The DGX Spark is an AI-specialized “supercomputer” in mini format. It promises high performance on paper, with 1 PFLOP on AI workloads. However, some have noted that this performance is not being met in practice. This is what John Carmack announced on X, along with Awni Hannun, lead developer at Apple.
The DGX Spark doesn’t hold its perfs, heats up and in some cases reboots?
John Carmack, co-founder of ID Software, takes the floor on X and points out that NVIDIA’s supercomputer doesn’t live up to its promises in terms of performance. He measured a maximum power consumption of 100W, whereas the system is rated at 240W. Subsequently, it states that the performance level observed is well below what was advertised by the chameleon. In fact, the device offers only 480 TFLOP in FP4, or around 60 TFLOP in BF16, where 1 PFLOP was advertised in FP4. But that’s not all, as Carmack also points out that the machine tends to overheat in these conditions. He also says he’s had feedback that the machine has even restarted in one case because of the heat.
But he’s not the only one, as Hannun reports similar results using PyTorch and MLX. Once again, the machine only managed to output 60 TFLOPs during BF16 matrix operations.
In this way, we might be tempted to say that NVIDIA has overestimated the performance of its hardware, but others have pointed out that the 1 PFLOP claimed can only be achieved under certain very specific conditions, since the DGX Spark hardware supports structured sparsity. In practice, however, this requires a particularly well-optimized workload, which is clearly not often the case. Without this sparsity, performance plummets, and that’s what we’ve seen here.
