The GTC conference on March 23 may be the window time for NVIDIA to officially announce the new generation of GPU architecture. This week, a newcomer named KittyCorgi released a lot of Ampere’s core technical information. Although the authenticity has yet to be proven, you may wish to take a look.
According to this revelation, the core area of Ampere GPU is as high as 826mm2. If it is built based on 7nm, the transistor scale will be quite terrifying.
Specifically, the “GA103” core corresponding to the RTX 3080 Ti is divided into 6 groups of GPC units, each group of GPC has 10 SM units, a total of 60 SM units, 3840 CUDA (if one SM unit is increased to 128 CUDA, That is 7680, based on the core area, the latter is more likely), the video memory bit width is 320bit. Previously, the TU102 of the RTX 2080 Ti had 72 SM units, 68 effective, for a total of 4608 CUDAs.
The “GA104” core corresponding to the RTX 3080 is also 6 groups of GPCs, each group of GPC is reduced to 8 SM units, a total of 48 SM, 3072 or 6144 CUDA, 256bit video memory width, all support PCIe 4.0.
Back to the SM unit, the INT32 unit remains unchanged, the FP32 scale doubled, Tensor Core doubled, L1 increased, and the optical tracking unit upgraded to RT Core Advanced.
Looking forward to Lao Huang’s announcement of Ampere and Hopper as soon as possible…
The Links: BSM50GD120DN2 FF600R06ME3