to leave a comment.
Semianalysis Report… "AMD, Google, etc. to Benefit from Easier Scaling"
An analysis suggests that the launch of Nvidia's next-generation AI server, from the leading AI semiconductor company, will be delayed by more than a year due to manufacturing difficulties.
Semianalysis, a semiconductor analysis firm, stated via social media service X (formerly Twitter) on the 6th (local time) that the launch of Nvidia's 'Kyber NVL144' has been delayed by more than 12 months, pushing it to 2028.
Semianalysis explained that this delay is due to difficulties encountered in printed circuit board (PCB) manufacturing.
Kyber is a server rack that combines 144 high-performance chips to function like a single computer, and it was slated to debut next year along with 'Vera Rubin Ultra', the next-generation AI system.
As an alternative, Nvidia proposed an 'NVL72x2 back-to-back' architecture, which would place two 72-chip racks side-by-side, but this was canceled due to customer backlash over its unusual design and immense operational burden.
Semianalysis also predicted that the 'NVL576' system, which connects 8 racks via optical communication, is likely to be delayed or limited to small-scale production due to technical challenges.
Furthermore, the launch of the Rubin Ultra chip with four computational dies was also canceled, leaving only the Rubin Ultra with two computational dies, performing at half the level, the analysis stated.
Due to these delays, Nvidia will find it difficult to scale up data centers for large-scale computation, and as a result, competitors like AMD and Google could gain a reflective benefit, Semianalysis projected.
Nvidia did not respond to CNBC's request for comment on this analysis.
Newsletter
Get key news delivered to your email every morning
to leave a comment.