John Carmack: "DGX Spark has only half the advertised performance" (twitter.com)

John Carmack — a respected engineer and public commentator on systems and AI hardware — posted that NVIDIA’s DGX Spark delivers only about half of its advertised performance in his hands. His claim, shared on X, says real-world throughput was roughly 50% of published numbers, sparking debate because DGX-class systems are used as reference platforms for large-model training and expensive infrastructure buys. Carmack’s critique emphasizes a growing mismatch between peak theoretical specs vendors publish and the sustained, end-to-end performance teams see on practical workloads. For the AI/ML community this is important: it highlights that advertised TFLOPS or peak bandwidth often don’t translate into lower training time or higher token throughput once precision modes, memory bandwidth, NVLink/PCIe topology, thermal behavior, software stack (CUDA/cuDNN/Triton) and model-parallel inefficiencies are considered. The practical implication is procurement and benchmark transparency — teams should validate vendor claims with reproducible, workload-specific benchmarks (tokens/sec, steps/sec, cost-per-token) rather than relying on peak metrics. The post is a reminder that vendors need clearer disclosure of test conditions and that software/stack optimization can be as decisive as raw silicon in achieving advertised performance.