AZisk Bolted an RTX Pro 6000 Onto a Strix Halo and Ran a 122B Model Across NVIDIA and AMD at Once

Our GMKtec EVO-X3 coverage flagged the OcuLink port as the interesting new trick on Strix Halo mini-PCs. Alex Ziskind (@AZisk) just took that trick to its logical extreme: he bolted a $10,000 NVIDIA RTX Pro 6000 onto a 128GB Strix Halo box and ran a single 122-billion-parameter model across an NVIDIA GPU and an AMD APU at the same time. His words: "AMD and Nvidia forced to do the one thing they aren't meant to."

What OcuLink actually opens up

OcuLink is four lanes of PCIe Gen 4 on an external port, about 63 Gbit/s, the same electrical connection a desktop graphics card uses, exposed on the back of a mini-PC. It lets you plug a real desktop GPU into a Strix Halo machine for the first time. Ziskind tested the range, and the results tell a clean story about when an external GPU helps.

The headline lesson is the RTX 5080 row: a fast 16GB card is only fast on models that fit in its 16GB. The moment a model spills past that, the built-in Strix Halo APU, with its full 128GB pool to draw on, wins outright. Capacity beats raw speed once the model is too big for the card, exactly the tradeoff our bandwidth explainer and 128GB matchup keep circling back to.

Why the 122B model ran faster than the 32B

The result that looks impossible: the Pro 6000 ran a 122-billion-parameter model at 121 tok/s, faster than it ran a 32-billion one. Ziskind explains it exactly right. The 122B model is an A10B Mixture-of-Experts, only 10 billion of its 122 billion parameters fire per token. Generation reads only the active slice, so a huge-but-sparse model decodes like a small one. As he puts it, "the model is huge, but it's also light on its feet." That is the whole thesis of our MoE-era piece, demonstrated live on a desk.

The party trick: one model, two vendors' GPUs

The finale split a 130GB model across the NVIDIA card (88GB, ~70% of the layers) and the AMD APU (34GB), decoding at roughly 33 tok/s. The trick to making rival silicon cooperate is Vulkan: NVIDIA's CUDA and AMD's ROCm do not mix, but Vulkan runs on both, so LM Studio can spread one model's layers across both chips. Every token passes through both GPUs in sequence, which is why the OcuLink link speed is the ceiling here, a faster interconnect would go faster still. One caveat from the build: the machine would not even boot with the Pro 6000 attached (the firmware choked on a card that large), so he had to hot-plug it after boot, a genuine no-no with OcuLink that happened to work.

What viewers are saying

• "So over 20k+ for this vs 10ish for m5 ultra at 256gb... we are waiting for that video, want to see prefill and tps comparison." @en4ble1337 on YouTube
• "The X2 has USB-4, which allows thunderbolt eGPU setups. I'm actually going to be testing that today when my AMD AI Pro 9700 cards come in." @Ang3l12_3D on YouTube
• "You could not have picked a better day to post this since Qwen 3.8 Next Flash just dropped. This exact setup would work for it." @peterallen1999 on YouTube

The takeaway

For a local-AI builder, the useful conclusion is not "buy a $10,000 GPU for your mini-PC." It is the shape of the tradeoff OcuLink exposes. A discrete GPU added to a Strix Halo box is a huge win for models that fit its VRAM (up to 6x here) and for prompt processing, but the built-in 128GB pool is what lets you run the big sparse models at all, and MoE architecture is what makes those big models fast enough to bother. The most sensible version of this build is a mid-VRAM card for prefill plus the APU's memory for capacity, not a five-figure Pro card. Size any model against your own hardware in our Can I run it? calculator before you start buying ports.

Sources and how we researched this

• Video: "I Plugged RTX Pro 6000 Into a Strix Halo... Didn't Expect This" by Alex Ziskind, August 26, 2026. All measured figures are his; we have not tested this setup first-hand.
• Hardware context: our GMKtec EVO-X3 / OcuLink coverage and the MoE active-parameter explainer.
• Viewer comments quoted verbatim and attributed from the video.

Related: The GMKtec EVO-X3 and its OcuLink trick · Every frontier open model is a MoE now · Bandwidth, Not TFLOPS

See what your machine can run →