Nearly two months later than originally planned, NVIDIA GTC 2020 finally returned on May 65438, 2004.
However, this time, developers can't download the series online, so they can only watch the keynote speech of "Leather Master" Huang Renxun through live online. Lao Huang finished this unique "Kitchen Tone" at his home in Silicon Valley.
Although it was held in the kitchen, NVIDIA exploded a "nuclear bomb" and released a new generation of GPU architecture Ampere.
In the direction of autonomous driving, NVIDIA has achieved unprecedented success through the combination of two Orin SoC and two GPUs based on Ampere architecture? Up to 2000? Robotaxi computing platform, the overall power consumption is? 800 watts.
Some people in the industry believe that the computing power required for L2 autopilot is less than 10 TOPS, L3 needs 30-60 TOPS, L4 needs more than 1000 TOPS, and L5 needs at least 1000 TOPS.
Now NVIDIA's autopilot computing platform has gone from? 10TOPS/5W，200TOPS/45W？ Arrive? 2000 tops /800W? The complete product line of, corresponding to the foresight module, L2+ADAS? And then what? Robotaxi？ All levels of applications.
In terms of product line, NVIDIA? Driving AGX? Will it fully benchmark Mobileye? EyeQ？ Series, hoping to become a key manufacturer of mass production supply chain.
1, A New GPU Architecture: Ampere
It's worth waiting for two months. At this GTC, Huang Renxun released NVIDIA's next-generation GPU architecture Ampere (Ampere) and the first GPU NVIDIA a 100 based on this architecture.
The overall performance of A 100 is 20 times higher than that of the previous generation products based on Volta architecture. This GPU will be mainly used for data analysis, professional computing and graphics processing.
Before Ampere architecture, NVIDIA had developed several generations of GPU architectures, all named after great men in the history of scientific development.
Such as Tesla, Fermi, Kepler, Maxwell, Pascal, Volta and Turing.
The upgrading of these core architectures is the key to improve the overall performance of various GPU products in NVIDIA.
For the first GPU A 100 based on Ampere architecture, Huang Renxun introduced its five core features in detail:
It integrates more than 54 billion transistors and is the largest 7-nanometer processor in the world. The tensor kernel of the third generation tensor operation instruction is introduced, which is more flexible, fast and easy to use. Using structured sparse acceleration technology, the performance is greatly improved; Support a single A 100 GPU to be divided into as many as seven independent GPUs, and each GPU has its own resources to provide different computing power for different scales of work; The third generation NVLink technology is integrated, which doubles the speed of high-speed connection between GPUs, and multiple A 100 can form a giant GPU with scalable performance.
These advantages add up, and finally the training performance of A 100 is improved compared with the previous generation GPU based on Volta architecture? 6 times, reasoning performance improved? Seven times.
Most importantly, A 100 can now be supplied to users, which is produced by TSMC's 7nm process.
Alibaba Cloud, Baidu Cloud, Tencent Cloud and other domestic enterprises are planning to provide services based on A 100 GPU.
2.Orin+ amp architecture GPU: to achieve 2000TOPS computing power.
With the introduction of NVIDIA's new GPU architecture Ampere, NVIDIA's NVIDIA Drive has also ushered in a performance leap.
As we all know, NVIDIA has launched several generations of Drive AGX autopilot platforms and SoCs before, including? Driving AGX Xavier and AGX Pegasus? And then what? Driving AGX Orin.
Among them, the Drive AGX Xavier platform includes two Xavier SoC, with a computing power of 30TOPS and a power consumption of 30W.
Tucki P7, which was launched recently, is equipped with this mass-produced computing platform, which is used to realize a series of L2-level automatic driving assistance functions.
Drive AGX Pegasus platform includes two Xavier SoC and two GPUs based on Turing architecture, with a computing power of 320TOPS and a power consumption of 500W W W.
At present, autonomous driving companies such as Wen Yuan Star are using this computing platform.
At the GTC China conference in February, 20 19, NVIDIA released the latest generation of self-driving computing SoC Orin.
