Nvidia’s largest clients for AI chips, such as Microsoft, Amazon, Google, and Meta, have allegedly reduced their orders for the latest Blackwell series of GPUs due to overheating problems. Nevertheless, this does not concern us regarding Nvidia’s newly unveiled RTX 50 lineup of gaming GPUs, which are also derived from the Blackwell framework.
The Information (paywall, via Reuters) asserts that overheating along with other associated “glitches” have prompted clients to postpone Blackwell orders or revert to Nvidia’s prior and presumably less troublesome “Hopper” generation of AI-optimized GPUs.
In October, Nvidia’s CEO Jensen Huang acknowledged that the new Blackwell series of AI GPUs contained “design imperfections,” which were resulting in delays in delivering AI racks to clients.
“The design flaw resulted in low yields. It was entirely Nvidia’s responsibility,” Huang stated. Those comments did not specifically mention overheating but instead referenced yields, which typically refers to the proportion of operational chips that can be produced from a larger wafer containing numerous GPU dies.
Consequently, these flaws may be entirely unrelated to whatever is causing the purported overheating, or they might be interconnected to some extent.
Indeed, reports about overheating Blackwell GPUs have been circulating since last November, with The Information revealing that Nvidia liquid-cooled racks housing 72 Blackwell GPUs had undergone multiple redesigns to address the issue.
Regardless, what we can definitively say is that there is no basis for assuming that any difficulties with Nvidia’s AI GPUs will lead to complications for its newly announced RTX 50 family of gaming chips at CES.
Indeed, these too are founded on the Blackwell architecture and constructed using the same TSMC N4 silicon. However, in terms of actual configuration and the number and balance of functional components, the gaming chips are fundamentally distinct.
Admittedly, there may be a concern, perhaps regarding the design of Blackwell’s Tensor cores, which might similarly affect the gaming GPUs. Yet, the likelihood of that is probably quite minimal. Additionally, the workloads and software that operate on a gaming chip are entirely different from those used in training or inferencing an AI model.
It is also worth mentioning that it is the same source, namely The Information, promoting this overheating narrative once more. Broader validation would lend credibility to the report. Nevertheless, according to Nvidia itself, there have been complications with Blackwell.
We will soon discover just how fast and indeed how hot Nvidia’s new RTX 50 gaming GPUs will be. Keep an eye on this space for our comprehensive reviews.