AI & VR Hardware Architecture: Portable 128GB Systems for Local LLMs and Wireless VR

Editor’s note: this started as a research summary from a conversation with Gemini exploring portable hardware for local AI + VR. The specs below have been checked against current vendor listings and independent reviews and corrected where they were wrong — corrections are called out inline, not silently fixed — with a recommendation added at the end.

The problem and the hard constraints

The goal: one travel-friendly setup that can host a local 70B-class model and drive high-fidelity wireless VR (Half-Life: Alyx, Skyrim VR) without compromising badly on either.

• The 128GB threshold. INT4 quantization of a 70B model needs roughly 35–42GB of raw memory (70B params × ~0.5 bytes/param at 4-bit, plus KV cache and context overhead). Once you add OS overhead and headroom for concurrent apps, 128GB unified/system RAM is the realistic floor to avoid OOM failures — the math checks out and is the right constraint to design around.
• The privacy mandate. Deep, reasonable distrust of default Windows 11 telemetry and features like Recall. Linux (kernel 6.12/7.0+) is the preferred path where drivers allow it; where they don’t, hard Windows-debloating is the fallback — see trade-offs below.
• The portability dilemma. Heavy copper/thermal solutions for sustained AI + VR loads fight against a light travel footprint. A split headless home server + ~1kg thin-client companion is a real third option, not just a compromise between the other two.

At a glance

1. Ultra-portable 13–14" laptops — ASUS ProArt PX13 (HN7306)

AMD Ryzen AI Max+ 395 (“Strix Halo”), up to 128GB unified LPDDR5X-8000 memory (256-bit bus), 13.3" 2.8K OLED touchscreen at 60Hz, Radeon 8060S integrated graphics. Confirmed against ASUS’s own spec sheet and Notebookcheck’s review — one correction: 128GB is the top configuration, not the default — most retail listings ship 64GB, so specifically confirm the 128GB SKU before buying.

• Storage: single M.2 2230 slot only — plan on cloning to a larger 2230 drive for model weights, since there’s no second bay.
• Linux: needs kernel 7.0+ for full Strix Halo support; the smart-amplifier audio firmware quirk (silent output until the .bin is copied from the Windows partition) is a real, documented issue on this generation of AMD APU laptops, not specific to this one model.
• Verdict: the best inference speed of any laptop here thanks to LPDDR5X bandwidth feeding the integrated GPU. VR is capped by the 60Hz panel and the 8060S’s shared thermal envelope — medium settings, not ultra.
• CUDA development: none. The Radeon 8060S is ROCm/Vulkan territory only — no CUDA path exists on this hardware. Fine for quantized-inference work (llama.cpp, GGUF), a real limitation if your work involves CUDA-only libraries, custom kernels, or training code written against NVIDIA’s stack.

2. Compact dedicated mini-PCs — Minisforum AtomMan G7 Pro

Correction, and an important one: independent reviews (TechPowerUp, Guru3d) and Minisforum’s own listing put this system’s memory ceiling at 96GB across two SO-DIMM slots (DDR5-5600), not 128GB. 64GB SO-DIMM modules exist, so 2×64GB may physically work, but it isn’t a Minisforum-validated configuration — don’t plan a 70B deployment around it without testing your specific board first. As configured, 96GB is below this brief’s own 128GB threshold.

• CPU/GPU: Intel Core i9-14900HX + NVIDIA RTX 5070 Laptop GPU (8GB VRAM), 200W performance mode, six-heatpipe dual-fan cooling.
• Storage: two M.2 2280 slots — a real advantage, simple second-drive install without touching the Windows install.
• Verdict: the clear VR winner in this list — dedicated NVIDIA silicon with full CUDA support handles Half-Life: Alyx at ultra settings over Wi-Fi 7 without argument. For 70B inference, once a model spills past VRAM into system RAM, the narrower SO-DIMM bus throttles hard (3–6 tok/s is a realistic range) — and the 96GB ceiling means a 70B model at INT4 is already tight before you even get to that bottleneck.
• CUDA development: yes — real CUDA/cuDNN, the only mainstream stack most PyTorch code, training tutorials, and CUDA-only libraries actually target. The catch is VRAM, not CUDA support itself: 8GB on the RTX 5070 Laptop GPU is enough for CUDA dev work, small-model fine-tuning, and inference experimentation, but not for training/fine-tuning anything close to 70B — that still has to happen elsewhere or in the cloud.

3. Fixed unified-memory boxes — MS-S1 Max / GMKtec EVO-X2 / ASRock AI BOX-A395

Same Ryzen AI Max+ 395 platform as the ProArt PX13, in a desktop-ish mini-PC shell: 128GB soldered LPDDR5X-8000 unified memory (confirmed, non-upgradable), 40-CU Radeon 8060S, 50 TOPS NPU.

Correction: the MS-S1 Max specifically ships with a built-in 320W internal power supply, not an external ~200W brick — check the exact model before assuming brick weight/size for a travel kit; this varies across the EVO-X2 / AI BOX-A395 family, which do use external adapters.

