John Carmack says NVIDIA DGX Spark draws under half its rated power, delivers half expected performance
Image courtesy of OnThisDay & Ryans
In a post on X, John Carmack, NVIDIA’s new DGX Spark system is not reaching its claimed performance targets. He and other researchers observed that the device draws only about 100W, less than half of its 240W rating, and achieves around 480 TFLOPS in FP4 or roughly 60 TFLOPS in BF16 workloads, which is less than half of the expected output.
Carmack, best known as the co-founder of id Software and for his work on early 3D graphics, wrote on X that the DGX Spark “ appears to be maxing out at only 100 watts ” and “ delivering about half the quoted performance. ” He also noted that the unit becomes hot even at that level, with at least one user reporting automatic restarts during long runs.
DGX Spark appears to be maxing out at only 100 watts power draw, less than half of the rated 240 watts, and it only seems to be delivering about half the quoted performance (assuming 1 PF sparse FP4 = 125 TF dense BF16) . It gets quite hot even at this level, and I saw a report…
— John Carmack (@ID_AA_Carmack) October 27, 2025
Hannun, the lead developer of Apple’s MLX framework who responded to Carmack, confirmed similar results in his own microbenchmarks. Using PyTorch and MLX he found the DGX Spark sustained only about 60 TFLOPS in BF16 matrix operations, well below expectations for a system advertised as offering 1 PFLOPS in FP4.
NVIDIA & Sparsity
As others have pointed out, NVIDIA’s claimed 1 PFLOPS performance for the DGX Spark is based on structured sparsity , a hardware feature that skips zero-value operations in neural networks, effectively doubling the theoretical compute rate. This figure assumes workloads fully optimized for sparsity, which from I read, rarely applies in general use. In normal dense operations without sparsity, the actual performance is about half that number, matching what tests above have shown.
NVIDIA DGX Spark is a ‘mini workstation’ system that runs on DGX OS. It is based on GB10 Blackwell SuperChip, which is part of NVIDIA N1-series of SoCs combining Mediatek ARM CPU and NVIDIA’s own Blackwell GPU. The system launched last week , and it costs $3999, with partner designs starting at $2999.
Source: X/Twitter
