The RDNA 3 architecture will mark a significant improvement in performance, and in efficiency


AMD continues to work on RDNA 3, its next graphics architecture and natural successor to the current RDNA 2. During the last weeks we have seen very interesting information that generally point in the same direction: this new architecture will mean the abandonment, by AMD, of the GPU designs with a monolithic core, and will mark the transition towards a totally MCM design, that is, a multi-chip module.

This transition will involve very interesting changes that, broadly speaking, will be similar to those we saw when Zen arrived. Until then, a processor integrated, in a single silicon chip, all the cores and I / O elements. Thus, for example, a Core i9-10980XE integrates its 18 cores on a single chip. With Zen, and especially with Zen 2, all that changed. AMD now uses small silicon chips known as chiplets to manufacture their processors.

Each chiplet has a total of 8 cores and 32 MB L3 cache, which means that to create a processor with 16 cores AMD has to join two chiplets. These chiplets are also supported by a third silicon chip that houses the I / O elements, and are interconnected by an Infinity Fabric communication system. A GPU with 10,240 shaders, for example, could be made with two chips of 5,120 shaders each.

Well, under a monolithic GPU design, all its elements are integrated into a single silicon chip. The RDNA 3 architecture will completely change that approach, bringing the chiplet concept to the graphics industry, meaning that a GPU will consist of at least two interconnected silicon chips, working together. AMD will have to overcome significant challenges in this regardAlthough with the experience you have accumulated with Zen, you shouldn’t have a problem. The launch of the first graphics cards based on the RDNA 3 architecture should occur between mid-to-late 2022.

RDNA 3 architecture would allow great improvements in performance and consumption

According to the latest information, the RDNA 3 architecture will be more efficient than Ada Lovelace, the architecture that NVIDIA will use in the RTX 40 and that will be the last that will maintain, in theory, a monolithic core design. Note that this information comes from unofficial sources, and therefore not absolutely valid, although it makes sense, as MCM designs have been shown to be more efficient than monolithic core designs.

AMD’s top of the range based on the RDNA 3 architecture will be the Navi 31 GPU, a chip that will, in theory, have 120 CUs (computing units) divided into two chips, that is, 60 CUs per chip, which would leave a total of 7,680 shaders. For comparison purposes, I remind you that the Radeon RX 6900 XT has 5,120 shaders. However, other sources say that AMD would be able to integrate Navi 31 into the GPU up to 160 CUs, which would equal 10,240 shaders, as long as the current division of 64 shaders per CU is maintained.

There are some contradictory opinions, but the important thing is that they all agree on four keys: RDNA 3 will have an MCM design, it will mark a huge increase in terms of power, a notable improvement in efficiency and will keep the monolithic core design on its mid-range graphics cards i.e Navi 33 chips which, for us to understand, will be the successor to the Navi 23, used in the Radeon RX 6600, which means that we are talking about the future Radeon RX 7600.

For the rest, it is rumored that AMD could mount up to 512MB of infinite cache on the Radeon RX 7900XT, provisional name by which its next top of the range is known based on the RDNA 3 architecture, and which will maintain the 256-bit memory bus, as well as the GDDR6 memory, although operating at higher frequencies. Nothing has been finalized, but it makes sense to think about speeds of 18 GHz, or maybe more. In terms of performance, RDNA 3 is rumored to outperform Ada Lovelace in performance under rasterization, in fact there is talk of up to 30% more, but at the same time it is said that NVIDIA would continue to win in ray tracing.

Like it? Share with your friends!


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