Nvidia’s hacked data seems to contain quite a bit of information about the next generation of video cards, which is expected to be called Ada (Lovelace). For example, earlier today we wrote aboutthe number of sm’s and the cuda cores linked to them. Meanwhile, it looks like the RTX 4000 cards will also have a hefty amount of L2 cache on board.
For example, Twitter user XinoAssassin demonstrates with a screenshot of alleged Nvidia code that Ada GPUs offer 16 MB of L2 cache per 64-bit memory bus, compared to 512 kB per 32-bit in the case of Ampere. The memory buses of the various AD GPUs have also been leaked, allowing the amount of L2 cache to be calculated.
Ada Lovelace (‘RTX 4000)
GPU Cuda-cores Bus L2-Cache
AD102 18.432 384-bit 96 MB
AD103 10.752 256-bit 64 MB
AD104 7.680 192-bit 48 MB
AD106 4.608 128-bit 32 MB
AD107 3.072 128-bit 32 MB
Ampere (RTX 3000)
GPU Cuda-cores Bus L2-Cache
GA102 10.752 384-bit 6 MB
GA103 7.680 256-bit 4 MB
GA104 6.144 256-bit 4 MB
GA106 3.840 192-bit 3 MB
GA107 2.560 128-bit 2 MB
The main specifications of Ada and Ampere.
The top model, AD102, is said to have as much as 96 MB of L2 cache, which is 16 times more than the 6 MB of the Ampere flagship GA102. Also, the lower segmented AD106 and AD107 chips would offer significantly more L2 cache than Ampere at 32 MB. How Nvidia achieves this jump is not clear, possibility a technique like AMD’s Infinity Cache is used. Combined with the higher amount of cuda cores, Ada at least seems to offer a solid performance uplift compared to its predecessor.
source: Hardware.info
