The Information’s exclusive report is considered nothing more than a repetition of old rumours.
A The Information recently came out with an exclusive report in which it was discussed that NVIDIA’s biggest customers, including Microsoft, Amazon, Google and Meta, are ordering fewer GB200 AI racks than planned because they overheat and do not ensure proper operation. Similar rumors came in a series the previous year, even NVIDIA admitted that there were problems during the design, but claimed that they had been fixed.
Most recently, in November, it was said that the DGX GB200 NVL72 rack server suffers from heat dissipation problems, which was also made public by The Information, mentioning that NVIDIA has redesigned the system’s cooling several times.
By the way, it is true that, considering the size of the design in question, it is a very serious task to dissipate 120 kW of heat, which means that you cannot build on previously proven cooling solutions, as they were not designed to handle this amount of heat. Therefore, it is certain that something very new had to be invented in order to optimize the cooling of such a system.
A Taiwan Economic Daily according to his report, at the same time, Taiwanese server manufacturers deny that there is any problem with cooling and claim that they are just repeating old rumors. According to them, everything is going according to plan in terms of production.
However, it is also true that Microsoft, Amazon and OpenAI cut back on their original orders, so they are buying and installing fewer GB200 designs than planned. For example, Microsoft originally counted on more than 50,000 cabinets, but they are currently asking for much less from systems based on the Blackwell architecture. At the same time, the affected customers did not comment on the exact reason why they did not claim their original orders, so the rumor about overheating has not been confirmed by official sources.
Source: prohardver.hu
