The Blackwell GPUs, set to replace the H100 GPUs, are expected to deliver a 30-times performance boost while reducing energy consumption by up to 25% on some workloads. However, the overheating issue could potentially delay the new launch date, slated for early 2025, and affect customers' data center infrastructure deployment plans. The delay could also prove costly for Nvidia, which plans to sell the GPUs for up to $70,000 each and complete server racks for over $3 million.
Key takeaways:
- Nvidia's upcoming Blackwell GPUs, which are designed for data center racks, are reportedly encountering overheating issues, causing concern among customers.
- The overheating problem occurs when the GPUs are connected together in server racks designed to hold up to 72 chips at once, resulting in excessive heat that can cause operational inefficiencies or hardware damage.
- Nvidia has attempted to resolve the issue by asking suppliers to alter the design of their racks, but so far without success. However, the company has downplayed the issue, stating that such engineering iterations are normal and expected.
- The Blackwell GPUs are set to deliver a 30-times performance boost versus the H100 chips while reducing energy consumption by up to 25% on some workloads. Any further delays in their launch could be costly for Nvidia, which plans to sell the chips for up to $70,000 each and complete server racks for over $3 million.