NVIDIA Tesla A100, this is the design of the new solution for AI and Deep Learning

Robert Sole 14 May 2020

1 1.083 1 minute of reading

Yesterday a video was published in which Jen-Hsun Huang, CEO of NVIDIA, took out of the oven a new HPC platform based on Ampere. This new solution is specially designed for AI, Deep Learning, Machine Learning and others. Now we have learned something more about the silicon that is integrated into this platform. Just a few hours before the GTC 2020 starts.

Specifically, a rendered image of the NVIDIA Tesla A100 has been leaked, which is accompanied by 6 stack of HBM2 memory. This Tesla A100 solution has been developed for the industrial and business sector. This solution is based on the Ampere architecture, which promises a brutal leap from the current Tesla architecture.

[amazon box="B07JBQ4DBV"]

First image of the NVIDIA Tesla A100

These NVIDIA solutions are typically used in data centers, high-performance computing systems, robotics, AI, and a host of specialized fields where high computing power is required. Like the solutions based on Volta, the solutions based on Ampere will be a reference in these sectors.

The Tesla A100 unit is based on the Ampere architecture GA100 silicon with 6 stacks of HBM2 memory. We see that it has been redesigned, since the mounting holes do not match those of the Tesla V100. Also, the DIE appears to be slightly larger than that of the GV100, with that of the GA100 being around 820-840mm.².

Something very interesting that has been revealed is that the GA100 GPU will have 54.000 million transistors and has been manufactured in 7nm. What we do not know is whether this lithograph is TSMC's or Samsung's, something that has not been indicated at the moment. We also do not know if the memories are HBM2 or HBM2E and the size of each of the stacks.