AMD Radeon: Chiplet-Based GPUs Like Ryzen CPUs?
Source: HW Upgrade added 04th Jan 2021
After CPUs, AMD seems intent to drastically change the design of their GPUs. The company is studying an MCM project that will allow multiple graphics chiplets to work properly together. The idea is in a newly published patent.
by Manolo De Agostini published 04 January 2021 , at 11: 01 in the Video Cards channel
AMD Radeon RDNA
AMD was the first to believe and adopt a MCM project for its microprocessors, from Ryzen consumer models up to the EPYC server proposals. The US company has highlighted how to separately produce the components of a CPU, untying the cores from the memory controller and other interfaces, not only favors a landing on the market quickly and predictably , but it also has advantages in terms of performance and consumption.
The MCM (multi-chip module) design, where a processor is composed of different blocks, also produced with different processes and connected via high-speed bus, allows for a more innovative approach to design, and this has allowed AMD to recover the gap with Intel and surpass it in different sectors of market, thanks to the possibility of increasing the number of cores through the creation of dedicated chiplets with advanced production processes. There is also an advantage in terms of production yields , as producing chiplets with the most appropriate process allows have fewer defects.
It is clear how the time of monolithic chips , i.e. a highly integrated project that contains all the various aspects of a microprocessor on the same die, both at the drips : it is less and less functional to innovations future and this not only applies to CPUs, but also to GPUs . Both AMD and Nvidia have long toyed with the idea of moving to MCM design and this transition may be closer than we think .
A patent issued by AMD titled “ GPU Chiplets using high bandwidth crosslinks “exposes the idea, or at least one of the ideas on the table, that the company led by Lisa Su is stroking to make the project of a MCM GPU . In the documentation, AMD illustrates some of the reasons it has not taken that route so far : the high latency of communication between chiplets, programming models and the difficulty in implementing parallelism.
However, the company’s engineers think they can bypass these obstacles using an in-package interconnection called “ high bandwidth passive crosslink “, able to allow any “chiplet GPU” to communicate directly with the CPU and at the same time with the other chiplets. Each GPU, in addition to the cores, would have its own cache and the necessary to work autonomously: in short, each chiplet would represent a complete GPU, fully manageable by the operating system.
Unlike what we observe with AMD CPUs, where the company has crammed the cor ex 86 in a dedicated chiplet and everything else in an I / O die, in this case the idea looks more like making smaller GPUs and making them work together . One of the growing problems nowadays is that the most powerful graphics chips occupy such a large area that they require advanced and very expensive production processes, with yields not always exceptional.
Creating smaller GPUs would reduce production costs and have better yields, while the new interconnection would preserve performance and at the same time the software could make the most of the hardware, without particular modifications. At the moment it is not clear if and when this project will become a concrete reality, but it is rumored that this will happen after the solutions based on RDNA 3 architecture, expected between this year and next.