At the scene, Huang Renxun mentioned that the Chinese market accounted for 1/3 of Nvidia's total revenue, and Nvidia had 3,000 employees in China.
Huang Renxun told Zhizhi that the technology announced yesterday to cooperate with ARM to build AI chips originated from DLA, DLA (Deep Learning Accelerator, deep learning accelerator) is essentially an ASIC chip built by Nvidia. Now many AI chips, such as Google’s TPU, essentially an ASIC hardware.
In addition, Huang Renxun confirmed in yesterday’s interview that Nvidia had suspended the test of driverless cars on public roads and explained the reasons.
By the way, today is exactly the day of the Adobe Summit. Huang Renxun also participated in an interview at the Adobe Summit in Las Vegas this morning. He talked about the Nvidia ray tracing technology, image processing technology, and so on. He returned to San Jose GTC this afternoon. on site.
Hand ARM, help chip manufacturers to create dedicated AI chip
Now that many Chinese AI chip startup companies are emerging, will it pose a threat to Nvidia?
Huang Renxun said that AI is the future of software. In all industries in the future, all software will benefit from AI, cloud, medical care, manufacturing, etc. The future of AI is very large. Its volume is much larger than that of a company. Nvidia cooperates with many AI startup companies. It will even work with AI chip startups, such as integrating DLA's TPU-like hardware into the ARM framework, allowing chip companies to build their own AI chips.
In yesterday's speech, Nvidia announced cooperation with chip giant ARM to build the IOT AI SOC, integrating NVIDIA's AI technology into ARM's Project Trillium framework launched in February this year, making it easier for chip manufacturers to build their own. AI chips, including mobile phone chips, consumer electronics chips, IoT chips, etc. This technology originates from the DLA Deep Learning Accelerator project open sourced by GTC on the GTC last year.
At the scene, Huang Renxun told Zhizhi that DLA (Deep Learning Accelerator) is essentially an ASIC chip built by Nvidia. Now many AI chips, such as Google's TPU, are essentially ASIC hardware.
Huang Renxun said that after this cooperation with ARM, after integrating DLA into ARM's IP framework, manufacturers can download AI architectures with different configurations from small to large, and build their own AI chips. However, this chip needs to run TensorRT software.
NVIDIA DLA vs Huawei Unicorn 970, Apple A11
Deepin Talla, Global Vice President of Nvidia, is responsible for the business of Robotics, Smart City, and DLA Deep Learning Accelerator. This project was the technical foundation for yesterday's cooperation with chip giant ARM to create a deep learning accelerator IP for smart chips.
Last year, Nvidia chose the open source DLA project, which allows major chip manufacturers to download the chip accelerator program for free and build their own low-power AI chips. However, there are still certain technical thresholds for using open source projects. At this GTC, Nvidia announced that it will once again lower the threshold for AI chip creation and cooperate with ARM.
Deepu Talla tells Chi Chih that DLA is an open source project and this cooperation will not charge ARM. The DLA project has been developed for three years. The reason why it chooses free and open source is that projects like the Internet of Things are not the focus of Nvidia's attention. .
In the recent half year, many smart dedicated AI chips have emerged. Among them, the two most closely watched are two mobile phone chips: Huawei Unicorn 970, Apple's A11. Deepu Talla spoke at the scene, and now there is no Nvidia Deep Learning Accelerator IP. Compared with the Unicorn 970 NPU, Apple A11 neural network engine parameters. However, after three years of DLA R & D, Nvidia found that creating a deep learning accelerator hardware is actually the simplest part of the entire project, the hard part is the software, how to Good support for various AI software, AI frameworks, how to better deploy neural networks, etc.
In these respects, many software engineers from Nvidia have invested many years of research, especially in neural network applications (inference) and deployment. In this GTC, Huang Renxun also introduced a new version of TensorRT 4.0, letting neural networks Deployment is easier and faster.
Four major reasons why GPUs are in short supply
At the scene, Huang Renxun also stated that the current demand for GPUs has exceeded the supply of Nvidia, for four reasons:
On the one hand, the game industry is developing rapidly, and several heavy game products have been released last year. The demand for GPUs has increased;
On the other hand, content creation is now in progress, and the content sharing market is also developing, especially content-based content creation;
The third aspect needless to say is artificial intelligence. With the development of artificial intelligence, Infineon’s GPU makes artificial intelligence even more popular, no matter who you are.
The last aspect is one of the hottest areas at present: Blockchain. Since billions of GPUs in the world use the same architecture, it can be regarded as the world's largest and most decentralized database.
Suspend testing of driverless cars on public roads
Not long ago, the case of pedestrian deaths caused by road tests on Uber's driverless vehicles affected the global auto driving industry. During the GTC, Nvidia formally confirmed that it will suspend testing of driverless cars on public roads.
Huang Renxun first said that Uber did not use Nvidia’s autonomous driving technology. In the previous interview, Huang Renxun also mentioned that this accident did cause people to feel sad, and the entire industry was also involved. Nvidia is also very concerned about the safety of everyone. , including Nvidia's own staff, because they also do various testing tasks in the test car.
Therefore, Nvidia decided to suspend the test of driverless cars on public roads. From this incident, they learned, studied, studied, and learned experiences to make their work safer. However, Huang Renxun also said that this suspension should not be too serious. Long.
In yesterday’s keynote speech, Huang Renxun also launched a 3D autopilot simulation test platform named DRIVE SIM and Constellation, which first generates sensor data (including cameras, radar, etc.) in the cloud, and then transmits these data to DRIVE Pegasus. In order to help train autopilot systems, it is equivalent to testing driverless cars in a virtual world, which is more secure.