New information from a major leaker and new controversy straight from Twitter, where something really surprising is specified that could somehow lead us to look into the past. Apparently AMD and NVIDIA have their “ultimate weapon” for the GPU gaming market in their arsenal, especially the red team, as they could showcase a GPU with 2 GDCs in a single interposer. Fact or fiction? doRX 7950 XT with Navi 31?
The leak emphasizes something important and that’s where all the controversy stems from, as it’s said that the chip that was destined to rule will be nothing more than the princess of the castle. The Navi 31 is therefore not the best that AMD would have in the pipeline, but how is it possible?
Specs TBD, no name, but it definitely wouldn’t be Navi 31
The 31 is not the flagship of the 3X, the fastest will debut in 2023.
Since RGT had already said a 2GCD product, I mistakenly thought that was the flagship. But it is certain that the specifications of the flagship are still to be determined.
— Greymon55 (@greymon55) June 27, 2022
And here is the first controversy. These are the two main AMD leakers confirming the same point of the news and how one can read it says very clearly that Navi 31 is not the flagship from AMD, but it would be another fully parallel chip that would take the performance crown.
So what would AMD hide? Well take enough salt, to taste, season and spoil, because apparently the fact that Navi 31 is a set of chiplets as seen almost a month ago doesn’t take away one iota of intent AMD to make the first desktop GPU purely on chiplets, i.e. two GCDs in one package with their corresponding MCD, which should be twice as many units or the same size but larger to hold everything needed for the Infinity cache and I/O array.
There are critical voices questioning this statement, mostly saying that there is no other chip beyond Navi 31 on the horizon, but why the credibility of such a famous leaker as greymon55?
The problem of 2 or more GCDs in the same package of RX 7950 XT and Navi
You have to think two ways, whether it’s true or not and whether you’re wrong. If you’re right, then what we would have is more power with much more than 2x consumption itself, unless the frequencies alleviate the deficit we’re going to talk about.
And does adding a second GCD in the interposer lie in very straightforward issues. We ignore the price because it is something logical, but we cannot ignore the fact that the SVT they will be much deeper because the substrate has to work with many more intersections at different levels. A large interposer to separate the two GCDs results in higher latency and also a much higher energy consumption, since the distance also influences the energy expenditure to maintain the integrity of the signal, which is obtained depending on the voltage and the best materials used.
Surely we would be talking about an I/O matrix that acts as an intermediary and a connector for the two GCDs and therefore either we have to talk about a thread manager in the driver or in the operating system, or in both . It is also possible for the MCDs to be shared by the two GCDs, which would simplify the architecture in exchange for a larger total size in terms of area and features.
Remember that we are not talking about a Navi GPU (whichRX7950XT?) Dual Core as before, but only one interposer with two large chiplets, which makes a lot of difference, especially in terms of efficiency.