It is clear that the RDNA architecture has completely changed the fortunes of the Radeon Technology Group, especially since the fiasco of AMD Vega, the last of the GCN architectures which has become its particular Bulldozer. Fortunately, with the two iterations of RDNA, they were able to achieve a competitive architecture, with which to face NVIDIA having much less market share and resources, and they were able to renew contracts with the game console manufacturers. video.
What we knew about RDNA 3
Absolutely nothing is known about RDNA 3 at the official level, except for a device from AMD itself where it is said to go to a more advanced node than the 7nm one where the first two generations of RDNA went. A more advanced node not only means higher density in terms of transistors, but also higher clock speeds and / or lower power consumption.
AMD promises in RDNA 3 a jump in performance compared to consumption at the same level as that between RDNA 1 and RDNA 2. Where not only technologies such as Ray Trac ing and Variable Rate Shading and Infinity Cache have been integrated, but also in addition, the higher the clock speeds were reached.
Everything indicates that RDNA 3 will be launched under the 5nm node of TSMC, specifically the N5P process. AMD has no motivation to use the less advanced 6nm node. As it will be released on a par with Zen 4 which will use the 5nm node, it makes no sense for us to think about using another node.
AMD RDNA 3 architecture begins to take shape
Navi 33 (?) = Navi 21 + New Generation IP Core
– Yuko Yoshida (@KittyYYuko) May 2, 2021
Through a tweet from Yuko Yoshida, who is one of “initiates»More famous in the hardware industry, we are informed that the GPU Navi 33 has the same configuration as Navi 21, i.e. 80 maximum configuration calculation units. Which means it would be a transfer from the same Navi 21 although using TSMC’s 5nm node for fabrication. The mention of the Next Gen IP Core? We don’t know what this means, but there are two possibilities:
- Refers to newer versions of GPU coprocessors such as video codec, graphics card, DMA players, etc.
- The GPU will be made up of two different chips instead of one.
A few months ago, it was rumored that RDNA 3 would not be a monolithic design, but that it was going to be made up of two parts. One part would be in charge of rendering the graphics, the other would carry the memory controller and the coprocessors. Either way, this is something that has never been confirmed and is contradicted by AMD’s own patents.
In Navi 33 is the key
AMD when naming the chips for a generation of its RDNA or also known as Navi, it uses the following nomenclature: Navi XY, where X is the generation and Y is the variation of the chip, the highest values lows of And they correspond to the high range and the lower ones to the low range.
Navi 33 would therefore not correspond to a possible AMD Radeon RX 7800 or 7900, but to a lower model like an AMD RX Radeon RX 7700 and even a 7600. Let us not forget that AMD recently patented a two-way intercom system. GPU in MCM mode.
Considering that NVIDIA Lovelace is going to have a huge setup in terms of the number of SMs, there is no doubt that the AMD Navi 31 could be a GPU that can hold up to 160 compute units. Work for the software as one, as described in the AMD patent and be the return of two GPUs.