Field Note
Local artificial intelligence hardware exists in a perpetual tug-of-war between raw compute speed and addressable memory. High-end consumer GPUs compute quickly but hit strict VRAM ceilings, while unified-memory APUs offer massive capacity at modest bandwidth. A recent engineering experiment by tech reviewer Alex Ziskind tested the boundary between these two approaches by pairing GMKtec's EVO-X3 mini-PC—powered by AMD's Ryzen AI Max+ 395 processor with 128 GB of unified memory—with external NVIDIA graphics cards over an OCuLink interface.
On paper, the combined pool represents up to 224 GB of graphics-addressable memory [00:00]. Making an AMD integrated GPU and a dedicated NVIDIA workstation card run a single model simultaneously, however, requires bypassing the proprietary software stacks that define both platforms [00:26, 09:47]. AMD relies on ROCm, while NVIDIA depends on CUDA [09:52, 09:58]. Because neither company designs their runtimes to share workloads with the other, the setup required llama.cpp compiled against the cross-platform Vulkan backend [00:26, 10:09, 10:24].
The baseline performance highlights the fundamental tradeoff of unified memory [02:50, 04:05]. On its own, the EVO-X3's integrated Radeon 8060S processed a dense 32-billion-parameter model at roughly 11.2 tokens per second [02:58, 04:12]. Connecting an external NVIDIA GeForce RTX 5080 boosted speed on smaller models to roughly 85 tokens per second—nearly triple the APU's pace [03:23, 03:28]. But the moment a workload spilled past the RTX 5080's 16 GB buffer, performance collapsed to 1.6 tokens per second as memory management stalled [04:05, 04:08].
To test true large-model execution without CPU fallback, Ziskind upgraded the external dock to an enterprise-grade NVIDIA RTX Pro 6000 Workstation Edition card with 96 GB of onboard VRAM [00:26, 06:18, 06:25]. On a standalone basis, the Pro 6000 handled large mixture-of-experts (MoE) models at blistering speeds [06:46, 07:02]. Running an active-parameter MoE architecture like Qwen 3.5 122B-A10B (quantized to 4-bit, roughly 76.5 GB on disk), the card delivered over 120 tokens per second because MoE designs only activate a fraction of their total parameters during any single token generation step [06:47, 07:02, 07:13].
The critical test came with an uncompressed, 8-bit quantized version of that 122-billion-parameter model, demanding roughly 130 GB of storage—exceeding the standalone capacity of both the 96 GB RTX card and the 96 GB maximum VRAM allocation of the AMD APU [08:34, 08:58, 09:01, 11:22]. By slicing the model across both devices via Vulkan—allocating 88 GB of layer weights to the NVIDIA card and 34 GB to the AMD APU—the frankenstein setup executed the full 130 GB model at an average of 37 to 39 tokens per second [09:03, 09:31, 09:36, 09:41].
While technically functional, the operational overhead reveals why heterogeneous computing remains rare in production [00:23, 09:47]. The OCuLink connection provides a PCIe 4.0 x4 link (roughly 63 Gbps), which introduces a physical communications bottleneck when passing layer activations between disparate memory buses [01:25, 11:00]. Furthermore, early platform firmware struggled to initialize the workstation card, requiring manual boot sequencing in Linux and custom startup scripts to hot-attach the eGPU dock safely [07:50, 08:07, 08:12].
The economics also demand clear-eyed scrutiny. The GMKtec EVO-X3 retails around $3,800 for the 128 GB configuration, while a workstation-class 96 GB RTX Pro card costs between $10,000 and $14,000 [06:19, 12:39]. Spending upwards of $15,000 on a mini-PC tethered to an external graphics dock makes little sense when compared to purpose-built multi-GPU server nodes or dedicated Apple Silicon studios [06:19, 12:20, 12:39]. For self-hosters who already own high-capacity APUs, however, the experiment proves that open runtimes like Vulkan provide a viable path to pooling hardware across manufacturer boundaries without relying on vendor lock-in [10:09, 10:14].