Celebrity

NVIDIA’s Next-Gen Blackwell GB100 GPUs Utilize Chiplet Design, Feature Significant Changes

admin
by admin September 17, 2023

NVIDIA's Next-Gen Blackwell GB100 GPUs Utilize Chiplet Design, Feature Significant Changes 1

NVIDIA's next-gen Blackwell GB100 GPUs for HPC & AI customers are rumored to go fully onboard with a chiplet design according to Kopite7kimi.

NVIDIA Rumored To Be Going All Onboard The Chiplet Train With Next-Gen Blackwell GPUs For AI, GB100 To Offer Significant Changes

The latest rumor states two things, the first is that NVIDIA is now expected to utilize its first chiplet design for the modern-day data center segment. So just as a recap, the NVIDIA Blackwell GPUs were initially anticipated to be the first family to go down the chiplet route until rumors reported that the company had decided against it and was meant to use a more standard monolithic design. Chiplet and Monolithic designs both come with their advantages and disadvantages but given the cost and efficiency required to achieve performance uplifts today, chiplet and other advanced tech for packaging are being utilized by competitors such as AMD and Intel.

NVIDIA GPU Roadmap. (Image Source: NVIDIA)

NVIDIA has so far proven that the industry can move forward without using chiplet with its Hopper and Ada Lovelace GPUs both excelling in delivering the best performance per watt & the highest margins that the company has ever seen. But moving forward, that's going to change, and starting Blackwell, we might just see NVIDIA's first chipset-packaged design. The Blackwell GPUs are so far scheduled for a 2024 release for the datacenter and AI segment.

Kopite7kimi puts the emphasis on the Data Center and AI GPUs when talking about Blackwell. This shows that NVIDIA may not yet move over to chiplets for its gaming GPUs codenamed "Ada-Next" but incorporate a level of advantage packaging technologies within its Data Center and AI GPUs to maximize chip output. As mentioned above, chiplets come with their disadvantages and those usually lay within sourcing the proper factories to package these chips.

TSMC's CoWoS is one of the key packaging techs that's available to GPU clients such as AMD and NVIDIA but it looks like both companies may be fighting to get access to TSMC's top tech. The fight usually comes down to who can offer the most sum of cash and the green team is currently swimming in AI money. Furthermore, there are other key components that need to be sourced depending on the level of chipset-based integration that NVIDIA wants to use. Both AMD & Intel are doing some advanced chiplet packages, integrating several IPs on a singular chip package so it will be interesting to see just how advanced NVIDIA's design is for its first-gen chiplet architecture on Blackwell.

The second part of the story comes down to the internal architectural structure of the Blackwell GPUs. It is reported that the number of units within Blackwell GPUs such as GPCs, TPCs haven't changed much from Hopper but the internal unit structure could be hinting towards the SM/CUDA/Cache/NVLINK/Tensor/RT count has significantly changed. Previously, we have seen at least two GPUs leak out which include the Blackwell GB100 and GB102. The second one can either be a data center or a gaming GPU but Kopite7kimi has already said that the consumer (gaming) parts will fall under the GB200 series and not the GB100 series.

There are also rumors that NVIDIA is evaluating the Samsung 3GAA (3nm) node which may enter mass production in 2025 though Kopite7kimi believes that NVIDIA may not change its plans and stick to TSMC for its next-gen GPUs. The same leaker previously reported that Blackwell won't be using a 3nm process node. Given the AI and Data Center progress that NVIDIA has achieved since its Pascal GPUs and the extreme success with Ampere and Hopper GPUs, Blackwell will mark one major advancement to NVIDIA's line of chips, propelling the company into the next era of AI and computing.

Written by Hassan Mujtaba

Post a Comment

0 Comments