Tech

John Carmack Slams Nvidia DGX Spark: AMD Steps In

TIG SEA14 hours ago

59 2 minutes read

John Carmack Slams Nvidia DGX Spark: AMD Steps In to Assist Developers

The artificial intelligence hardware market is currently experiencing a significant upheaval across the globe. A major controversy recently erupted around the DGX Spark AI system. Legendary developer John Carmack sharply criticized the $4,000 device. This criticism immediately brought attention to severe performance failures and overheating issues. The developer’s comments quickly validated widespread user reports of device instability.

DGX Spark appears to be maxing out at only 100 watts power draw, less than half of the rated 240 watts, and it only seems to be delivering about half the quoted performance (assuming 1 PF sparse FP4 = 125 TF dense BF16) . It gets quite hot even at this level, and I saw a report…
— John Carmack (@ID_AA_Carmack) October 27, 2025

Carmack’s Critique and Overheating Reports

Carmack did not mince words regarding the DGX Spark inference machine. He claimed the device does not achieve its advertised performance targets. Furthermore, the unit reportedly suffers from severe overheating during extended use. According to Carmack, the machine maxes out at a mere 100W power draw. This low power draw suggests that the hardware is significantly underutilized. The developer also pointed to the growing dissatisfaction among the community. Developer forums have been overwhelmed with reports. Many users are complaining about system crashing and sudden shutdowns during heavy workloads. This widespread instability suggests major thermal and power delivery problems.

Technical Breakdown of Performance Issues

Industry reviews provide crucial technical context for the performance disappointment. The DGX Spark is a compact machine designed for local AI inference. It features a unique unified system memory architecture. This system shares a massive 128 GB of LPDDR5x memory between the CPU and GPU. While the memory size allows exceptionally large models to load, the limited memory bandwidth creates a bottleneck.

The LPDDR5x memory offers a limited bandwidth of only 273 GB/s. This limitation severely impacts the token generation throughput. For AI inference tasks, the DGX Spark is significantly slower. Comparisons show it is far behind a high-end RTX 5090 consumer card. The consumer card boasts much higher memory bandwidth, leading to faster token decoding performance. Ultimately, the unified memory bandwidth is the primary limiting factor in the system’s performance.

AMD Swoops In to Offer Developer Assistance

The public controversy provided a timely opportunity for a competing hardware manufacturer. AMD has reportedly “swooped in to help” the frustrated developers. AMD’s intervention involves providing direct support and assistance to those struggling with the DGX Spark’s technical issues. This move is a strategic and highly visible challenge to the market dominance of Nvidia. It aims to support the wider AI development community. The situation further highlights the competitive nature of the hardware industry.

This situation underscores the fierce competition in the AI inference hardware space. It also highlights the importance of real-world performance validation over marketing claims.

Source: Tomshardware