Tesla's self-developed 7nm super chip can replace NVIDIA Ampere GPU

Tesla’s self-developed 7nm super chip can replace NVIDIA Ampere GPU

Tesla’s self-developed 7nm super chip can replace NVIDIA Ampere GPU

Tesla’s self-driving technology development is at the forefront, and there is no doubt that there is a large-scale supercomputer behind it as a development training platform.

A few days ago, Tim Zaman, general manager of Tesla engineering, shared the latest specifications of Tesla’s artificial intelligence training system.

A year ago, it was equipped with 5,760 NVIDIA A100 GPUs, and now it has increased by 1,600 to 7,360, an increase of 28%.

A100 is also a complete Ampere core, integrated 80GB HBM2e video memory, has 2TB/s bandwidth, and consumes up to 400W.

According to Zaman, Tesla’s supercomputer is the 7th largest server computing platform in the world in terms of GPU size.

Of course, Musk is not satisfied with buying chips from NVIDIA. In fact, Tesla is developing the Dojo supercomputing, and the GPU chip is also a self-developed D1 and TSMC 7nm process.

D1 is a semi-custom ASIC, which serves artificial intelligence inference operation training.

The single-precision floating-point performance is as high as 22.6T, integrating 50 billion transistors, and the power consumption is also as high as 400W.