Why ASUS won't publish the one spec that decides if the ProArt GR1X is worth buying

A mini PC the size of a paperback book, packed with 128GB of memory, capable of running a 120-billion-parameter language model on your desk instead of in the cloud. That's the pitch behind the ASUS ProArt GR1X, and on paper it reads like the end of AI subscriptions as we know them. No more API bills. No more rate limits. No more waiting for someone else's GPU cluster to free up.

But there's a number missing from ASUS's spec sheet, and it's the number that decides whether the GR1X is a genuine workstation replacement or a very expensive way to run a chatbot slowly. Before you get excited about ditching your monthly AI subscriptions for a box on your desk, here's what you actually need to know.

What is the ASUS ProArt GR1X?

The ASUS ProArt GR1X is a compact desktop workstation built around NVIDIA's new RTX Spark platform — the chip previously known by its engineering codename N1X. ASUS describes it as combining an NVIDIA Blackwell RTX GPU and high-performance NVIDIA Grace CPU for running personal AI agents, advanced creating, and gaming, equipped with up to 128GB of unified memory.

The whole thing fits into a chassis that delivers remarkable power in a 150 x 150 x 51 mm compact design that fits anywhere, effortlessly tackling heavy AI tasks and complex 3D rendering from a device that fits in the palm of your hand. That's roughly the footprint of a Mac mini, which is exactly the comparison ASUS wants you drawing.

Under the hood, the RTX Spark superchip is a groundbreaking superchip combining a NVIDIA Blackwell RTX GPU (up to 6,144 cores) and a 20-core Grace CPU. ASUS rates the system at 1 Petaflop of FP4 performance, letting you run 120 billion-parameter models locally, and the broader RTX Spark platform is designed to render 90GB+ 3D scenes, generate 4K AI videos and run 120B-parameter LLMs with up to 1 million tokens of context.

On the connectivity side, the desktop version brings up to 128GB unified memory, 140W thermal headroom, 10GbE networking and M.2 PCIe Gen5 x4 expansion.

What's the missing number, and why does it matter so much?

Here's the catch. ASUS published core count, memory capacity, thermal headroom, even the exact millimeter dimensions of the chassis. What it didn't publish anywhere on the spec sheet is memory bandwidth — the speed at which that 128GB pool actually moves data to the GPU.

This isn't a minor omission. Capacity and bandwidth are two completely different things, and conflating them is how people end up disappointed with hardware that technically has "enough memory." Capacity tells you whether a model fits. Bandwidth tells you how fast it runs once it's loaded.

Look at NVIDIA's own DGX Spark — the developer-focused predecessor built on the earlier GB10 chip, which shares the same unified memory architecture as the new RTX Spark platform. NVIDIA's official documentation confirms the system offers Unified Memory: 128 GB of high-bandwidth memory for large models. That sounds great until you look at the actual throughput number, which independent testing has pinned down clearly: The DGX Spark's 273 GB/s memory bandwidth is modest compared with a data-center GPU.

That's the trade-off in plain terms: Spark has the memory capacity to fit large models that other desktop hardware cannot, but it is not built to deliver the highest possible inference throughput. Which means a system like this is genuinely good at one thing and mediocre at another, and marketing copy rarely draws that line for you.

How does the ProArt GR1X compare to AMD's Strix Halo boxes?

This is where the competitive picture gets interesting, and where ASUS's silence on bandwidth starts to look less like an oversight and more like a strategic choice.

AMD's rival platform, the Ryzen AI Max+ 395 (codenamed Strix Halo), already ships in several mini PCs at the same 128GB memory tier. AMD's own technical materials describe it as AMD Ryzen™ AI MAX+395 with 128GB, a groundbreaking SoC that sets new standards for an APU in memory capacity and generative AI capabilities. Independent reviews put its real-world memory bandwidth at up to 128 GB of LPDDR5X-8000 through a 256-bit bus delivering up to 256 GB/s of theoretical memory bandwidth, around 215 GB/s measured in practice.

