# GMKtec EVO-X3 Tested by Wendell: The Fastest Strix Halo Yet, and Its Unholy OcuLink Trick > Source: > Published: 2026-08-08 14:15:37+00:00 The 128GB unified-memory tier is where local AI's action is right now, and GMKtec just reshaped its best-known box. The EVO-X3 takes the same Ryzen AI Max+ 395 platform as the [EVO-X2 we covered](https://vettedconsumer.com/gmktec-evo-x2-guide-the-sub-1-500-mini-pc-that-runs-70b-models-locally/) and rebuilds it as a tower with one genuinely new trick: an OcuLink port. Wendell of Level1Techs spent a review getting, in his words, "up to no good" with it, plugging NVIDIA and AMD GPUs into the same Strix Halo machine and using a local model to debug his own cluster. His verdict on the box itself: "Is this one of the fastest Strix Halo machines that I've tested? Yes." ## What Wendell found in the box The core is familiar Strix Halo: Ryzen AI Max+ 395, Radeon 8060S, and 128GB of Micron LPDDR5X-8000 in one pool. His AIDA64 run measured **118 GB/s of real memory bandwidth** at a not-fabulous 143.8ns latency, "about as much as a quad-channel Threadripper system" from a small tower. Two spec choices push it to the top of the Strix pack: DDR5-8000 memory and a **140W CPU power limit**, the ceiling of what the chip supports. His caveat on that 140W is the single most useful sentence in the review, and it matches everything we have written about [bandwidth-bound decode](https://vettedconsumer.com/bandwidth-not-tflops-what-sets-your-local-llm-speed-and-why-the-newest-card-isnt-always-fastest/): "for AI performance like inferencing, token generation, it's not a limitation of the CPU. It's the memory speed. So the 140W is not going to help you there. It'll help you with prompt processing speed." Extra watts buy faster prefill and compute, not faster generation. On the compute side he adds that the 16 Zen 5 cores "run circles around the 20 cores that you get in the DGX Spark for CPU type operations." Measured output where it counts: running AMD's new Lemonade stack (the one-install path to the platform's GPU and NPU), he showed Qwen 3.6 generating at **nearly 60 tokens per second** on Windows. He also notes a quiet platform win: the newest Windows 11 builds handle unified memory properly (a fix Microsoft shipped while preparing for NVIDIA's Windows-on-ARM RTX Spark), so you no longer need to pre-carve the 128GB into "system" and "VRAM" halves. The days of the 64/64 BIOS split are ending. Housekeeping notes from the review: the tower cooling stays "basically inaudible" under 100W, but the default BIOS fan curve only maxes at 95C ("too hot for me," and he shows his more aggressive settings); the 2.5GbE Realtek NIC is a letdown he already flagged on the X2; the MediaTek Wi-Fi 7 radio is the 160MHz variant; SR-IOV is exposed in BIOS; and the Windows preinstall doesn't force a Microsoft cloud account if you set up offline. ## The OcuLink trick: CUDA and ROCm in one machine The X3's real differentiator over the X2 is the OcuLink port: four real lanes of PCIe Gen 4, 8GB/s each way, no protocol overhead. As Wendell puts it, it's "a kludge... not meant for mass consumption," and it is also the fastest external GPU path a Strix Halo machine has. (Standard Strix PCIe slots only deliver 25W of power, which is why eGPU docks solve a real problem here.) What he did with it borders on the unholy, his word: a 5060 Ti 16GB (NVIDIA, CUDA) and an ASUS Turbo R9700 32GB (AMD, around $1,000 to $1,200) attached to the same 128GB Strix box, with LM Studio's multi-runtime support juggling the backends. The logic is the split our [prefill-vs-decode explainer](https://vettedconsumer.com/prompt-processing-vs-generation-why-your-box-is-fast-at-one-and-slow-at-the-other/) describes: a fast discrete GPU handles the compute-heavy parts while the 128GB pool holds the big sparse models, the same division of labor behind his observation that Mac-cluster builders "have been putting things like NVIDIA's DGX Spark in front of their Mac cluster so the DGX Spark can do the prompt processing." The part of the review that feels like a preview of 2027 homelabs: his X3 spends its days running a Qwen 3.6 agent equipped with custom tools that build, tear down, and benchmark the Thunderbolt-RDMA links between his other Strix Halo machines. His framing of small local models is worth quoting against the hype: "It's not a frontier model. It doesn't have a ton of intelligence... But what it can do really well is run tools." The clustering write-ups land on the [Level1Techs forum thread](https://forum.level1techs.com/t/gmktec-evo-x3-ai-workstation-lemonade-benchmarks-and-specs/253262?ref=vettedconsumer.com), along with his full benchmark tables. ## The price situation, dated carefully This is where the X3 needs context, because