According to recent rumors, the next generation of Nvidia's Ada GPU architecture will receive a huge increase in the size of the L2 cache. Leakers @harukaze5719 and @xinoassassin1 on Twitter have revealed this information unconfirmed. They claim that the L2 cache sizes of these graphics cards have increased on the new architecture by about 16 times compared to current Nvidia GeForce RTX 30-series GPUs, and are on par with Nvidia's Radeon RX 6000-series GPUs. AMD with its technology infinity cache.
The flagship die of the Ada Lovelace, the AD102, it would supposedly receive the largest cache size of the entire range at an impressive 96MB. The AD103 and AD104 models could receive 64MB of cache instead, the AD106 will have 48MB and the AD107 will come with the least amount of cache at 32MB.
The Ada Lovelaces would be brown beasts of L2 cache
Even the AD107's 48MB cache size would outstrip Nvidia's most powerful GPUs. The GA102 comes with "only» 6 MB L2 cache. Previous reports indicated that the Ada Lovelace could have up to 71% more cores than the Ampere series. The more cores, the more memory bandwidth is needed to constantly feed each core with data. Rather than moving to a wider memory interface, improving bandwidth utilization with a larger cache is a good option.
This would also be a response to AMD, which introduced large internal caches on GPUs with the Radeon RX 6000 series based on RNDA2. With RDNA2, AMD introduced its Infinity cache, which added a huge L3 cache to the GPU along with the other L0, L1, and L2 caches. The top RDNA2 array has 128 MB of Infinity Cache.
In tests with GPUs like the RX 6800 XT and RX 6900 XT, it was found that these huge cache sizes could completely alleviate the GPU's 256-bit memory bus bottleneck. That bus pales in comparison to that of Nvidia's RTX 3080, which has a 320-bit one, and the 3090's has a 384-bit one. coupled with Nvidia's RTX 3080 and 3090 using faster GDDR6X memory, while AMD has stuck to regular GDDR6 ICs.
Cache a boon for modern GPUs
Despite the apparently less powerful memory configuration of the RX 6800 XT and RX 6900 XT, have been able to stand up to their Nvidia counterparts thanks to the large reserve of Infinity Cache. this will be what would have made NVIDIA want to dedicate so much prominence to them.
NVIDIA's Ada Lovelace will be the first GPUs to follow an AMD novelty. According to the leak, Nvidia will not only use large L2 caches, but also keep the same bus width as the current generation of Ampere GPUs. It is seen that Nvidia will rely more on internal caches to maintain high memory bandwidth, just like AMD.
Source: Tom's Hardware
