For an event NVIDIA delegated Danny Shapiro, who is responsible for the development of the automotive systems segment, and he adhered to one interesting point during the speech, related to the simulation of tests of “robo-cars” using the DRIVE Sim platform. The latter, we recall, allows you to simulate in a virtual environment testing a car with active driver assistance systems in different lighting conditions, visibility and traffic intensity. NVIDIA representatives are convinced that the use of the simulator will accelerate the development of safe automatic vehicle control systems many times over.
In this process, it is not the number of miles that the prototype has covered that matters, Shapiro explains, but their quality. In this context, we mean the concentration of those conditions that make it possible to determine the behavior of the control system in critical situations. When automakers test conventional prototypes on public roads, they may not encounter critical situations for a long time, and therefore learning is slow. In addition, to search for some specific meteorological conditions, it is necessary to send testers to a remote area, where no one can guarantee the constant presence of the factors necessary for testing the algorithms: rain or snow will stop, the fog will dissipate, and the tests will have to be stopped. The simulator allows you to work out all this in a virtual environment.
NVIDIA is by no means going to replace real tests with virtual ones, they should complement each other. That is why the company uses for simulation the same set of equipment that is put in real prototypes of “robo-cars”, it’s just that their sensors and cameras receive not real data, but simulated ones.
Tesla remains a partner NVIDIA, but there are also contradictions
When it came to relations with Tesla, Mr. Shapiro emphasized that it remains a client and partner of NVIDIA, as it continues to use server components of the same name. At the same time, a number of Tesla statements regarding the performance of their own processor for accelerating neural networks NVIDIA continues to challenge. Representatives of Tesla, according to Shapiro, distort NVIDIA data, resorting to incorrect methods of comparison.
Based on the new in-house processor, Tesla's on-board computer delivers 144 trillion operations per second, while NVIDIA's DRIVE AGX platform at its highest configuration achieves at least 320 trillion operations per second, according to an NVIDIA spokesperson.
Disputes NVIDIA and Tesla's statements about the energy efficiency of their processor. All market players, according to Shapiro, obey the same laws of physics, and it cannot be that Tesla suddenly took and developed a processor that would be significantly more efficient in terms of speed and power consumption.
The introduction of "robomobiles": there is no need to rush
Danny Shapiro made a very important recognition for the entire industry. He said that early in the development of automated vehicle systems, market participants made many ambitious statements about the timing of fully autonomous vehicles on public roads. NVIDIA itself in the past also sinned with this, but as we delved deeper into the study of the problem, it became clear that the creation of such systems would take much more time than it seemed initially. NVIDIA does not want to bring something "raw" and insecure to the market, like many other companies involved in transport automation.
By the way, Shapiro stressed that NVIDIA itself is not going to produce “robomobiles”. Yes, she has several prototypes that travel on public roads in different regions of the planet, but these cars are only used to work out the algorithms in practice. Toyota, one of the largest automakers in the world, has begun cooperating with NVIDIA, and it will purchase not only components for on-board vehicle systems, but also server systems. In general, Shapiro believes that sales of server components for car control systems will become NVIDIA's main source of revenue in this area in the future. At least the profit margin here is higher than when selling components for final on-board units.
About competition with Intel and the Need for Acquisitions
For the sake of participating in the creation of components for a car "autopilot", Intel Corporation some time ago acquired the Israeli company Mobileye, which at first supplied Tesla electric vehicles with its components. When the partners parted ways, Israeli developers found shelter under the wing of Intel. NVIDIA assesses the competitive potential of Intel in the automotive sector as follows: the last of the companies has many disparate components (Mobileye cameras, Xeon server processors, Nervana neural network accelerators, Altera programmable matrices, and even a supposed discrete graphics processor), but NVIDIA itself can counter all this vertically integrated open ecosystem.
When Denny Shapiro was asked if she was considering a takeover of any sensor developer for "autopilot" systems (the same lidars, for example), he objected that such a deal would complicate fair interaction with all other optical radar developers. For this reason, NVIDIA prefers to maintain an equal relationship with all of them and will not buy anyone to form its own, more closed ecosystem.
About prices for "autopilot" options: from a few hundred to several thousand dollars
The NVIDIA representative at the RBC Capital Markets conference repeated the thesis already voiced by the company's executive director. "Autopilot" will add to the cost of cars from several hundred to several thousand dollars, depending on the level of autonomy of the system. The difference in price will be determined not only by a different set of components, since more “independent” cars will need more sensors, but also by the complexity of the algorithms. NVIDIA reminds that it now prioritizes the development of its software over hardware, and therefore more difficult to drive vehicles will require higher software costs.
But the cost of "automatic" options will not depend on the size of the cars, since both the truck and the compact passenger car will require a single set of components. Unless the sensors and cameras will be placed differently, but this will not have a decisive impact on the cost. By the way, NVIDIA is convinced that trunk transportation will become one of those areas where transport management automation will be implemented in the first place. In the end, logistics companies and their customers are interested in this, as this will reduce the transport costs for the delivery of all goods.
Source: 3dnews.ru