Moore’s law is dead, or so the NVIDIA CEO claimed a few years back. The fact that chip miniaturization has started hitting the ceiling is no surprise. Intel has been struggling to move to the 10nm node over the past 4-5 years, and even though the first Ice Lake chips based on it have started shipping the yields aren’t exactly great.
AMD, on the other hand, has moved to the 7nm node and doesn’t seem to have any major supply issues. Why? Well, the fact that they outsource the fabrication to TSMC does greatly help, but the reason for the impressive price-performance ratio is the company’s MCM (Multi-chip module) or chiplet design.
This has allowed AMD to bring HEDT levels of performance to the consumer market without charging a fortune for it. The Epyc CPUs are also challenging Intel’s supremacy in the server spaced for the first time in over a decade.
Knowing your enemy is half the battle or so they say. NVIDIA is doing just that and then taking the next step. Using the same technology as rival AMD in their next-gen Hopper graphics cards (set to launch after Ampere in 2022).
This is still a rumor so take it with a grain of salt, but analyzing industry trends I’d say it’s very likely. Furthermore, this design suits GPUs far more than CPUs. GPUs are more parallel in nature while CPUs tend to be sequential. In other words, CPUs pack a dozen or fewer cores and are focused single-threaded performance while GPUs consist of up to several thousand cores, running in tandem, doing different calculations (graphics, compute, BVH, etc) simultaneously by dividing the workload between the various components.
The MCM design does induce a latency penalty which can inversely impact sequential tasks that favor higher IPC and faster response times. The result is that AMD chips are relatively slower in games compared to Intel’s. Although GameCache greatly makes up for this issue, it is still no doubt not completely eradicated.
With graphics cards, such a latency increase won’t have as much of an impact as GPUs focus on parallel workloads rather than a sequential speedy pipeline. One of the main benefits of this design will be better yields, resulting in larger, more powerful GPUs at reduced costs.
WCCFTech did a small experiment and found that migrating a medium-sized 815mm² monolithic GPU to the MCM-design results in a 116% yield gain with much lower wastage. The benefits obviously increase as we move to larger, more powerful chips. This may still be a rumor, but we’re quite confident about the legitimacy and seeing the success of AMD’s Zen microarchitecture, I’d say it’s only a matter of time before other chipmakers adopt the same technique.