In the meantime, also in view of the 16 GB memory expansion of the new Radeon graphics cards, a GeForce RTX 3080 Ti with 20 GB memory expansion was long rumored, although the first leaks had already spoken a different language in the summer of 2020, well before the actual launch. Because there was talk of a card with 11 GB or 12 GB memory expansion and a memory interface of 352- or even the full 384-bit, as I had suspected in my news back in June 2020 and later specified (see link below).
Well then? It’s interesting that various leaks surfaced exactly yesterday, so on 05.02.2020 late in the afternoon (our time), so on the exact day when NVIDIA’s FAE (Field Application Engineer) informed the board partners (AIC) about the details and this story then immediately today, so one day later also found its spread via Twitter. Thus, one should be able to locate the leak rather in the NVIDIA headquarters, if one still takes into account the time difference. Because it also fits to this picture that at least one of the leakers writes about 10240 FP32 units.
What’s piquant about this, though, is that board partners have so far only been told that there’s a new, different SKU and that they need to plan on 12GB of memory expansion. On the other hand, a precise list of specifications has not (yet) been sent to AIC! That’s exactly why I won’t participate in this speculation about possible shaders with concrete numbers for the time being, because the possibility of pruning using hardware straps to create manifold variations is almost infinite. Therefore, the question remains open whether 80 SM are really set (where rumored) or whether it could be even a few less to move the card more in the middle between GeForce RTX 3080 and RTX 3090. But I’ll stay on top of that, of course.
The reasons for this are now also the subject of intense speculation among the board partners. On the one hand, it’s the very obvious lack of GDDR6X modules that should make such a card seem more interesting, and on the other hand, you can also significantly reduce the cost and effort of the board (and cooler) used. Because the 3080 Ti with 12 GB would only be equipped on one side, which can also be seen in the clearly simpler back (symbol image, below)
The victim of the deletion and the backport to the simpler board would then also be the accompanying deletion of the NVLINK port, which should then continue to be reserved for the GeForce RTX 3090 alone.
On the other hand, if the leaked number of shaders turns out to be correct, then NVIDIA has a solid internal problem, because the leak would then come directly from their own employees and not from the AIC area. If you add the time, then the group of people can even be narrowed down quite well. Whether the leaker did himself a real favor with the exact number of shaders remains to be seen.