Nvidia released the world's first on-board AI supercomputer DRIVE PX 2 at the just CES event. But what is even more shocking is its parameter specifications, which is completely worthy of the name of supercomputer. It is a server running on a car:
Based on the 16nm FinFET process, the power consumption is up to 250W, and the water-cooled heat dissipation design is adopted. Support 12 channel camera input, laser positioning, radar and ultrasonic sensor;
CPU part: two new generation NVIDIA Tegra processors, including 8 A57 cores and 4 Denver cores (probably the first ARM architecture processor with A57 as a small core);
GPU part: NVIDIA's new generation GPU architecture Pascal (Pascal, which claims that its performance is 10 times that of the previous generation Maxwell architecture) was launched for the first time. Its single precision computing capability has reached 8TFlops, surpassing that of TITAN X, and has more than 10 times the deep learning computing capability of the latter.
Not to mention that the A9X on the iPad Pro has been blasted to dust. Even on the desktop level, it is also a blooming flower/monster. The desktop level TITAN X is also "only" 7TFLOPS, and Intel's core i7 is only 280GFLOPS. This performance is equivalent to the sum of 150 MacBook Pro.
The rest is industry and usage:
NVIDIA launched the first generation DRIVE PX last summer, which has been used by more than 50 automobile manufacturers, primary suppliers and research institutions to carry out automatic driving development. The technical R&D personnel of Audi, BMW, Mercedes Benz, Ford and other companies all gave high praise.
And Mr. Huang Renxun, the CEO of Nvidia, said that in order to make cars drive automatically, cars need to sense the surrounding environment through various sensors, generate accurate maps and precise positioning, make accurate predictions in an extremely short time, and also need to rely on in-depth learning. This is the purpose of an unprecedented processor.
NVIDIA will build an end-to-end in-depth learning training platform called NVIDIA DIGITS for autonomous vehicles. The platform is deployed through DRIVE PX 2, so that cars can better understand the whole world, and each car can build its own in-depth learning network through the network. The next level is Create a "deep neural network NVIDIA DRIVENET". After half a year of training, its object detection accuracy in KITTI dataset has reached 88%.
