(Credit: Nvidia)
Don't miss out on our latest stories. Add PCMag as a preferred source on Google.
Nvidia's new DGX Spark mini AI workstation has had its "Can it run Crysis?" moment and passed. Kind of.
Although not designed to run games, the miniature system does have a GB10 superchip equipped with a powerful 20-core ARM CPU and a Blackwell GPU with 6,144 CUDA cores, roughly equivalent to an RTX 5070. With 128GB of LPDDR5X memory, it has the raw power to run games. It turns out that with the right software hacks, it actually can.
A Reddit user who invested $4,000 in the Nvidia reference DGX Spark design (complete with a spackled front panel) managed to get Cyberpunk 2077 running on it using Box64, an x86-to-ARM translation layer. As VideoCardz reports, they compiled Box64 0.3.8 with Box32 support and installed Steam on the device. After a little more finagling, they managed to get the game running at an average of around 50fps at 1080p with medium settings. Ironically, DLSS support wasn't viable, so multi-frame generation is out too.
That's not a lot for the price tag, but then again, the DGX Spark was designed to accelerate AI workloads and serve as an AI developer machine, not play games.
Our Reddit friend wasn't the only person who thought it could be used to play games. ETA Prime on YouTube performed an unboxing and review of the MSI EdgeXpert AI, a third-party device that sports Nvidia's new DGX Spark design, including the GB10 superchip. Priced at a comparatively low $3,000, it still features the same CPU and GPU cores, as well as 128GB of internal memory, so the hardware is more than capable.
This time around, however, they used RPCS3 and Xemu to emulate PlayStation 3 and Xbox games, such as Skate 3 and the original Forza Horizon, which ran smoothly at 30fps. That's no challenge for a hardware setup like this, though, and ETA Prime plans to test more demanding games via emulation, including several PC titles, in the future.
As Doom modders have shown, gamers will get their games to run on just about anything if you give them enough time. The question now becomes whether Nvidia wants to support this. Official drivers would make the whole experience far better, but it would be catering to an extreme niche within an already-niche product segment. That seems unlikely, considering Nvidia has far bigger tasks in mind as it looks to dominate global data center hardware supply chains.