Specifications of NVIDIA's RTX series appear to be pretty hazy right now with everything from GA104 to GA102 being a possible suspect for this particular iteration. One of the most reliable leakers (for this series) has just posted that they believe the RTX 3080 will actually ship with a GA102-200 core that has 4352 CUDA Cores and 10GB vRAM. While the leaker is very reliable and I am sure we will see this GA102-200 GPU in one form or the other, I have tagged this post as a rumor since we aren't yet sure of the nomenclature.
NVIDIA RTX 3080 graphics card to get the GA102-200 treatment with 4352 CUDA cores, 10GB vRAM, and at least 15 TFLOPs of graphics horsepower?
We just recently published the first picture of the NVIDIA RTX 3080 and it is clear that it is open season for leaks. We are finally entering into the crunch time before the launch of the next generation of NVIDIA's RTX GPUs and I for one can't wait. The leaker in question, Kopite7kimi has been on the money for pretty much all of their NVIDIA Ampere leaks and while I am sure we will see this GPU under one name or the other (RTX 3080 SUPER/Ti for eg), they aren't a 100% sure of the nomenclature yet. Regardless, it is clear that NVIDIA is getting ready to launch some extremely powerful graphics cards.
RTX3080, GA102-200, 4352 CUDA, 10GB VRAM. I hope so. https://t.co/IbFO8VzfIU
— kopite7kimi (@kopite7kimi) June 6, 2020
Since the RTX 3080 is speculated to ship with a GA102-200 core, this would mean that it actually houses 4352 CUDA cores. At a conservative clock speed of just 1.75GHz, which NVIDIA has easily achieved in the past, this would mean it will output 15 TFLOPs at the bare minimum (TFLOPs is a function of Clock Rate * Core Count * IPC (which in this case is 2) / 1000000). That is an insanely high amount and would result in a very very powerful card - not to mention actual IPC gains. Initial leaks had indicated that the RTX 3080 would simply have a GA104 core, while later leaks suggested a GA103 core. All of that seems to have changed (and for the better) because NVIDIA is clearly prepping multiple flavors of the GA102 and one of these could potentially trickle down to the RTX 3080 (assuming NVIDIA doesn't add more naming schemes to their lineup - SUPER for example - which is a possibility).
An RTX 2080 is currently a $699 MSRP card and outputs around 10 TFLOPs of graphics horsepower. If NVIDIA keeps pricing the same, this would mean the RTX 3080 becomes 50% more powerful at the same price. NVIDIA can choose to do this for their entire RTX 3000 series lineup and something that would strongly incentivize upgrading to the new series by RTX 2000 series serial upgraders and those that had skipped a generation (or more). This also means that the RTX 3080 Ti flagship is shaping up to be a massive 18TFLOPs+ behemoth just based on these specifications alone - one which will ship with 21 Gbps GDDR6X memory. Couple this with a super-fast PCIe Gen4 SSD and you have a next-gen console slaying combination.
I can't help but think that NVIDIA being this aggressive with specifications has to do with the fact that the next generation consoles' specs have been revealed and the Xbox flagship ships with a 12 TFLOPs GPU. In order to maintain their performance supremacy and the ease of which they maintain it, NVIDIA would like to have at least two cards that can handily beat the console out. Jensen does not like being second and shifting to the 7nm process would have accorded him enough breathing room to make such a decision. Competition is good and it looks like the consumer will win again. It could also potentially have to do with their expectation of how AMD Big Navi will perform - since that is also expected to land this fall.