AMD is now in a very strong moment, but above all it has a close relationship with TSMC after the decision of some large companies to turn to other suppliers such as Samsung, as have NVIDIA and Qualcomm. On the other hand, with Apple having made the jump to 5nm, that leaves TSMC’s 6nm node fully available to AMD.
What is TSMC 6nm node?
TSMC’s 6nm manufacturing process, known as N6, was announced by the Taiwanese mega-foundry in April 2019 as an easy transition node to using EUV technology.
Unlike the N7 + node, the N6 node allows manufacturers to use designs for the standard 7nm manufacturing process, meaning AMD can redeploy its portfolio of 7nm engineering cores under the new node.
TSMC did not give any advantage in terms of speed and consumption to use the new node, but in terms of density where node N6 is 18% denser than N7, which means being able to remove more chips per wafer or being able to add additional elements to create more powerful variants of known CPUs or GPUs.
AMD plans go through 6nm
Although it is not official, in some cases the leaks are controlled by the manufacturers themselves and we have evidence through a version of Zen 3 under TSMC’s 6nm node.
This version is called Zen 3+. We don’t know how it differs from the standard version of the Zen 3 or if AMD took the opportunity to make some last minute tweaks that raise the CPI a bit, what we do know is that we’re going to see a new one. AMD series of processors and SoCs use Zen 3+ kernel.
On the one hand we have the Rembrandt-H SoCs, which should replace the Cezanne-H which will be presented in a few days, on the other hand Warhol, which would replace the current Ryzen 5000 desktop.
Is 2001 a year of transition for AMD?
A few years ago AMD and GLOBALFOUNDRIES entered into an agreement to manufacture the IOD, which includes the Northbridge and Southbridge of the Zen 2 architecture and which was also used in the Zen 3. It is precisely in this year 2021 that the contract between AMD and Global Foundries will end.
If Warhol processors are to be compatible with both DDR5 and PCIe 5 memory, then a new IOD is needed, different from Global Foundries, this leads us to building a new Zen processor that we could see on sale at fin 2021 or early 2022.
This means that AMD instead of going for a direct generational switch to adopt DDR5 in its processors, Alder Lake, AMD taking advantage of the chipset configuration of some of its models AMD Ryzen can easily build DDR5 compatible variants. without making a drastic change to the chip.
¿AMD RDNA 2+ at 6 nm for 2021?
One of the things that has filtered through the roadmap is SoCs with RDNA 2 integrated GPUs, which shouldn’t surprise us given how easy it is to port designs from TSMC N7 to node N6. .
In the same way that AMD apparently ported Zen 3 to said node, it can also do so with RDNA 2 and take advantage of the additional 18% density to launch a new generation of AMD graphics cards. After all, an annual launch of new graphics cards is not a crazy idea, not always under the same architecture but it has always been done traditionally.
But is a configuration beyond the Navi 21 possible? Yes it is possible and we have the answer in Navi 14 used in the RX 5500 and RX 5300 which had a configuration of 12 compute units per Shader Array instead of 10, if the same configuration was applied in an RX 5900 XT then we would have a 96 compute unit GPU.
The other possibility is AMD RDNA 3
From the presentation they made last October, they talked about RDNA 3 under an advanced manufacturing node more advanced than 7nn, the problem is that the most advanced node claim can include 6nm and the time map in the slide is part of the product launch period under node N6 by AMD.
AMD can incorporate a number of enhancements into RDNA 2 in order to improve architecture performance and be able to beat NVIDIA on all fronts, not only in rasterization, but also in Ray Tracing as well as Artificial Intelligence. .
An RDNA 3 for 2021 is not impossible, we know that Navi 31 has been around since it was filtered in the pilots a month ago, but without going into the impossible, which we can expect from the third generation RDNA 3 are things like the following:
- The ray accelerator unit would have been improved, now instead of performing 4 box intersections and 1 triangle intersection, it can perform 4 box intersections and 2 triangle intersections.
- AMD would have integrated CDNA “Tensor Cores” in this RDNA 2 improvement.
- GDDR6X instead of GDDR6, but with densities of 2 GB per chip.
What we doubt the most is doubling the number of ALUs per computing unit, this implies much deeper changes in the GPU that require a much longer development time, but AMD, improving some points of RDNA 2, can deploy a third generation of your RDNA 3 in less than a year.