NVIDIA's next-generation Blackwell GPUs for HPC and AI will use chiplet design

NVIDIA's next-generation GPU architecture, codenamed Blackwell, is the successor to the current Ada Lovelace generation and the beginning of a new era of GPU design for the company. This is due to Blackwell being used for both the company's data center and AI side in addition to consumer GPUs for PC gaming and creators - a shake-up from the Hopper and Ada mix we see today.

2

Information on Blackwell has been swirling around the internet for a few years, with the early word indicating that the next generation of GPU architecture will adopt a chiplet design similar to AMD's approach with RDNA and the flagship Radeon RX 7900 Series. New info from NVIDIA insider kopite7kimi points to NVIDIA's Blackwell being very different at the top end - between the data center and consumer versions.

GB100, rumored to be the flagship Blackwell GPU, will adopt a chiplet or Multi-Chip Module (MCM) design, pointing to NVIDIA making a major change to packaging. With new info outlining that the consumer flagship GeForce RTX 5090 will use GB202, the PC gaming side will remain monolithic.

• Read more: GeForce RTX 5090 will be 1.7X more powerful than the RTX 4090 and feature a 2.9 GHz Boost Clock

Further differences indicate that CUDA Core counts will remain relatively unchanged on the data center side, which is surprising. However, the new GPU clusters are undergoing significant changes to accommodate the new MCM approach, leading to more flexibility to customize GPUs for various clients.

Interestingly, with most people now looking at 2025 for Blackwell, kopite7kimi does hint that it could still be on track for the second half of 2024. However, this is in relation to the data center side, with the consumer GPUs potentially eyeing a later release window.