According to Kopite7kimi, it is rumored that Nvidia’s next-generation Blackwell GB100 GPU for HPC and AI customers will fully adopt Chiplet design. The latest rumors have two points: First, Nvidia is expected to use chipset design for the first time in the modern data center field.

Visit the purchase page:

JD.com NVIDIA series product summary

So, we can recall that Nvidia Blackwell GPUs were originally expected to be the first product line to go the chipset route, until rumors surfaced that Nvidia decided not to go the chipset route and instead go with a more standard monolithic design. Chipsets and monolithic designs have pros and cons, but given the cost and efficiency currently required to achieve performance improvements, chipsets and other advanced packaging technologies are being adopted by competitors such as AMD and Intel.

So far, Nvidia has proven that both its Hopper and AdaLovelace GPUs deliver the best performance per watt and highest profit margins the company has ever seen without using a chipset. However, that will change over time, and starting with Blackwell, we may see Nvidia's first chipset package designs. So far, Blackwell GPUs are scheduled to be released in 2024 for the data center and artificial intelligence fields.

When talking about Blackwell, Kopite7kimi focused on data centers and AIGPU. This suggests that Nvidia may not yet adopt a chipset for its gaming GPU, codenamed "Ada-Next," but will use a certain degree of advantageous packaging technology in its data centers and AIGPUs to maximize chip output. As mentioned above, chipsets have their drawbacks, and those drawbacks often lie in finding the right factory to package these chips.

TSMC's CoWoS is one of the key packaging technologies available to GPU customers like AMD and Nvidia, but it looks like both companies may be vying for TSMC's top technology. The battle often comes down to who can provide the most funding streams, and NVIDIA is currently swimming in AI money. Additionally, depending on the level of chipset-based integration Nvidia wants to use, other key components will need to be sourced. Both AMD and Intel are doing some advanced chipset packaging that integrates multiple IPs on a single chipset, so it will be interesting to see how advanced Nvidia's first-generation chipset architecture design is with Blackwell.

The second part of the rumor is the internal architecture of the Blackwell GPU. It is reported that the number of GPC, TPC and other units inside Blackwell GPU has not changed much compared with Hopper, but the internal unit structure may imply that the number of SM/CUDA/Cache/NVLINK/Tensor/RT has changed significantly. Prior to this, we have seen at least two GPUs leaked, including the Blackwell GB100 and GB102. The second one could be a data center or gaming GPU, but Kopite7kimi has already said that the consumer (gaming) part will belong to the GB200 series, not the GB100 series.

In addition, there are rumors that Nvidia is evaluating Samsung's 3GAA (3nm) node, which may enter mass production in 2025, but Kopite7kimi believes that Nvidia may not change its plans and stick to using TSMC to produce the next generation of GPUs. The same leaker also previously reported that NVIDIA will not use the 3nm process node. Given the progress Nvidia has made in artificial intelligence and data centers since the launch of Pascal GPUs, and the huge success of Ampere and Hopper GPUs, Blackwell will mark a major advancement in Nvidia's chip product line and propel the company into the next era of artificial intelligence and computing.