John Carmack, the renowned founder of id Software and former CTO of Oculus VR, has unleashed a scathing critique of Nvidia's DGX Spark mini AI-supercomputer, sparking debate among tech enthusiasts and developers. Carmack's assessment reveals a stark contrast between Nvidia's promises and the actual performance, leaving many underwhelmed.
Nvidia's DGX Spark, unveiled in mid-October, boasts an impressive array of features, including a 20-core ARM-based Nvidia Grace CPU and a Blackwell GPU with thousands of CUDA cores, equivalent to an RTX 5070. However, Carmack's findings paint a different picture.
According to Carmack, the DGX Spark falls short in several aspects. It consumes approximately 100W of power, significantly less than Nvidia's stated 240W at full capacity. Moreover, its AI performance is estimated to be around half of the promised petaflop, and it tends to overheat during extended usage.
These findings are supported by the lead developer of Apple's MLX framework, who reported that the DGX Spark achieved only 60 TFLOPS of performance, a stark contrast to the expected 240 TFLOPS. This discrepancy raises concerns about Nvidia's accuracy in its marketing materials.
The performance disparity could be attributed to Nvidia's tendency to exaggerate performance during product launches, as evidenced by the RTX 5070's claims of RTX 4090-like capabilities. The DGX Spark's performance is further complicated by Nvidia's use of structured sparsity, a technique that enhances compute rate but may not be optimal for real-world workloads.
Despite these challenges, the DGX Spark remains a formidable AI development tool. However, it is essential to approach Nvidia's claims with a critical eye, recognizing that the actual performance may not match the advertised specifications. As the debate continues, developers and enthusiasts eagerly await further insights and responses from Nvidia.
