Information that NVIDIA did not plan to reveal for a few more weeks (months?).
A few days ago, NVIDIA was the target of a cyber attack. After a slight hesitation, the graphics card specialist confirmed the attack and gave some (rare) information about the data that was stolen. NVIDIA wanted to be reassuring, but things are gradually gaining momentum as new information is released on the Net. Thus, after the disclosure of the source code of the super sampling technique of NVIDIA - the famous DLSS - it is the specifications of the future range of GPU that are released in the wild.
NVIDIA has already mentioned its work on GPUs several times and we know that after the GeForce RTX 3000 currently on the market and based on the Ampere architecture, graphics cards using the Ada Lovelace architecture should arrive. According to the latest rumors, NVIDIA could distribute these new cards - which are already called GeForce RTX 4000 - as early as September 2022, but more interestingly, the technical specifications of the GPUs of these cards have been revealed. Of course, this information is marked with the seal of rumor, so it is advisable to remain cautious.
Technical specifications of NVIDIA 'Ada Lovelace' GPUs
Five references of GPUs have been unveiled from the AD102 to the AD107 through the AD103, AD104 and AD106. GPUs that, according to WCCFTech, are distinguished from the outset by the considerable increase in the second-level cache. On the Ampere generation, NVIDIA was satisfied with a maximum of 6 MB, even on the most powerful of its GPUs. On the other hand, we are talking about offering up to 96 MB of L2 cache on the AD102, the future flagship of the range. This GPU, which will probably be integrated into the GeForce RTX 4090, is also distinguished by the number of its Cuda cores (18,432). It will be equipped with 144 stream processors (SM) and a 384-bit memory bus.
The range continues with the AD103, admittedly less powerful, but still well equipped with its 10 752 Cuda cores, its 84 SM, its L2 cache of 64 MB and its 256-bit memory bus. The AD104 has 7,680 Cuda cores, 60 SM, 48 MB of L2 cache and a 192-bit memory bus. Surprisingly, the leaks do not mention an AD105, but they mention a more modest AD106 with 4,608 Cuda cores, 36 SM, 32 MB of L2 cache and a 128-bit memory bus. Finally, the AD107 closes this series of GPUs with 3,072 Cuda cores, 24 SM, still 32 MB of L2 cache and a memory bus that remains in 128-bit.
NVIDIA Ada Lovelace GPU specifications compared to NVIDIA Ampere GPU specifications
WCCFTech publishes a second small table to give a small idea of the power of these different GPUs compared to the current generation. We can see that the Ada Lovelace architecture promises remarkable gains, but it may also be accompanied by a very high power consumption.
