On Sunday local time, The Information reported that Nvidia's new generation Blackwell processors have serious overheating issues in high-capacity server racks. These issues have led to design adjustments and project delays, causing concerns among major clients such as Google, Meta, and Microsoft about whether Blackwell servers can be deployed as planned.
Insiders revealed that Blackwell GPU is designed specifically for AI and high-performance computing (HPC), but it has experienced overheating issues in servers configured with 72 processors, with each rack consuming up to 120 kilowatts of power. The overheating issue has forced Nvidia to modify the rack design multiple times, which not only limits GPU performance but may also damage hardware. Customers are concerned that these technical issues may delay the processor deployment process in the data center.
The report stated that according to Nvidia employees who have been committed to solving this problem, as well as customers and suppliers who are familiar with the issue, the chip manufacturer has repeatedly requested its suppliers to change the rack design to address the overheating problem, but the report did not disclose the names of the suppliers.
NVIDIA is partnering with leading cloud service providers, which is an essential part of our engineering team and processes. Engineering iterations are normal and expected, "a spokesperson for the company said in a statement to the media.
In October, Nvidia CEO Huang Renxun stated that with the help of long-term partner TSMC, the design flaws of Nvidia's latest Blackwell AI chip have been fixed. He expected that Blackwell chips would be shipped in the fourth quarter. According to media reports, the improved Blackwell GPU is expected to be shipped as early as the end of January next year. This means that the new generation of Blackwell processors may be delayed for three months due to overheating before delivery.