GPUs

NVIDIA’s Next-Gen GPUs to be Called Hopper: An MCM Design (Like AMD’s Ryzen) with as Many as 10,000 Cores?

These days everyone is talking about AMD’s Ryzen 3000 chips. One of the main building blocks behind the 7nm Zen 2 was the MCM or chiplet design. Multiple CCDs (Core Complex Dies) and a Central I/O Die, connected by the Infinity Fabric Interconnect. This increases yields and reduces wastage significantly, essentially making twelve-core and sixteen-core consumer CPUs possible. Regardless of what Intel might say, they’ve already adopted the MCM design in various projects (Lakefield, & Ponte Vecchio).

Moore’s Law and the Chiplet Approach

As Moore’s law dies a slow death, most chipmakers are following AMD’s lead and that includes arch-rival NVIDIA as well. Uncle’s Jensen’s team is busy designing the next-gen Hopper GPUs (set to launch after Ampere in 2022) using the multi-chip module (MCM) approach. This means multiple dies connected by an Infinity Fabric-like Interconnect, resulting in a beastly GPU packing as many as 10,000 cores.

This isn’t that inconceivable. After all, Intel’s Ponte Vecchio GPUs for Exascale will also leverage the same design. Two Sapphire Rapids CPUs and six Xe GPUs connected by various interconnects, supported by a ton of cache.

MCM Design and the Muli-GPU Problem

There is one core problem with this design, namely how to sync the various chiplets. In CPUs, that’s not a problem as the workloads are sequential but GPUs are highly parallel in nature. Thousands of cores in tandem working on different phases of the pipeline to deliver those life-like graphics. Latency is a key concern here. Synching multiple GPUs has always been a challenge, and that’s exactly why multi-GPU technologies like SLI and XFX have largely failed in the past.

Creds: WCCFTech

However, there’s still ample time for the industry to work towards a solution and usher in the next generation of GPUs, based on the MCM design. This will greatly increase scalability, reduce costs and as well as enable complex effects such as real-time ray-tracing. Not just for lighting but the whole scene. This was unimaginable till now, but with that kind of firepower, it won’t be hard to implement.

Tags

Areej

Computer Engineering dropout (3 years), writer, journalist, and amateur poet. I started Techquila while in college to address my hardware passion. Although largely successful, it suffered from many internal weaknesses. Left and now working on Hardware Times, a site purely dedicated to. Processor architectures and in-depth benchmarks. That's what we do here at Hardware Times!

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Back to top button
Close