NVIDIA GeForce RTX 4000, large cache and memory bus information

After the rumors about the number of Streaming Multiprocessor of the GPU Ada Lovelace Of NVIDIAi.e. the alleged range GeForce RTX 4000 which should see the light in the last part of the year (September?), new information emerges on the amount of L2 cache of the chips and the memory bus.

The information comes from (via Videocardz) from the data stolen from NVIDIA in the now well-known attack of the past few weeks. As for the L2 cache, the new architecture should do a clear step forward compared to Amperewith the flagship GPU AD102 he should have 96MB L2 cache compared to the 6 MB of the current GA102 of the RTX 3090. Un 16-fold increasewhich is reminiscent of what AMD did with RDNA 2 GPUs and Infinity Cache (albeit with a different cache hierarchy).

I chip AD103, AD104, AD106 e AD107 they should instead boast respectively 64, 48, 32 and 32 MB of L2 cache. The current analogue GPUs GA103, GA104, GA105 and GA107 instead integrate 4, 4, 3 and 2 MB.

The AD102 GPU should have un bus di memoria a 384 bit, AD103 a 256 bit, AD104 a 192 bit e AD106 e AD107 a 128 bit. The increase in L2 cache comes alongside that of CUDA cores. AD102 should have a maximum of 144 Streaming multiprocessors, which means it could deliver 18432 CUDA cores, 71% more than the 10752 CUDA cores integrated in the GA102.

As for the other GPUs, AD103 would have a maximum of 84 SM (10752 CUDA core), AD104 would stop at 60 SM (7680 CUDA core), while AD106 and AD107 would integrate 36 and 24 SM, that is 4608 and 3072 CUDA cores respectively.

GeForce RTX 3000 (Ampere)					GeForce RTX 4000 (Available)
GPU	SM	L2 cache	Bus	CUDA Core	GPU	SM	L2 cache	Bus	CUDA Core
GA102	84	6 MB	384 bit	10752	AD102	144	96 MB	384 bit	18432
GA103	60	4 MB	256 bit	7680	AD103	84	64 MB	256 bit	10752
GA104	48	4 MB	256 bit	6144	AD104	60	48 MB	192 bit	7680
GA106	30	3 MB	192 bit	3840	AD106	36	32 MB	128 bit	4608
GA107	20	2 MB	128 bit	2560	AD107	24	32 MB	128 bit	3072

Read This Now: Nvidia won't release the RTX 2060 12GB founders and it's perfect for mining

The rumors about the higher consumption compared to the current generation, even if to be taken with a grain of salt regarding the rumors about the extreme values of 800-850W, acquire further foundation. For this reason, the new GPUs – which should be produced at 5nm by TSMC – will be powered by a 16-pin (12 + 4) cable capable of carrying up to 600W.