Nvidia Reportedly Rejects Samsung HBM3 Chips Due to Overheating and Power Issues

Published by

teaser
According to internal sources cited by Reuters, Nvidia has tentatively declined to use Samsung's most recent High Bandwidth Memory (HBM) chips in its artificial intelligence (AI) GPUs due to issues with overheating and excessive power consumption. This development concerns Samsung’s HBM3 and the upcoming HBM3E chips, which represent the fourth and fifth generations of the HBM standard, respectively. These chips are primarily utilized in GPUs tailored for AI applications. The sources reveal that these HBM3 and HBM3E chips are slated for release this year, with production expected to commence in the first half of 2024. The specific reasons behind the chips' underperformance in Nvidia’s evaluations have not been publicly disclosed by either company. However, insiders claim the chips exhibit rapid overheating and high power usage, which are critical drawbacks for efficient GPU performance.

Samsung has been engaged in efforts to meet Nvidia's testing criteria since last year. Despite these efforts, recent tests conducted in April on the 8-layer and 12-layer HBM3E chips were unsuccessful. The feasibility of addressing these technical issues remains uncertain, with sources expressing concern about Samsung's competitive position. Industry competitors like SK Hynix and Micron have also begun production of HBM3E chips, intensifying the pressure on Samsung. In response to queries from Reuters, Samsung emphasized that HBM technology is a specialized memory product that necessitates extensive optimization in collaboration with its customers. The company reiterated its commitment to optimizing its products through close partnerships with its clientele. Following the initial Reuters report, Samsung refuted claims that its chips were failing due to thermal and power management problems, asserting that its testing process with Nvidia was progressing smoothly and according to plan.

Image

This situation highlights the challenges in the high-stakes market of high-performance memory chips for AI applications, where efficiency and reliability are paramount. 

Source: Reuters

Share this content
Twitter Facebook Reddit WhatsApp Email Print