, it was found from documents leaked last week that the Ada Lovelace architecture GPU will have 16MB of L2 cache per 64-bit memory bus, which is higher than the current Ampere architecture GPU’s 512KB per 32-bit memory bus. I don’t know if Nvidia refers to the idea of AMD’s Infinity Cache? This means that Nvidia will equip GeForce RTX 40 series GPUs with up to 96MB of L2 cache, which is far more than the 6MB of the current GeForce RTX 30 series.
In addition to the flagship AD102’s 96MB, the AD103 and AD104 are both 64MB, while the AD106 and AD107 are 48MB and 32MB respectively. Combined with the previous information,
AD102, AD103, AD104, AD106, and AD107 correspond to 144, 84, 60, 36, and 24 groups of SMs, respectively; the number of CUDA cores is 18432, 10752, 7680, 4608, and 3072 respectively; video memory widths should be 384, 256, 256, 192 and 128 bits.
It is rumored that AD102 is a large chip of 600mm², and its area is similar to that of GA103. Thanks to TSMC’s 5nm process, AD102 is equipped with 18,432 CUDA cores, which is a 71% increase on the basis of GA102 with basically the same area. Obviously, the new generation of GPU has a major upgrade in architecture, and it is likely that there will be a qualitative leap in performance. However, there are still many unclear places, such as the clock and the power of the new generation of graphics cards.