NVIDIA's compact AI supercomputer, the DGX Spark, has made its official debut on the market, boasting a starting price tag of $3,999 (equivalent to roughly 28,396 yuan). Nevertheless, John Carmack, the co-founder of id Software, took to social media to reveal that the actual power consumption and performance metrics of the DGX Spark fall considerably short of NVIDIA's official assertions. The supercomputer's rated power consumption stands at 240W, yet in reality, its peak consumption hovers around a mere 100W, marking a less-than-half achievement compared to the stated figure. When subjected to a BF16 workload, the system's performance registers at approximately 60 TFLOPS, while under an FP4 workload, it reaches about 480 TFLOPS—both figures failing to meet anticipated levels.
