Tesla self-driving supercomputer powered by NVIDIA A100 GPU

 Tesla self-driving supercomputer powered by NVIDIA A100 GPU

NVIDIA GPUs are being leveraged for Tesla's autopilot and autonomous driving capabilities.

Andrej Karpathy, senior director of AI at Tesla, at CVPR 2021, the world's largest computer vision conference, held online from June 19-25, U.S. It unveiled its own supercomputer used for training. The cluster utilizes 720 nodes of eight NVIDIA A100 Tensor Core GPUs (5,760 GPUs total) for industry-leading 1.8 exaflops of performance.

Nvidia A100 GPU

“This supercomputer is truly groundbreaking,” said Andrea Carpati, director. In fact, in terms of flops, that supercomputer is fifth in the world.”

With unprecedented levels of computing in the automotive industry, Tesla is helping autonomous vehicle engineers do their jobs more efficiently with cutting-edge technology.

NVIDIA A100 GPUs power data centers around the world, delivering industry-leading performance with acceleration at any scale. Based on the NVIDIA Ampere architecture, the A100 GPU delivers up to 20x performance improvement over the previous generation, and can be split into up to 7 independent GPU instances on demand.

The supercomputer is part of Tesla's vertically integrated approach to autonomous driving, with more than 1 million cars already on the road, constantly improving and building new capabilities.

Support from automobiles to data centers

Powered in ‘shadow mode,’ Tesla’s DNN silently detects and makes predictions while driving, even without actually controlling the vehicle. These predictions and any mistakes or misidentifications are recorded. Tesla engineers use these instances to create training datasets for complex and diverse scenarios to improve DNNs.

With over 1 million clips of approximately 10 seconds in length recorded at 36 frames per second, accumulating a massive amount of data totaling 1.5 petabytes (PB), the scenario is repeatedly run in the data center until the DNN works without errors. do. It is sent back to the vehicle and the process begins again.

“Training a DNN in this way and storing such a large amount of data requires a huge amount of computing,” said Andrea Carpati, director. So, Tesla is building a state-of-the-art supercomputer with a high-performance A100 GPU.”

Director Andrea Carpati introduces a supercomputer with NVIDIA GPU.

endless iteration

Tesla's supercomputers give autonomous vehicle engineers the power they need to experiment and iterate during development. “The DNN structure we are building now allows a team of 20 engineers to work simultaneously on a single network, and isolate different functions for parallel development,” explained Andrea Carpati, Director. This DNN is driven through the training dataset at a much faster rate than previously possible iterations.

“Computer vision is at the heart of what we do, enabling autopilot. This requires training a huge neural network and doing a lot of experimentation. That's why we're investing so heavily in computing.”