reported 40 to 60 tokens per second on 70-billion-parameter models at four and six-bit quantization, and close to 60 tokens per second running a Qwen 3.6 quant through AMD’s Lemonade stack.

His criticisms were specific. The single Realtek 2.5GbE port drew the sharpest line: “2 and 1/2 gigabit in this day and age, GMK Tech didn’t learn from Evo X2.” The MediaTek RZ717 Wi-Fi 7 radio runs 160MHz channels rather than 320MHz. And the default thermal behavior needs attention, since “the balanced configuration can leave a little bit to be desired on thermals” and, in the fan menu, “By default, it will not go the full fan speed until you hit 95. That’s too hot for me, so I like to adjust this.” With a more aggressive curve set, he found the acoustics reasonable: “this thing basically stays inaudible in almost every scenario. When you’re using more than 100 watts, you can definitely hear fan noise. It’s more whooshing of air than high-pitched whine.”

OCuLink was what drew him to the machine. He connected both an NVIDIA RTX 5060 Ti and an AMD R9 700 through an OCuLink dock, mixing CUDA and ROCm runtimes on one host, and explained why the connector matters: “This is truly four lanes of PCIe Gen 4 because it is physically PCIe Gen 4.” He was equally clear about who it is for, calling it a server connector and warning that “It’s not a connector that’s meant for mass consumption.” On value, he placed the EVO-X3 at the low end of the Strix Halo market rather than the high end, describing it as “one of the least expensive Strix Halo machines you can get” against comparable 128GB configurations.

Conclusion

The EVO-X3 makes sense for one buyer in particular: someone who wants 128GB of unified memory for local model work and wants a real expansion path off the box when the integrated GPU runs out of room. OCuLink is the feature that justifies the redesign, and no amount of USB4 tunneling matches four physical lanes of PCIe Gen4. The cooling appears to answer the fan-noise complaint Tom’s Hardware raised against the EVO-X2, since both reviewers described a machine that stays quiet until it is genuinely working hard, and the full-metal vertical chassis is a clear step up in build.

The case against it is just as clear. The memory is soldered, so the configuration is permanent, and there is only one configuration on Amazon. The rear I/O is thin for a machine this size, with one Ethernet port where the previous generation had two, no DisplayPort, and no card reader. The Wi-Fi radio is a half-width Wi-Fi 7 implementation. And the chip is the same Ryzen AI Max+ 395 that has been shipping since early 2025, at a price well above what the EVO-X2 launched at, which means a lot of what you are paying for is the chassis and the OCuLink port rather than new silicon.

If you are weighing this against the rest of the platform, Starry Hope’s guide to running large language models on Strix Halo covers where the real bottlenecks sit, and the Mini PC comparison chart puts the EVO-X3 next to every other box in the catalog. Buyers who mainly want the memory pool and can live without OCuLink should compare it directly against the EVO-X2, which offers more memory tiers and a second 2.5GbE port at a lower position in the range.

Frequently Asked Questions