Nvidia’s unveiled its latest catchily-named supercomputer, the DGX SuperPOD, claiming the robot-overlord will play a key roll coordinating and training its self driving car fleet.
DGX SuperPOD was unveiled today and according to Nvidia is the 22nd fastest supercomputer on the planet, featuring 1536 Nvidia V100 Tensor Core GPUs and offering 9.4 petaflops of processing capability – to non-techies, trust us, that’s a lot.
The company plans to use the new supercomputer to help develop and “train” the AI (artificial intelligence) infrastructure needed to manage and run a fleet of self driving cars.
It’ll apparently do this by collecting roughly 1 terabyte of data per hour from the test self-driving cars it manages and using it to create “training” algorithms on the rules of the road for them. This will reportedly help ensure the cars are “safe” and continuously improving, just like a real world driver.
Clement Farabet, vice president of AI infrastructure at Nvidia, said the high power demands of the training scheme forced the company to create a custom supercomputer to handle it.
“AI leadership demands leadership in compute infrastructure. Few AI challenges are as demanding as training autonomous vehicles, which requires retraining neural networks tens of thousands of times to meet extreme accuracy needs. There’s no substitute for massive processing capability like that of the DGX SuperPOD,” he said.
Nvidia claims it was able to train a ResNET-50 “neural network” running the cars in less than two minutes using the new DGX SuperPOD. The process reportedly took 25 days in 2015.
The tech is available to customers now, though Nvidia hasn’t confirmed which car manufacturers have signed up to use it.
Nvidia is one of many companies experimenting with self driving car tech. Google’s been working on a fleet of self-driving cars for years, as has Uber. Apple is also rumoured to be working on a fleet of autonomous automobiles, though the company has never formally admitted this.