Little by little we seem to be getting closer to the truth. The leak, although sometimes failing, has led to this and for months in several buildings today, it has been reaffirmed.
And it looks like NVIDIA will provide everything from scratch with a TOP graphics card promising to clear from the blows seen with Trance.
NVIDIA RTX 3090: 12GB of 18Gbps VRAM with 384-bit bus
Filtering is more focused on VRAM memory than at any other stage, but also leaves a very important brushstroke that we will analyze over time. As we saw in the last two weeks, NVIDIA looks like it's building the latest modules in between Samsung at GDDR6 at 18 Gbps, and will try to close GAP that it has always had with AMD and its high-tech HBMs in terms of bandwidth.
And that in this data we have to add another leak suggesting that NVIDIA will move the tab to the VRAM size part, because it means that the RTX 3090 will come with it 12 GB instead of 11 as usual.
For pure statistics, this should leave a 384 bit bus, added to previous data will provide full bandwidth 864 GB / s. To obtain this data, the RTX 2080 Ti currently offers 616 GB / s, which indicates that the bandwidth at Ampere over the RTX 3090 will increase 40.25%.
Up to 39 TPC on the GA102 chip, a new Shaders count
The leak is progressing and they say that NVIDIA will provide a new TPC figure for this GA102 chip. And that this RTX 3090 will come with it 39 Configuration User Interface Collection to do the calculation 4992 CUDA Cores (Shaders).
This is quite surprising, since it only involves 3.21% more than what was seen on the RTX 2080 Ti and could only be said to be 3 TPC than the chip it covers (TU102). So, and if that's true, NVIDIA would have to provide a much higher frequency compared to trying to Make enough difference to ensure a jump from Ampere architecture, IPC, and unique architecture development.
And that they don't foresee such a big change, far from it, like the one imagined by Pascal, but rather, a small adjustment of the whole set at a much higher speed that will eventually be used at levels similar to those seen so far.
Thus, the data suggest an increase in the frequency of an increase close to 2 GHz and an equivalent use of 250 watts of this model, something NVIDIA should be able to accomplish without too much trouble under 10nm from Samsung.
In any case, the data do not look as much as expected, as shaders are expected to rise significantly, and for them the units of TMU, ROP, RT Cores and Tensor Cores are "only" very often.
This way and if this is true, everything will be in the hands of this parameter and its MHz. Will this be enough to stop AMD and its new GPUs?