The network tells us about the GA103 and GA104 cores of the Ampere architecture - in a 7 nm manufacturing process סמסונג And with a big possible leap in performance
Exciting new rumors that have found their way to various corners of the network purport to provide us with preliminary data on the cardsGeforce RTX The Future of Series 3000 - With theRTX 3080 andRTX 3070 first, based on cores in architecture Ampere Enhanced and in the 7 nm advanced production process.
The RTX 3070 may be based on a kernel named Code GA104, with 48 SM type clusters and 3,072 CUDA units - which will make a big leap over just 2,304 CUDA units in the TU106 core of RTX 2070, without regard to the improved performance thanks The transition to a significantly efficient and minimized production process. Along with the growth in the number of processing units we may also encounter a doubling of the graphic memory volume - when a 256-bit memory controller combined with modern GDDR6 chips can allow us to see graphic cards With 8GB of memory Dynamic built-in default and 16 GB of memory Dynamic in the leading models.
The RTX 3080 will allegedly be based on a kernel named Code GA103, which will include up to 60 processing clusters and 3,840 active CUDA units - another massive leap compared to the current Turing generation, where 2,944 processing units and 3,072 processing units are found on RTX 2080 and RTX 2080 Super respectively. If that is not enough, the memory controller for this model may be 320-bit configuration, which will allow us to get built-in GDDR6 memories at extremely impressive volumes of 10GB or 20GB - so it would not be overkill to estimate that this Ampere model will provide Performence Significantly even better compared toGeforce RTX The company's leading 2080 Ti today (for gaming as a primary use).
If these core names match reality, we will probably also get a kernel named GA102 which will be the flagship product of NVIDIA In the new generation, you can use theGeforce The RTX 3080 Ti and even more exotic and expensive versions of the Titan, Quadro and Tesla brands - and may include over 5,000 active CUDA units, based on 80 SM clusters and above, considering the differences we could see in the generationGeforce RTX 2000.
It should be made clear that many of these details are purely speculative and speculative, and should therefore be approached with great skepticism - but if it is true that the announcement is already planned in about two months we expect to receive many more pieces of information to help verify or disprove the data in the previous paragraphs in no time . We will continue to monitor and update you as always!