Detailed chips information in NVIDIA RTX 40

The Kopite7Kimi insider, which is known for many leaks that are true, published a block diagram for the NVIDIA AD102 graphic processor, which will be used in RTX 4090 and probably RTX 4080.

AD102 will consist of 12 graphics processing clusters (GPC), which is 70% more than in GA102. Each GPC consists of 6 TPC and 2 streaming multiprocessors (SM), and this is exactly the same configuration as the GA102. What has changed is the configuration of the FP32 and Int2 of units inside, which will now be 128+64. Each streaming multiprocessor includes 512 FP32 and 256 Int32 units. Thus, the AD102 chip, which consists of 24 SM, will receive 12288 FP32 and 6144 Int32, which is 18432 nuclei.

AD102 will receive 4.5 MB of the L1 cache, which is 50% more than that of the GA102, as well as 96 MB of the L2 cache (16 times more than that of GA102). And the chip has tensor nuclei of the fourth generation and the kernel of trace of third -generation rays.