(Credit: Nvidia)
Don't miss out on our latest stories. Add PCMag as a preferred source on Google.
John Carmack, id Software founder and former Oculus VR CTO, has released a blistering criticism of Nvidia's new DGX Spark mini AI-supercomputer, claiming that it delivers around half the rated performance, pulls less than half the claimed power requirements, and may even be overheating during long runs. In short, he's not impressed, and other developers are chiming in with their own underwhelming experiences.
Nvidia launched the DGX Spark in mid-October. Powered by the new GB10 superchip, it features a 20-core ARM-based Nvidia Grace CPU and a Blackwell GPU with thousands of CUDA cores. That's roughly equivalent to an RTX 5070, but with 128GB of shared LPDDR5X memory instead of VRAM, and 4TB of NVMe storage. It's a powerful little system for AI training and inference, with access to Nvidia's full suite of AI developmental tools, CUDA compatibility with various other programs and applications, and all in a tiny footprint.
Except, Carmack claims Nvidia's numbers don't add up.
Nvidia says the DGX Spark should draw around 240W at full tilt, but Carmack claims it's more like 100W. Nvidia says it can deliver up to a petaflop of AI performance, but Carmack claims it's closer to half that. It also gets very hot. VideoCardz quotes the lead developer of Apple's MLX framework as backing up Carmack's claims, with the DGX Spark only achieving 60 TFLOPS of performance, where it would expect close to four times that.
That's a big disappointment for hardware that costs $4,000, although admittedly, some partner designs are more affordable, with a starting price of $3,000.
Part of the reason for this performance disparity could be with Nvidia fudging the numbers during its big reveal of the new mini workstation. This is something it's been known to do during graphics card unveilings for gamers, claiming that the RTX 5070 would offer RTX 4090-like performance, for example, where it barely beat out the last-generation 4070 Super in reality.
With the DGX Spark, Nvidia claims a 1 petaflop performance rating for the system, but that's when basing the numbers on structured sparsity. As VideoCardz details, this is where the hardware skips zero-value operations in neural networks, which can double its effective compute rate. Great for performance numbers and benchmarks, not so useful in the real world where workloads aren't optimized for that sort of sparsity.
Although it's not confirmed that's what is happening in Carmack's case, it sounds about right. It doesn't mean the DGX Spark is useless; it's still a powerful AI dev tool, but it's probably not as good as Nvidia made it seem.
Nvidia did not immediately respond to a request for comment.