NVIDIA's mini AI supercomputer DGX Spark was finally officially launched not long ago, with a starting price of US$3,999 (approximately 28,396 yuan).However, John Carmack, co-founder of id Software, recently stated on social media that the actual power consumption and performance of DGX Spark are far lower than the official promotional value.
He and some researchers observed that NVIDIA DGX Spark appeared to fail to meet its stated performance goals,The DGX Spark is rated for 240W power consumption, but Carmack found that the device's power consumption only topped out at around 100W, less than half of the rating.
The system only achieved around 60 TFLOPS of performance under the BF16 workload and around 480 TFLOPS under the FP4 workload, well below the expected output.
Not only that, but it added that even at 100W levels, the device can get very hot, and at least one user has reported automatic reboots during long runs.
Replying to Carmack, Hannun, the lead developer of Apple's MLX framework, confirmed that he had obtained similar results in his own benchmarks.
Hannun found that when using PyTorch and MLX, the sustained performance of DGX Spark in BF16 was only about 60 TFLOPS, which was far lower than the 1 PFLOPS of FP4 performance advertised by NVDIA.
DGX Spark is positioned as a "mini workstation" system based on the GB10 super chip, combining MediaTek's ARM CPU and NVIDIA's own Blackwell GPU.