PNY DGX Spark NVDGXSPARK-PB 2025, ARM, 128GB, 4TB SSD, Gold

128GB of unified memory and a 4TB SSD in a 1.2kg box, with up to 1 petaFLOP of FP4 AI performance. It's a 99th percentile RAM monster built specifically for local AI work. Just know the reliability score is a rough 11th percentile, and at $5,300 it's a specialist's machine, not a general desktop.

The DGX Spark is built around the GB10 Grace Blackwell Superchip, pairing a 20-core ARM CPU (10 Cortex-X925 and 10 Cortex-A725 cores at 3.2GHz) with a Blackwell GPU carrying 128GB of VRAM. That's not a typo. 128GB of VRAM in a mini PC. NVIDIA claims up to 1 petaFLOP of AI performance at FP4 precision, which is the kind of number that makes data scientists do a double take. The unified memory architecture means the CPU and GPU share that 128GB pool, so you're not bottlenecked by PCIe transfers when moving large model weights around.

For real-world AI work, the 128GB of memory is the star. It's enough to load a 70B parameter model with room to spare, and the 4TB NVMe SSD gives you space to keep multiple quantized models local. The ConnectX-7 Smart NIC with 10GbE means pulling datasets from a NAS or pushing checkpoints to a server won't feel like watching paint dry. Our benchmarks put the GPU at the 85th percentile, which is strong but not top-of-the-charts. The CPU at the 93rd percentile actually outranks the GPU relatively speaking. For a machine this size, the performance density is genuinely impressive. Just don't expect it to replace a multi-GPU server for training runs.

Against the Apple Mac Studio M3 Ultra, the DGX Spark holds its own on memory capacity and AI-specific hardware, but Apple's ecosystem is far more mature for creative work. The Mac Studio scores better for creators overall, while the Spark's 82.6 AI score shows where it's aimed. The Lenovo Legion 90YA003GUS and HP OMEN GT16-0364 are gaming desktops first, with discrete GPUs that crush traditional gaming benchmarks but can't touch 128GB of unified memory for model inference. The Dell Aurora ACT1250 and MSI Aegis ZS2 are similar stories, strong general-purpose machines that would feel out of their depth running a 200B parameter model. The Spark's 94th percentile port selection also beats most of these competitors, with Thunderbolt, DisplayPort, HDMI 2.1, and four USB-C ports. But every one of those competitors likely beats the Spark's 11th percentile reliability score, and that's worth weighing carefully.

At $5,300 USD, the DGX Spark is not cheap. But context matters. A traditional workstation with 128GB of VRAM would cost you multiple times that, assuming you could even find a single GPU with that much memory. The Mac Studio M3 Ultra with comparable unified memory runs significantly higher. For AI developers who need to run large models locally without renting cloud GPUs by the hour, the math can work out. The 4TB of storage and 128GB of RAM mean you're not immediately spending more on upgrades. But if you're not actively working with LLMs or training models, this is a lot of money for a machine that scores 61.7 for business use. The value proposition is razor-focused: it's a bargain for AI work and a splurge for everything else.

The headline numbers here are wild for a box this size. You're getting 128GB of unified LPDDR5x memory and a 4TB NVMe SSD in a 1.2kg chassis that sips power compared to a full workstation. Our database puts the RAM at the 99th percentile and storage at the 98th, which means this little machine out-specs almost every mini PC we've ever tracked. The 20-core ARM CPU lands at the 93rd percentile too, so it's not just a one-trick GPU pony. But the reliability score is rough, sitting at the 11th percentile, and that's a real conversation starter for a $5,300 machine.

PNY is shipping this with NVIDIA's DGX OS preinstalled, which is a developer-first Linux environment, not a casual Windows box. The pitch is simple: run AI models up to 200 billion parameters locally, fine-tune them, then push to the cloud when you're ready. For AI and LLM work it scores 82.6 out of 100 in our system, with developer use at 81.5. It's clearly built for one audience, and it nails that audience. Everyone else should probably keep scrolling.

If you're not actively working with AI models, skip this. The 61.7 business score and ARM architecture mean everyday productivity software may not even run natively. Gamers should look elsewhere too, the GPU's 85th percentile ranking is against mini PCs, not gaming rigs, and DGX OS isn't built for Steam. The 11th percentile reliability score is also a red flag if you need a machine that just works without tinkering. And if your AI workloads are small enough to run on a standard desktop with 32GB or 64GB of RAM, spending $5,300 on 128GB of unified memory is overkill. This is a machine for people who know exactly why they need it.

The DGX Spark is a specialized tool that happens to look like a mini PC. If you're an AI developer, researcher, or serious hobbyist who wants to run large models locally, the 128GB of unified memory and petaFLOP-class FP4 performance make this a compelling buy despite the $5,300 price. The 99th percentile RAM and 98th percentile storage rankings back that up. But the 11th percentile reliability score is a genuine concern, and the 61.7 business score tells you this isn't a general-purpose machine. Buy it for AI work and you'll probably love it. Buy it as a cool desktop and you'll wonder where your money went.