Another delay. This time, it appears that Nvidia's Blackwell AI GPUs have fallen victim to last-minute delays.
Engineering samples of the B100 and B200 GPUs have already been delivered, but it appears that enterprise customers such as Microsoft, Meta, and xAI will have to wait for their orders. Tom's Hardware reports that two anonymous sources working on the chip first broke the news of the pushback, which was later corroborated by Bloomberg with anonymous Microsoft sources.
Previously, it was believed that Nvidia was aiming for a late 2024 release. Since several high-profile customers have placed large orders for Nvidia's latest mega-GPUs, with prices estimated to be $70,000 per "superchip," a significant delay is unlikely to be received favorably.
While these sources point to design flaws as the reason for the backlash, Dylan Patel, chief analyst at independent research firm SemiAnalysis, has posted a tweet telling a different story.
According to an update distributed to SemiAnalysis clients on July 22, CoWoS-L (Chip on Wafer on Substrate) start-up problems are the primary cause, and TSMC's AP3 packaging plant will be converted from CoWoS-S production to CoWoS-L production need to shut down in order to convert to CoWoS-L production. [This process is not expected to be completed until the second quarter of next year. The tweet also mentions rumors of overheating and firmware issues, but calls them "overblown". The update mentions various "teasing pains that need to be fixed," but SemiAnalysis does not appear to see them as the main cause of the delay.
Whatever the cause, a significant shipping delay could shake confidence in Nvidia's ability to deliver expensive AI processing GPUs in large volumes and on schedule.
Nevertheless, Nvidia is unlikely to worry about losing customers due to significant delays; AMD has its own competing AI GPU, the Instinct MI300, but AMD CEO Lisa Su recently said that the company is selling as many as it can currently make, and that the company will be able to deliver the product in time, She noted, "The overall supply chain is tight and will remain tight through 2025."
Given Nvidia's dominance in this space and the reported performance potential of the Blackwell series AI GPUs, it is unlikely that any major customer will switch to another hardware provider in the face of potential delays.
Once again, TSMC holds the key to the world's chip manufacturing capacity, and if it needs to reconfigure to keep up with demand, the world's other leading high-tech companies may just have to wait in line. Or perhaps there really is a design problem with NVIDIA's latest AI hardware that is slowing things down.
Still, at least it doesn't seem to be a printing press mistake.
Comments