NVIDIA's DGX Spark machine, designed as the ultimate AI box for local and fast AI prototyping, is reportedly operating at half the expected power and performance levels. John Carmack, founder of AGI-focused Keen Technologies and former CTO of Oculus VR, claims that the DGX Spark mini PC is not meeting its specified performance. NVIDIA rates the DGX Spark mini PC at 240 W of system power, but Carmack's benchmarks indicate it only draws about 100 W, effectively halving the power draw and performance. The DGX Spark's peak throughput is approximately 31 TeraFLOPS for FP32 and around 1,000 TOPS with NVIDIA's NVFP4 reduced-precision format. At BF16 dense compute, it is supposed to achieve 125 TeraFLOPS, but these targets are not being met. The measured compute is about 480 TeraFLOPS at FP4 and only about 60 TeraFLOPS at BF16.

After facing multiple delays, NVIDIA's DGX Spark has finally reached developers. However, many are reporting software and firmware issues on NVIDIA's end. There may also be thermal throttling problems, causing the chip to reduce frequency and voltage to prevent overheating. In some cases, the system has rebooted, potentially due to inadequate cooling. The GB10 SoC is rated for a 140 W TDP, and the 128 GB configuration of LPDDR5X could add several dozen additional watts. Therefore, a 100 W power draw doesn't seem feasible for the DGX Spark. It remains to be seen whether a software or firmware update will address these issues, or if NVIDIA will provide an aftermarket cooling solution for its $3,999 machine if it continues to overheat.