• GTT memory allocation (Linux/ROCm): by default the GPU-addressable GTT pool is roughly 50% of system RAM; reaching a 96GB GPU allocation needs a small BIOS framebuffer (512MB) plus a kernel-level bump via amdgpu.gttsize or the amdttm.pages_limit parameter, verified with mem_info_gtt_total. Leave 4–12GB for the CPU/kernel — pushing GTT too high has been reported to cause large-model loads to hang rather than OOM cleanly, so this needs actual testing on your kernel/ROCm version, not just applying the flag and trusting it.
• Key features: MS-S1 Max adds dual 10GbE (useful for a data-pipeline/ home-server role); ASRock’s AI BOX-A395 adds a physical dual-BIOS switch, useful for testing GTT configs without risking a bad flash.
• Verdict: the most efficient way to get a real 128GB AI micro-server, including 70B and MoE architectures up to ~120B. VR runs natively at medium settings — same integrated-GPU ceiling as the ProArt PX13.
• CUDA development: none, same as the ProArt PX13 — Strix Halo is a ROCm/Vulkan platform. PyTorch 2.7+ does support ROCm 6.3 as a first-class backend and llama.cpp/Ollama run well on it, but training and diffusion workloads still trail CUDA meaningfully (roughly 20–30% behind on comparable training benchmarks, per current ROCm-vs-CUDA writeups), and some CUDA-only tooling has no ROCm path at all. Fine as a 128GB inference server; expect real friction if the job is CUDA development specifically.

4. Enterprise AI boxes — NVIDIA DGX Spark / ASUS Ascent GX10

NVIDIA GB10 Grace Blackwell superchip (20-core Arm CPU + Blackwell GPU), 128GB unified LPDDR5X, up to 1 PFLOP FP4 compute. Confirmed weight is ~1.48kg, not 1.2kg.

• Verdict: unmatched local tensor throughput and a real path to NVIDIA’s enterprise/cloud tooling for fine-tuning up to ~200B parameters — genuinely a different class of device from everything else on this list. Completely off the table for VR: it runs DGX OS (Arm Linux), headless by design, with no SteamVR or Oculus Link driver path. Don’t shortlist this if VR matters at all — it’s a pure compute appliance, not a dual-purpose machine.
• CUDA development: the strongest option on this list, not just a “yes.” DGX OS ships the full NVIDIA stack (CUDA, cuDNN, TensorRT) out of the box, on the same Blackwell architecture NVIDIA’s own enterprise tooling targets — this is the one system here built specifically for CUDA-native AI development, not adapted to it. The Arm CPU is the one thing worth checking per-tool: most modern CUDA/PyTorch tooling has Arm builds now, but a small number of x86-only packages still don’t.

My take

None of these four is a free lunch — the trade-off is structural, not a gap someone forgot to engineer around. Unified memory (LPDDR5X shared across CPU/GPU) is what makes 128GB physically possible in a small, efficient box, but the same integrated GPU that benefits from that bandwidth tops out around “medium VR settings” — and, on AMD, means no CUDA at all, ROCm only. Dedicated NVIDIA VRAM (RTX 5070, or Blackwell in the DGX-class boxes) is what makes VR-ultra and real CUDA development possible, but it’s a separate, smaller, faster pool that a 70B model doesn’t fit in — so it falls back to a comparatively narrow bus the moment it spills over, or in the DGX Spark’s case, isn’t meant for VR at all. Adding CUDA suitability into the picture doesn’t change that trade-off, it just adds a third axis to sit on top of it: AMD unified- memory boxes (the ProArt PX13 and the MS-S1 Max family) are 128GB-elite for inference but shut out of CUDA-native development entirely; the two NVIDIA-based options are the only real CUDA path, at the cost of either limited VRAM (AtomMan) or no VR capability whatsoever (DGX Spark).

Given that, the realistic answer depends on which of the three axes — LLM capacity, VR, CUDA development — actually matters most for your work:

• If local 70B inference + VR is the actual pair you need (my pick): an MS-S1 Max or ASRock AI BOX-A395 as a home-based, always-on 128GB inference server, reached over Tailscale from a light thin client for everyday carry, with a laptop or the AtomMan handling VR locally. This is the only path that gets genuinely elite LLM performance without forcing a VR compromise onto the travel device — but it’s inference only; there’s no CUDA development happening on this hardware.
• If CUDA-native AI development is actually the priority (writing and training against real CUDA/PyTorch code, not just running quantized GGUF models), the AMD unified-memory boxes are the wrong category entirely regardless of their RAM size — the AtomMan G7 Pro (real CUDA, but 8GB VRAM and a 96GB system ceiling) or the DGX Spark / Ascent GX10 (the strongest CUDA environment here, zero VR capability) are the only two options that actually fit.

I’d avoid planning around the AtomMan G7 Pro for the 70B inference target specifically — 96GB officially supported is short of this brief’s own 128GB floor, and I wouldn’t bet a purchase on an unvalidated 2×64GB config working on that board — but it’s the best single box here if CUDA development and VR both matter more to you than hosting the largest possible local model. The DGX Spark / Ascent GX10 is worth knowing about but isn’t really competing in the VR half of this comparison at all — it’s a compute appliance for CUDA development specifically, not a dual-purpose machine.

Further reading

• Notebookcheck — ASUS ProArt PX13 GoPro Edition review
• TechPowerUp — Minisforum AtomMan G7 Pro review
• ServeTheHome — Minisforum MS-S1 Max review
• ServeTheHome — ASUS Ascent GX10 review
• AMD ROCm documentation — Strix Halo system optimization
• Thunder Compute — ROCm vs CUDA: GPU Computing Comparison