That's already meaningfully faster than the 273 GB/s figure the older DGX Spark posts — and AMD isn't stopping at 128GB. The next-generation "Gorgon Halo" chips are already confirmed to go further: AMD's Ryzen AI Max Pro 400 series arrives with Strix Halo, you could pack up to 128GB of unified memory, but AMD is pushing that boundary higher with Gorgon Halo, and industry reporting confirms the next Ryzen AI Max PRO 400 series is already planned with up to 192GB of unified memory.

So on the one spec that actually determines whether a local LLM box is useful for your workflow — how much you can load into memory at once — AMD's roadmap is already half a generation ahead. NVIDIA's counter-argument isn't capacity. It's CUDA, TensorRT, and an entire software ecosystem that AMD's ROCm stack is still catching up to for many production AI workflows.

Why is Windows on ARM the real risk here?

The silicon is only half the story. The RTX Spark platform, including the GR1X, runs on Windows 11 on ARM because the Grace CPU inside it is Arm-based, not x86. That's a bigger deal than it sounds.

Windows on ARM has improved dramatically, but it's still not universal. Plenty of legacy Windows software, drivers, and creative plugins simply don't have native ARM builds yet, and emulation layers introduce performance penalties that can eat into the very compute advantage you bought the machine for. Adobe rebuilding core apps like Photoshop and Premiere specifically for this new wave of ARM-based creative machines is a signal worth paying attention to — it tells you the ecosystem is still catching up, not that it's already there.

If your workflow depends on a specific x86-only plugin, an older CAD tool, or niche creative software without ARM support, that's a much bigger practical risk than a few hundred GB/s of missing bandwidth.

Who should actually consider a box like this?

Based on how NVIDIA positions the underlying GB10/Spark architecture, this class of hardware makes the most sense for a specific kind of user. NVIDIA's own DGX Spark product page frames it as ideal for developers who want to build and run autonomous AI agents securely and locally with up to 200 billion parameters, with 128GB of unified system memory. The company also positions it for iterative work: improve the performance of pre-trained models by fine-tuning on NVIDIA DGX Spark, with 128 GB of unified system memory, fine-tune models up to 70 billion parameters.

That's a fairly specific profile: AI developers running local agents, creators doing heavy 3D rendering, and researchers who need to prototype with large open models before deploying them elsewhere. It's explicitly not built for the highest-throughput production inference, competitive gaming performance, or replacing a rack of data-center GPUs — that's the honest trade-off NVIDIA itself points to when talking about where DGX Spark fits versus where it doesn't.

If you're mostly running smaller models, doing occasional AI-assisted editing, or just want a fast desktop for everyday creative work, the memory-bandwidth ceiling won't bother you much because you're rarely stressing it. If you're trying to run a 70B+ parameter model as a daily driver and expect snappy responses, the missing bandwidth number is exactly the thing you need answered before you spend the money.

How do you find the real numbers before buying?

Don't take a spec sheet at face value when a number is conspicuously absent — that's usually not an accident.

1. Check NVIDIA's DGX Spark documentation for the closest official baseline on the underlying architecture, since RTX Spark shares its unified memory design.
2. Compare against NVIDIA's DGX Spark product page for the official framing of what workloads it's actually built for.
3. Look at AMD's Ryzen AI Max+ 395 technical brief as a direct point of comparison for bandwidth and real-world generative AI throughput.
4. Wait for independent benchmarks. Spec sheets tell you what a company wants to highlight. Third-party reviewers running actual tokens-per-second tests on real models tell you what you're buying.

What's the bottom line?

ASUS didn't kill AI subscriptions forever — not yet, anyway. What it did was ship a genuinely interesting piece of hardware while leaving out the one number that would let you compare it honestly against the competition. That's not unusual in tech marketing, but it's exactly the kind of gap a smart buyer should notice.

The unified memory capacity is real. The CUDA advantage is real. The Windows-on-ARM risk is real too, and so is AMD's memory-capacity lead on the horizon. Before you preorder anything with "unified memory" printed proudly on the box, ask the boring question nobody wants to answer in the press release: how fast does that memory actually move? That number, not the one on the glossy spec sheet, is what will decide whether your local AI box feels like a supercomputer or a very well-cooled paperweight.