GeForce RTX 40 “Ada” GPUs to feature very large L2 Caches, NVIDIA’s own “Infinity Cache”?

NVIDIA Leaks: GeForce RTX 40 to have large L2 cache

While the company is still trying to figure out how the hackers gained so much information, the data that was already posted provides significant information of NVIDIA’s next architecture codenamed “Ada”. 

NVIDIA RTX 40 GPUs will be equipped with up to 96 MB of L2 Cache, up to 90 MBs more than Ampere RTX GPUs. This is the latest information that has been discovered in the files that were leaked last week. The data supposedly confirms that “Ada” GPUs would have 16MB of cache per 64-bit memory bus, significantly more than Ampere with 512 KB per 32-bit memory.

According to this information, the flagship Ada GPU codenamed AD102 would have 96 MB of cache, AD103 and AD104 would be getting 64 MB each, while AD106 and AD107 would be equipped with 48 MB and 32 MB respectively.

New leaked information, combined with insights from reliable NVIDIA leaker Kopite7kimi, have been detailed enough for this proposed GPU block diagram to be confirmed for the upcoming AD102 GPU. It is said that this chip would have a die size of 600 mm², about just as big as GA102.

NVIDIA AD102 GPU, Source: @ftiwvoe

The RTX 40 GPUs will be getting up to 71% more cores than RTX 30 GPUs, but that’s assuming each chip will see its full implementation released to the market. Naturally, there are many still many things that are not clear such as the clock speeds and power requirements for new cards, but it is quite obvious that RTX 40 GPUs will be a significant upgrade over Ampere. The same architectural leap is expected from AMD RDNA3 GPUs.

The information on large L2 cache has been first reported by XinoAssassin:

Source: @harukaze5719, @kopite7kimi, @xinoassasin1