New data from China, juicy new leaks that suggest that we will have a real generational leap in the new next-generation graphics cards. At NVIDIA with Blackwell we already saw the first signs, and today it is AMD’s turn. The data is really difficult to believe, since it represents an impressive leap forward, but for that, take salt and let’s go to trouble, because it is concrete and very relevant data. These are the first details of the AMD RX 8000: Shaders, clock and VRAM.
As we know, AMD has not wanted to compete with NVIDIA in this generation of graphics cards in pure performance, and has focused on a better performance/consumption/price ratio, which is giving it better sales than the green ones. There was talk at the time of internal problems of architecture. Failures that we have not been able to reveal, and that point to latency as a determining factor. If we put all this together we are left with RDNA 3, but with RDNA 4 it seems that AMD is going to go with everything.
AMD RDNA 4, is such a brutal performance jump possible?
We don’t want to get into unnecessary hype so we’ll try to keep the soufflé (F1 fans to me) in check. With RDNA 3 vs. Ada Lovelace we have seen several things. The first is that NVIDIA has a better node and that AMD preferred a better price with N5. The second is that the new node for RTX 50 and RX 8000 seems to be the N4 in an evolved version similar to that of NVIDIA. Finally, under this node the greens managed to include up to 18,432 Shaders.
All this is relevant for today’s leak, because according to this the RDNA 4 architecture can determine how far it will go, and we will explain. Starting with the node, the TSMC N3 it seems that it will finally be available to arrive in high volume for both companies next year according to this rumor. This will imply that AMD can really scale in the number of minimum units and with it in Shaders, without compromising consumption.
RX 8000: up to 18,432 Shaders, much higher frequency and much more VRAM
And is that if this is true, the changes in the higher-end RX 8000, namely, RX 8900 XTX, they can know even a little. The number of UC increases up to 144, and here we have to talk about something really interesting. The number of Shaders is specified as 18,432, exactly the same figure as the GH100 at 4 nm. Why is the number striking? Because AMD still counts individual CUs as having 64 Shaders. Therefore, if the new Navi 41 has 144 CU the normal thing would be to get 9,216 Shaders and not 18,432 in the largest of the RX 8000, which, as you may have observed, is exactly half.
AMD makes up for it by saying that they now have wave64 thanks to the second ALU FLOAT with Matrix as SIMD32, so to do the performance calculations FP32 multiply by four instead of two for the final value. NVIDIA internally changed the SMs and now counts INT and FT units as the same, since it ensures that they can work with integers and float at the same time regardless of the specific unit. So it goes from 64 to 128 Shader count.
After this, what does it matter if the number is one or the other? Well, it would mean that there are very important changes under the hood in RDNA 4 precisely because AMD would have gone from one count to the other. That is, there would be a second complete Float/Int/Matrix unit with double execution capacity, and therefore, it does not multiply by 4 the performance in FP32, but rather directly in the number of Shaders.
MCM 2.0 architecture, latency reduction and more VRAM for the RX 8000
There is more data to process. First of all, we already talked on several occasions about the overclocking potential of the architecture and how AMD had lowered frequencies in pursuit of lower consumption, but the advantages per MHz were there. There is talk that the reds will increase their frequency for the RX 8000 to 3.5GHzwhich would mean a rise in Exact 1 GHz on Boost for the hypothetical RX 8900 XTX.
As if that were not enough, AMD continues with its philosophy of improving a 50% performance per wattat least, and as it will also continue with architecture LCM based on dcm (at 5nm?) and GCDwe can only think that the increase in Shaders will come with a change and improvement in access times from external caches based on a larger data bus such as Infinity Fabric.
This is also determined by the increase in VRAM, which is specified as 48GB in the top model (RX 8900 XTX) and 32 GB in the bottom model (RX 8900 XT) depending on the MCDs that are integrated. It is not specified if this VRAM will be GDDR6, GDDR6X or directly GDDR7which is most likely said in passing.
Finally, the performance jump seems that, if what has been said is fulfilled, it will be great. From the RX 6950 XT to the RX 7900 XTX there was an approximate jump of between 20% to 25% depending on where you look at it. Given these data, regardless of the Shader count as such between 9,216 or 18,432, it seems that we will have a higher GAP, and without a problem, at 50% real. We will see if the leaks are correct with these RX 8000.