This chip consists of 65.438+0.7 billion transistors, which integrates NVIDIA's new generation GPU architecture, Arm Hercules CPU core and brand-new deep learning and computer vision accelerator, and can run up to 200 trillion calculations per second.
Compared with the previous generation Xavier, the performance has been improved by 7 times.
Today, NVIDIA has further improved the computing power of the self-driving computing platform, and achieved an amazing 2000TOPS computing power by integrating two Orin SoC and two GPUs based on Ampere architecture.
Compared with Drive AGX Pegasus, its performance is improved by more than 6 times, and correspondingly, its power consumption is 800W W. ..
According to the computing power of Orin SoC 200TOPS, the computing power of GPU based on Ampere architecture reaches 800TOPS.
Because of its high computing power, this platform can handle the higher resolution sensor input and more advanced autopilot depth neural network required for fully automatic taxiing operation.
For the development of advanced autonomous driving technology, NVIDIA relies on Orin SoC and Ampere GPU architecture to lead the whole industry in computing platform.
Of course, as a software-defined platform, NVIDIA Drive AGX has good scalability.
Especially with the introduction of Ampere GPU architecture, the platform has been able to achieve all-round coverage from the entry-level ADAS solution to the L5-class self-driving taxi system.
For example, NVIDIA's Orin processor series has a low-priced product that can provide the computing power of 10TOPS, and the power consumption is only 5W, so it can be used as the computing platform of vehicle-mounted forward-looking ADAS.
In other words, developers using NVIDIA Drive AGX platform can develop autonomous driving systems adapted to different market segments based on only one architecture on a single platform, which saves the high cost of independently developing multiple subsystems (ADAS, L2+ and other systems).
However, manufacturers who want to adopt the Orin processor will have to wait for a while, because this chip will provide samples of 202 1 until? It will not be put into production and started to supply until the second half of 2022.
3. NVIDIA's self-driving "circle of friends" has expanded again.
At this GTC, NVIDIA's self-driving "circle of friends" continued to expand.
Pony.ai, a domestic autonomous driving company, and an American electric vehicle startup company? Cano? Faraday will join NVIDIA's autonomous driving ecosystem in the future, adopting NVIDIA's Drive AGX computing platform and corresponding supporting software.
Ma Xiao Zhixing will build a new generation Robotaxi model based on Drive AGX Pegasus computing platform.
Previously, Ma Xiao Zhixing had already received a $400 million investment from Toyota. I don't know if its new generation Robotaxi will be built based on Toyota models.
Canoo, an American electric vehicle startup, has launched an electric mini-bus specially designed for * * * to enjoy the travel service, and plans to put it into production in the second half of 20021year.
In order to realize the series of driver assistance functions, this model will be equipped with NVIDIA Drive AGX Xavier computing platform. Not long ago, Canoo also reached a cooperation with Hyundai Motor to jointly develop an electric vehicle platform.
Faraday Future, as a special existence in the global new car manufacturing circle, has also joined NVIDIA's autonomous driving ecosystem this time.
The autopilot system on FF's first production car FF9 1 will be built on the computing platform of Drive AGX Xavier, and the whole car is equipped with as many as 36 sensors.
Faraday's future official said that FF9 1 is expected to start delivery at the end of this year. I don't know if there will be any more ticket skipping.
NVIDIA, as the absolute overlord in the field of GPU, has built a complete software-defined autopilot platform for data acquisition, model training, simulation testing, remote control and real vehicle application with the blessing of high computing power data center GPU and high-performance extensible autopilot computing platform, and realized a complete end-to-end closed loop.
At the same time, its autonomous driving ecosystem is also expanding. Hundreds of enterprises in the autonomous driving industry chain, including automobile manufacturers, tier-one suppliers, sensor suppliers, Robotaxi R&D companies and software start-ups, have been developing, testing and applying autonomous vehicles based on NVIDIA computing hardware and supporting software.
In the future, in the entire autonomous driving industry, with computing chips as the core advantage, NVIDIA will have a deeper reach and have the opportunity to become an indispensable supplier in the industrial chain.
This article comes from car home, the author of the car manufacturer, and does not represent car home's position.