two true-sounding claims point in opposite directions. At launch in early July, coverage was rough: [VideoCardz reported](https://videocardz.com/newz/gmktec-evo-x3-with-ryzen-ai-max-395-to-launch-at-3600-comes-with-128gb-memory?ref=vettedconsumer.com) a $3,600 starting price (128GB/2TB), roughly double what the EVO-X2 launched at, and [Notebookcheck called it](https://www.notebookcheck.net/GMKtec-launches-repackaged-Strix-Halo-mini-PC-with-2x-higher-price.1335259.0.html?ref=vettedconsumer.com) a "repackaged Strix Halo mini PC with 2x higher price." Wendell's counterpoint in this review is that against *same-configuration* competitors in today's memory-crisis market, it's "over a thousand dollars less," and per [VideoCardz](https://videocardz.com/newz/gmktec-pc-with-max395-and-128gb-costs-400-less-than-amd-ryzen-ai-halo?ref=vettedconsumer.com) it undercuts AMD's own Ryzen AI Halo by about $400. Both readings are right: the X3 is expensive against early-2026 Strix prices and cheap against what 128GB boxes cost now. The wrinkle is that the older-shape competition still exists: Beelink's GTR9 Pro 128GB lists around $1,985 at the official store as of this writing, without OcuLink or the 140W tune. One commenter's lament sums up the whole market: "When the first Strix Halo mini PCs started arriving, $2k felt a little spendy so I figured I'd wait for prices to drop a bit over time..." They went the other way. ## What viewers are saying - "GMKtec announced that a version of this with a 495+ and 192GB of RAM is expected to be launched before the end of the year. Also be careful with your eGPU card choice." *@r.l.royalljr.3905 on YouTube*(the 192GB variant is also[reported by TweakTown](https://www.tweaktown.com/news/112071/gmktec-evo-x3-mini-pc-is-coming-with-oculink-support-and-a-ryzen-ai-max-plus-pro-495-variant-packing-192gb-of-memory/index.html?ref=vettedconsumer.com)) - The skeptic's case, for balance: "Literally the same machine as the x2 in a different form factor, the only difference is the occulink... and you can use a m.2 occulink adapter on the x2 so MEH." *@BirdManPhil on YouTube* - "Imagine running Windows in 2026" *@michaelgleason4791 on YouTube*(Wendell's Linux testing is the promised follow-up) ## The verdict The EVO-X3 doesn't change what Strix Halo is: 128GB of unified memory at ~256 GB/s that runs the [MoE-era models](https://vettedconsumer.com/every-frontier-open-model-is-a-moe-now-what-that-does-to-your-hardware-math/) this tier was made for. What it changes is the ceiling for tinkerers: the fastest Strix configuration Wendell has tested, plus a legitimate PCIe path for mixing a discrete GPU into the box, which is the difference between "capable mini PC" and "small lab." If that describes you, it earns its premium over the cheaper boxes. If you just want the tier at the lowest price, the older-shape machines still undercut it by a wide margin, and the 192GB variant on the horizon is a reason for the patient to stay patient. Whatever box you pick, check your shortlist of models against it in [the calculator](https://vettedconsumer.com/can-i-run-it/). ## Sources and how we researched this - Video: ["Checking Out The GMKTec X3 Strix Halo: More Strix Halo Shenanigans!"](https://www.youtube.com/watch?v=cRNb1JyNdDw&ref=vettedconsumer.com)by Level1Techs (Wendell), August 3, 2026. All hands-on findings and measured figures above are his; we have not tested this machine first-hand. Full benchmark tables in the[Level1Techs forum thread](https://forum.level1techs.com/t/gmktec-evo-x3-ai-workstation-lemonade-benchmarks-and-specs/253262?ref=vettedconsumer.com). - Launch pricing and positioning: [VideoCardz](https://videocardz.com/newz/gmktec-evo-x3-with-ryzen-ai-max-395-to-launch-at-3600-comes-with-128gb-memory?ref=vettedconsumer.com)($3,600 launch),[Notebookcheck](https://www.notebookcheck.net/GMKtec-launches-repackaged-Strix-Halo-mini-PC-with-2x-higher-price.1335259.0.html?ref=vettedconsumer.com), and[VideoCardz on the Ryzen AI Halo comparison](https://videocardz.com/newz/gmktec-pc-with-max395-and-128gb-costs-400-less-than-amd-ryzen-ai-halo?ref=vettedconsumer.com). Beelink GTR9 Pro pricing per the official Beelink store, checked August 8, 2026. Prices in this class are moving weekly. - The 192GB Ryzen AI Max+ PRO 495 variant: [TweakTown](https://www.tweaktown.com/news/112071/gmktec-evo-x3-mini-pc-is-coming-with-oculink-support-and-a-ryzen-ai-max-plus-pro-495-variant-packing-192gb-of-memory/index.html?ref=vettedconsumer.com). - Viewer comments quoted verbatim and attributed from the video's comment section. *Related: **GMKtec EVO-X2 guide (and its price history)** · **Strix Halo vs DGX Spark, per owners of both** · **Every frontier open model is a MoE now** · **Prompt processing vs generation*