ASUS ProArt Mini PC With NVIDIA RTX Spark Packs 1 Petaflop of AI Into 150mm

The ASUS ProArt Mini PC is the spec sheet I've been waiting to dissect all week, and at Computex 2026 in Taipei, ASUS finally let me. Unveiled June 1, this is a desktop the size of a coaster — 150 x 150 x 51 mm — built around NVIDIA's brand-new RTX Spark superchip. The headline number is the one that matters: 1 petaflop of AI performance on your desk, sipping from a single power brick. Let me walk you through the silicon.

Inside the RTX Spark Superchip

The RTX Spark is a two-part beast fused into one package. On the CPU side you get a 20-core Grace Arm processor. On the GPU side, a Blackwell-generation RTX part with 6,144 CUDA cores and 5th-gen Tensor cores that support FP4 precision — the low-bit math that makes those petaflop figures achievable on local hardware.

The clever bit is the interconnect. Grace and Blackwell talk over NVLink-C2C, a chip-to-chip fabric that's an order of magnitude faster than shoving data across a conventional PCIe lane. That tight coupling is what lets the two halves share memory without a bottleneck, which brings me to the number every AI builder will care about.

128GB of Unified Memory

This Mini PC ships with 128GB of unified memory — one pool the Grace cores and the Blackwell GPU both address directly. No copying tensors back and forth across separate CPU and VRAM banks. For local AI work, unified memory is the difference between "it fits" and "it doesn't," and 128GB is a genuinely large playground for a box this size.

What 1 Petaflop Actually Buys You

Raw FLOPS are fun, but I care about workloads. ASUS rates the ProArt Mini PC to run 120-billion-parameter large language models with up to 1 million tokens of context locally. That's a serious frontier-class model and a context window long enough to feed an entire codebase or a stack of research papers in one pass — all without a cloud round-trip.

It's not just text, either. The thermal and memory headroom let it edit 12K video in 4:2:2 color and render 3D scenes north of 90GB. For a machine you could mistake for a Wi-Fi router, that's a remarkable creative envelope.

Thermals, Networking, and Expansion

Compact machines live or die on cooling, so here's where the engineering shows. ASUS built a dedicated thermal architecture rated for up to 140W of thermal headroom, which is what sustains long render passes and AI training runs without throttling. The connectivity matches the ambition: 10GbE networking to move those 90GB scene files fast, and a PCIe Gen 5 x4 M.2 slot for high-speed storage expansion.

To be clear on lineage — this is a distinct product from the ASUS Ascent QN10 I covered earlier, which rides on Qualcomm's Snapdragon X2 Elite. The ProArt Mini PC is a separate, NVIDIA-powered class of machine aimed squarely at local AI workstation duty.

Why This Matters for Local AI

The trend I keep tracking is compute migrating back to the desk. A 20-core Grace CPU, 6,144 CUDA cores, 128GB of unified memory, and a petaflop of FP4 throughput in a 150mm cube means a single developer can prototype, fine-tune, and serve large models without renting GPU time. The 1M-token context ceiling is the standout spec — long-context inference has historically demanded data-center memory, and here it is running on something you can hold in one hand.

Availability

ASUS says the ProArt Mini PC arrives in fall 2026 in select regions, with full configuration and pricing details closer to launch. I'll be first in line to benchmark it against the petaflop claim. Until then, the spec sheet alone makes this one of the most exciting small-form-factor reveals of Computex 2026.

Sources: ASUS Press Release (June 1, 2026); Notebookcheck (June 2026); PCWorld (June 2026)