NVIDIA’s new generation of Ada GPU core map exposure: RTX 4090 doubles performance without pressure-NVIDIA, Ada Lovelace, RTX 4090

At 11:00 a.m. on May 24th, NVIDIA will hold a keynote speech at the Taipei Computer Show. Although Lao Huang is absent, the senior vice president of GeForce business is on the list, and there is still hope to announce the news of the RTX 40 series graphics card or even release it in advance.

Before that, the whistleblower Kopte7kimi shared the core design drawing of the so-called AD102 GPU.

AD102 is also the sub-top core of the Ada Lovelace family, and it is also the configuration of the flagship RTX 40 series game card. It is likely to correspond to RTX 4090 Ti and RTX 4090 graphics cards.

According to the analysis, AD102 has built-in 12 sets of GPC (display computing cluster), which is 70% more than the previous generation GA102. Each set of GPCs consists of 6 TPCs (2 SMs) and each SM unit consists of 4 sub-cores, which are all the same as Ampere, but the difference is that each SM sub-core contains 128 sets of FP32 cells, plus IN32 integer cells total reached 192.

The complete AD102 includes 24 groups of SMs, all of which are 12288 FP32 units plus 6144 INT32, which is 18432 CUDAs in an easy-to-understand way.

In terms of cache, in the AD102 core, each group of SMs enjoys 192KB L1, a 50% increase over Ampere, totaling 4.5MB. L2 increased to 96MB, 16 times the ampere.

Correspondingly, the scale of ROP and RT light chasing units is naturally rising. The AD102 has a maximum of 384 ROPs, and the RTX 3090 Ti is only 112. In addition, the light tracking unit was upgraded to the third generation, and the Tensor unit was upgraded to the fourth generation.

Based on this, RTX 4090’s final performance doubling does not seem to be unrealistic. As far as FP32 single-precision floating-point is concerned, the outside world is expected to reach 90T, while RTX 3090 Ti is only 40T, at the cost of over 600W of power consumption…

Hashtag: RTX 4090NVIDIAAda Lovelace