On October 29th， according to KuaiKeJi， NVIDIA has finally launched its personal desktop AI supercomputer， the “DGX Spark.” While numerous OEM products are also emerging， this mini development machine priced at $4000 has drawn criticism from the esteemed John Carmack， who has questioned its actual performance and power consumption.
AMD， seizing the opportunity， has proposed offering a Strix Halo mini workstation as an alternative for Carmack’s evaluation.
Carmack stated in a post that the DGX Spark claims a power draw of 240W， but in practice， it can only reach up to 100W， and even then， it remains very hot. This suggests a significant discrepancy between advertised thermal design power (TDP) and actual operational performance， potentially indicating thermal throttling or inefficient power delivery even at reduced loads.
Regarding performance， the device is advertised to achieve 1 PFlops in FP4 precision with sparsity. However， its FP16 dense compute performance is claimed to be close to 125 TFlops， an eightfold difference. More critically， actual measured data reportedly falls significantly below even this FP16 figure. This disparity highlights a common industry practice of marketing peak theoretical performance， often achieved under highly specific， optimized conditions (like low-precision FP4 with sparsity)， which may not translate to real-world， general-purpose AI workloads. The substantial gap between FP4 and FP16 dense performance implies that the underlying architecture might be heavily optimized for sparse computations， which are not universally applicable.
Since the introduction of NVIDIA’s Blackwell architecture， the advertised AI compute performance has largely relied on FP4 data format with sparsity support， often utilizing 2:4 structured sparsity techniques. While this can theoretically double throughput， it is only suitable for specific matrix operations and workloads. In FP8 and FP16 precisions， the performance significantly degrades. This observation by Carmack aligns with a broader trend where newer architectures boast impressive figures based on nascent or less common data formats， potentially misleading users about their capabilities in standard development environments.
Furthermore， Carmack noted that the DGX Spark exhibited issues with sudden restarts after prolonged operation， leading him to question whether the performance was deliberately throttled before release. Such stability issues， especially in a high-performance computing device， are concerning and can severely impact usability and workflow. The possibility of pre-release performance limitations for stability or thermal management reasons is a serious accusation that warrants investigation.
Framework， a company that has been actively developing open-source hardware alternatives， responded to Carmack’s post by offering to provide him with a device equipped with AMD’s Strix Halo processor for testing. This positions the Strix Halo as a direct competitor to the NVIDIA DGX Spark.
AMD’s Vice President of AI Software， Anush Elangovan， openly welcomed the situation， stating they were ready to provide any support to Carmack to help him understand the capabilities of the Strix Halo. This move by AMD signifies a competitive strategy， leveraging Carmack’s expertise and public platform to highlight their own offerings against NVIDIA’s perceived shortcomings.
Furthermore， independent testing by ServeTheHome on a retail version of the DGX Spark reportedly showed that the combined CPU-GPU power consumption remained below 200W， and it was unable to reach the advertised 240W under any load. This corroborates Carmack’s observations about power limitations and suggests that the device may not be capable of sustaining its advertised performance profiles， even under less demanding scenarios than peak load.
In addition， multiple posts on NVIDIA’s developer forums have pointed out issues with the DGX Spark， including GPU crashes and unexpected shutdowns under sustained load. These user reports， combined with Carmack’s criticisms and the independent firm’s findings， paint a concerning picture of the DGX Spark’s stability and real-world performance. Such widespread reports of instability can significantly erode consumer confidence and raise serious questions about product readiness and quality assurance.
NVIDIA has not yet responded to these criticisms and reports.
