Lights and shadows in the RTX 40 that could surprise everyone, or at least almost everyone. We had been anticipating it when we said that it was too hasty for the brand and manufacturers to launch headfirst into PCIe 5.0 when the bandwidth in their new GPUs would not even meet the requirements of the previous version, and it seems that it will finally be so. And it is that the RTX 40 will only have support for PCIe 4.0, where the news could be in their cores.
Intel‘s support with Alder Lake for PCIe 5.0 is merely testimonial at this time and so much so that there is no product on the market that supports it. Many thought that the new RTX 40 would change this, but the reality is that it will be the SSDs already in 2023 that will have the leading role under this version of the bus, because those of Huang have very different plans.
NVIDIA RTX 40 – PCIe 4.0 only, but with PCIe Gen 5 connectors
Intel’s turn to position itself as the most innovative company in the sense of the new DDR5 and PCIe 5.0 is simply a mirage that at the same time is being its death. The memories are through the roof and it seems that neither AMD nor NVIDIA are going to keep up with it on the most coveted bus of all.
Although the Ryzen 7000 will catch up with the Core 12 in this section, it seems that the graphics will be more conservative, starting with the green ones.
PCIe Gen4
— kopite7kimi (@kopite7kimi) April 24, 2022
And it is that according to one of the best leakers today, the RTX 40 will be PCIe 4.0 and not 5.0 as such.
This implies that the bandwidth available to them goes from 128 GB/s or 64 GB/s on each channel to half this speed, that is, 64 GB/s and 32 GB/s bidirectional . But this is only a glimpse of what is to come and should not surprise us as such despite the inconsistency of PCIe 4.0 for the slot and the Gen 5 connector up to 600 watts , needs of the script we could say in this case.
The three NVIDIA paths for the RTX 40
The most important thing is not this leak itself, but the fact that the AD102 chip (and we understand the rest of them as well) is described as not having been preconceived for something as simple as increasing the number of shaders in FP32.
In other words, the door is left open to the possibility that NVIDIA has many surprises up its sleeve, but in what sense? Well, here begins the purest speculation that we are going to channel into three different options and we will see if we are right in any later in time.
The first could be defined as a generational change in the SM, substantial modifications that could indicate changes in one or more lower units such as RT Cores or Tensor Cores. Second could be a change to the FP32 engine outright, perhaps increasing the number of FP32s and thereby modifying the total Shader count.
Finally we would have the changes in the third engine with FP32 and INT32 . Explained this we will understand where NVIDIA can go, because it is most interesting. When the Ampere SM was presented, NVIDIA graphically biased it into four main engines plus the RT Cores, which are independent within the rendering although really in the Ray Tracing work pipeline to show the results of the pixel and FPS in full is not such, because it is a rendering of Ray Tracing and mixed BVH tree and not individual.
Attention, I never talked about the details of SM.
Don’t preconceive that AD102 has a simple 18432-FP32.— kopite7kimi (@kopite7kimi) April 22, 2022
Here enter the three arguments focused in different ways. The new graphics of the SMs do not show 4 independent engines as such, but three. A unified engine in FP32 and INT32 , another pure FP32 and the Tensor Cores for AI, more then separate the RT Cores that work separately as we have explained.
SM reunification
What if NVIDIA restructured SM? That is to say, it could unify the FP32 and integers thereby increasing the total count to the rumored 18,432 Shaders (very much up in the air right now) and the theoretical space that the unification would leave using it with the RT Cores and Tensor Cores as closer engines given the supposed changes in the caches.
This would mean a surprising improvement in Ray Tracing and would bring it closer to a pure rendering, which is what is sought, but at the same time it implies a very profound change in the SMs that would affect the TMUs and possibly the ROPs really, not so much in number. , but in structure, hierarchies and of course its performance. Given how hermetic NVIDIA has become, we will have to wait almost until the presentation to see if we are focused on this topic.
