It is said that Nvidia will switch to a completely chipset design, or more commonly known as a chiplet design, in its new generation Blackwell GB100 graphics cards for HPC and artificial intelligence customers. This claim, coming from a reliable source, indicates that Nvidia will make serious changes in the new generation.Nvidia is also getting on the chiplet train
According to the latest rumors from reliable hardware leaker Kopite7kimi, two things stand out: First, Nvidia is expected to use the first chiplet design for today’s data center segment. To briefly recap, Nvidia Blackwell GPUs were initially expected to be the first family to go the chiplet route, until rumors suggested that the company was opting out and would use a more standard monolithic design.
Chiplet and Monolithic designs both have their advantages and disadvantages, but considering the cost and efficiency required to achieve performance increases today, chiplet and other advanced packaging technologies are used by competitors such as AMD and Intel.Nvidia has proven so far that the industry can move forward without chiplets with its Hopper and Ada Lovelace GPUs, both of which deliver the best performance per watt and the highest margins the company has ever seen. However, it seems that this situation will change in the future and the chiplet train will be started starting from Blackwell. Blackwell GPUs are scheduled to be released in 2024 for the data center and AI segment.
To use the chiplet design, it is necessary to bring together appropriate production types (such as TSMC 3nm, Samsung 4nm). Therefore, although it is a cost-effective process, it is critical that there are no problems with supply. TSMC’s leading packaging technology, CoWoS, is one of the key packaging technologies available to AMD and Nvidia, but both companies appear to be fighting for access to TSMC’s best technology. This fight often comes down to who can offer more money. So Nvidia has the trump card.
There are also other important components that need to be procured depending on the level of chipset-based integration Nvidia wants to use. Both AMD and Intel are making some advanced chipset packages that integrate several IPs on a single chip package, so it will be interesting to see how advanced Nvidia’s design for the first-generation chipset architecture in Blackwell is. The second thing that stands out is about the architecture of Blackwell GPUs. It is noted that the number of units such as graphics and texture clusters on Blackwell GPUs will not change much from Hopper, but the internal volume structure (SM/CUDA/Cache/NVLINK/Tensor/RT) will change significantly. Therefore, it is aimed to spread the performance gain to these foundations.
44705
