NVIDIA’s Ampere AMPeE GPU is about a month old, and it is now a monster-class GPU with the A100 core for the data center market, 54 billion transistors and 826mm2 core area. NVIDIA says it took them four years to develop the AMPE, and that the core has reached tsmcon7nm limit. In a blog post on NVIDIA’s official website today, Jonah Alben, co-head of GPU engineering, talks about a lot of inside information about Ampere’s GPU.
According to Jonah Alben, it took four years to develop the monster, which is 20 times the performance of the previous generation of GPUs during the development of ampere GPUs, and the actual performance improvement is not so exaggerated. )
The A100 GPU uses TSMC’s 7nm process, integrates 54 billion transistors and has a core area of 826mm2, which is almost the limit of the 7nm process, and Jonah Alben’s challenge is to integrate more functional units into such a core, otherwise the GPU architecture will be redeveloped.
In order to make full use of these conditions, NVIDIA has designed a number of redundant units on the A100 GPU, which is also the reason why the existing A100 GPU is not full of blood, the complete 128 sets of SMX, 8196 stream processors are currently only enabled 108 sets of SMX units, 6912 CUDA cores, 6 sets of HBM2 memory also enabled only 5 sets of HBM2, so the total bit capacity of 40 GB, the total bit is 150-10.
As for whether the future will unlock the full version of the core, NVIDIA is not a statement, but from the previous situation, the big probability will push another wave of Super series.