Since Nvidia's internal network has been hacked, more details are emerging about the Ada Lovelace 3D cards - the ones that are replacing the current Ampere. Yesterday we wrote about the number of streaming multiprocessors and, accordingly, CUDA cores in new GPUs, and now it is reported that the GeForce RTX 40 will receive simply gigantic amounts of cache memory.
If the Ampere GPU has 512 KB for 32 bits of the bus, then Ada Lovelace will receive 16 MB for 64 bits. Assuming that the memory bus width in the new models will match the memory bus width of the current models, it turns out that the GeForce RTX 4090 on the AD102 GPU will have 96 MB of L2 cache, the mobile GeForce RTX 4080 Ti on the AD103 will have 64 MB, and the same number - for the GeForce RTX 4070 on AD104. The GeForce RTX 3060 on the AD106 will get 48MB of cache, while the AD107-based graphics cards will have 32MB.
Specifications GeForce RTX 40 and GeForce RTX 30 Number of SMs
Number of kernels of the CudAshire of the tire of the memory tire, Bit bunch of Cash memory L2, MBgeforce RTX 40 (Ada Lovelace) AD102 144 18432 384 96ad103 84 10752 256 64ad104 60 7680 256 64ad106 36 4608 192 48ad107 24 3072 128 32GeForce RTX 30 (ampere) GA102 84 10752 384 6GA103 60 7680 256 4GA104 48 6144 256 4GA106 30 3840 192 3GA107 20 2560 128 All technical details about the GeForce RTX 40 based on the latest leaks are summarized in the table. This makes it easier to get an idea of the capabilities of the GeForce RTX 40 in comparison with the GeForce RTX 30.
bbabo.Net