Rumor: Nvidia’s RTX 40 GPUs get much more L2 cache than RTX 30 GPUs

Nvidia’s upcoming RTX 40 GPUs will have significantly more L2 cache than the current RTX 30 GPUs. At least, that seems to emerge from data allegedly from the recent data breach that Nvidia had to deal with.

Twitter user XinoAssassin shows a screenshot of data allegedly from the data breach. This would indicate that the successors of the current Ampere GPUs from Nvidia will have 16MB L2 cache per 64bit memory bus. The website Videocardz suggests that that would be considerably more than the amount of L2 cache that the Ampere GPUs have, namely 512KB per 32bit.

The new RTX 40 GPUs have been referred to in the rumor circuit for some time with the name Ada, which refers to Ada Lovelace. That name also came up in the data from the recent data breach, but has not yet been officially confirmed. The fastest Ada GPU, codenamed AD102, would have a 384bit memory bus, meaning 96MB of L2 cache is present. The GA102 GPU used for the RTX 3090 and 3080 Ti also has a 384bit memory bus, but has to make do with 6MB L2 cache. The GA102 GPU is also the basis of the RTX 3080, although this video card has a 320bit memory bus and 5MB L2 cache.

With this move, Nvidia appears to be broadly following AMD’s lead. With the arrival of the Navi 21 GPU for AMD Radeon RX 6000 video cards, AMD introduced Infinity Cache, a cache memory that is intended to relieve the memory bus and increase the performance per watt. The company applied 128MB Infinity Cache to the chip, in combination with a 256bit memory bus. As a result, AMD was able to deliver more than twice the bandwidth of a 384-bit memory bus, with the advantage of a narrow and cheaper memory bus and lower power consumption. Nvidia will want to achieve the same by expanding the L2 cache to 96MB. A small side note here is that with Infinity Cache, AMD actually added L3 cache and thus expanded the cache memory in that way.