엔비디아 DGX 스파크 일상 사용기

I bought an NVIDIA DGX Spark! It’s actually quite a competent general purpose computer while being super tiny and power efficient. It basically just has an ARM CPU, a decent NVIDIA GPU (with the same number of CUDA cores as the RTX 5070), 128 GB of unified memory, and a 4 TB SSD. DISCLAIMER: I work at NVIDIA but I bought this with my own money and the opinions are my own. I didn’t get anything for writing this review. When I bought my DGX Spark, it only cost me $4000, but its price has since increased to $4700. The DGX Spark runs “DGX OS” but it is in fact just plain old Ubuntu 24.04 with some additions. If you want, you can just install another Linux distribution easily (Fedora works well), although there may be a couple of weird bugs with the Realtek Ethernet driver so the NVIDIA version of the Linux kernel has a couple of patches. Unlike some other ARM devices, the DGX Spark is all ACPI rather than device tree based, so regular Linux builds for arm64 work just fine. Generally, it is a power-efficient and quiet box (it gets slightly audible under load, but no annoying whine, and is silent when idle). It’s really nice how it doesn’t heat up the room or make noise. Stuff like web browsing and general programming work are super fast. The CPU has a Geekbench 6 score roughly on par with an M3 Pro 14 core and an AMD Ryzen 9 7950X: 2983 single core and 18183 multi core . Weirdly enough, my results were slightly worse than Geerlingguy’s Dell Pro Max with the same GB10. Perhaps it’s just thermals or random fluctuations. At any rate, it’s better than my SFFPC with its AMD Ryzen 9 7945HX, which scored 2869 and 16973 . 1 Can it run Crysis? It turns out, it can totally run Crysis. The DGX Spark is quite decent at gaming in fact. This is Crysis Remastered running at 3200 x 1800 resolution with everything set to High and DLSS set to Performance. The game has a platinum rating on ProtonDB , but like other people running NVIDIA graphics, I had to set the Proton version to Experimental for it to work. The Steam Snap for arm64 is maintained by Canonical and specifically targets gaming on DGX Spark. It’s probably the best avenue to play most games. It uses FEX to translate x86_64 to arm64, fulfilling a similar role as Rosetta 2 on Apple devices. It also bundles DLSS, RTX, and stuff so you can expect to play modern AAA titles with ray tracing turned on. Obviously, the “bang for the buck” isn’t great compared to dedicated gaming machines. After all, the DGX Spark is quite expensive. You probably wouldn’t want to buy the DGX Spark just for gaming. My small form factor PC with a dedicated NVIDIA graphics card is better at gaming at a lower price. The GPU of the DGX Spark is somewhat comparable to the RTX 5070 but with much slower memory bandwidth and much lower power, and there’s some overhead from FEX (whereas the performance penalty of Proton is usually rather tiny, or sometimes games are even faster in Proton than on Windows), so you can expect maybe two-thirds of the performance. Many modern Steam games with great graphics work just fine out of the box, including: Cyberpunk 2077 Portal RTX Indiana Jones and the Great Circle Running Train and many more! Generally, modern games on common game engines like Unreal, Unity, and Godot have a very high chance of just working, with bells and whistles like RTX and DLSS working too. The main categories of games that don’t work are: Games with aggressive anticheat that don’t work on Linux, such as League of Legends Some older games that need tinkering to get working (see the software that needed tinkering section) Non-Steam games have a lower chance of working 2 Local LLM inference The main reason for most people to buy the DGX Spark is probably to dabble with local LLM inference. I am happy to report that the DGX Spark has aged like fine wine in this regard. It just keeps getting better! Whereas early reviews lambasted it for apparently mediocre performance (bottlenecked by the memory bandwidth), a litany of novel tricks have enabled it to be literally several times faster now than when it first came out! NVFP4 gave a huge boost. It wasn’t originally supported on DGX Spark when it first came out, to much chagrin, but it became available soon afterwards. NVFP4 KV-cache quantization brought a big time-to-first-token speedup. Multi-token prediction gives a free speedup of up to 2x on many models. NVIDIA Model Optimizer gave another big boost to quantized models, resulting in a release on 2026-06-30 of a version of Qwen 3.6 27B that was way faster and better performing than earlier quantizations. DFlash uses a diffusion-based drafter to predict an entire block of future tokens at once, further improving upon speculative decoding. There’s a Spark Arena leaderboard with recipes and benchmarks that’s pretty useful. Anyway, right now, you can expect around 30-40 tokens per second (single concurrency, single Spark) with an NVFP4 quantization of Qwen 3.6 27B dense model, which is highly capable for everyday programming. It’s quite useful for general coding and agentic automation. At concurrency level 4 (four agents running in parallel, for example) you can expect a total token output rate of around 100 per second. For the Qwen 3.6 35B mixture-of-experts model, you can expect much faster performance, up to around 170 tok/s decode, so that’s incredibly useful for quick chatting and automation, although perhaps slightly worse at coding. 3 Software that needed tinkering 3.1 Darktable Darktable’s arm64 build works out of the box, but is extremely sluggish. It seems there’s some issue with the NVIDIA driver (at least on version 580, as tested) where transferring data from the GPU to CPU is very slow in Darktable’s OpenCL pipeline. This affects both in-app interactivity as well as export speed. I created the pull request “prefault OpenCL readback destinations” which fixes the issue. The pull request “preheat host memory for OpenCL” also fixes it, in a perhaps slightly nicer way (this PR got reverted, so the fix still isn’t available in mainline Darktable). 3.2 Rhinoceros 3D Rhinoceros 3D is a 3D modeling application for Windows x86_64 and for macOS (on Apple Silicon). I bought a license, and used this for my other projects such as my line scan camera case . Unfortunately, the latest version is well-known for not working on Linux, with a salty thread on the Rhino forums about it. Getting the Windows version to work is generally considered more straightforward than the macOS version. Anyway, I asked Codex to get it working, and it actually did! Here are the AI-generated instructions: Rhino setup instructions Please note: Wine does not accept AI-generated code, and I’m not knowledgeable enough about the fixes here, so I can’t really upstream my patched fork. 3.3 Need for Speed Porsche Unleashed Need for Speed Porsche Unleashed is one of my favourite video games of all time. I used to play it a ton in elementary school and have great nostalgia when I play it again. There is something special about just driving through a scenic European street without any distractions like “influencer voices”, experience confetti, billboards and ramps, and so on. At the same time, the pretty atmosphere is chill compared to super hardcore racing sims on laser-scanned tracks surrounded by nothing but tires, chain link fences, and ads. It turns out it just works well with mainline Wine for arm64 with FEX as the ARM64EC compatibility backend. I installed Verok’s patch and used the OpenGL3 renderer which works great! Now the graphics are even better than the original game (which ran only at really low resolutions of like 1024x768). One weird quirk: GB10 has 10 fast Cortex-X925 cores and 10 slow Cortex-A725 cores. I had to explicitly pin it to the 10 fast cores rather than the 10 slow cores, otherwise the framerates would be really low. I guess perhaps the game from 1999 uses so few resources that the scheduler doesn’t realize it needs to be running in real time. On Wind