Recently, Horizon, a leading provider of intelligent driving computing solutions in China, held a technology open day in Shenzhen, its new product Horizon Journey 5. As the leading smart chip developer in the current market share, what are the advantages of Journey 5 chip in the technical roadmap? What is the market application prospect of Horizon Journey 5? What is the future trend of intelligent driving development? As a next-generation intelligent terminal, Dr. Huang Chang, co-founder and CTO of Horizon, and Dr. Yu Yinan, vice president of Horizon, made wonderful introductions and interpretations on these issues.
Hashrate or algorithm? Computing is king
The Journey 5 chip is based on Horizon's latest generation BPU Bayesian architecture, which is synergistic with deep neural networks and is a data-driven deep learning acceleration engine. The biggest feature of Journey 5 is that it can achieve high effective computing performance under the condition of limited physical computing power (128TOPS), and the image rate can reach up to 1531FPS. In this indicator, it is a leading position in the industry.
How to understand this? To give a perhaps inappropriate example, if we imagine the physical computing power of a smart chip as the maximum power of an engine, and the algorithm as its transmission system, the user's real perceived computing effect is equivalent to the actual performance of the vehicle equipped with this powertrain.
An optimized algorithm equivalent to a more efficient gearbox, such as Porsche's PDK, maximizes the power generated by the engine, which in turn translates into faster acceleration. The advantage of only physical computing power does not have a better algorithm, just like a high-performance engine matched with an automatic transmission with lower transmission efficiency, which cannot fully release the power of the engine, and the acceleration time is even slower than that of a lower-power engine.
In other words, Horizon's algorithmic advantage is actually to maximize the full physical potential of the chip and use every computing power more efficiently. Therefore, Horizon's tendency in smart chips is very clear: simply stacking computing power can not completely solve the problem, algorithm optimization can improve the upper limit of user experience, and lean computing efficiency is the optimal solution, especially in the context of slowing down semiconductor development, the gap in computing performance of algorithms is far larger than the gap in computing power between chips.
This can actually be verified in other car companies, such as Tesla's HW3.0 using a self-developed FSD chip with a total computing power of 144TOPS, which is less than 7 times higher than the 21TOPS theoretical maximum computing power of HW2.5 NVIDIA Drive chip, but the image processing power is increased by 21 times. In addition, at present, the domestic TOP2 level intelligent driving assistance system Xpeng NGP, their P5 model only uses a 30TOPS computing power NVIDIA Xavier chip to achieve quite eye-catching urban NGP functions.
Heavy perception and light maps, injecting "Common Sense" into machines
Emphasizing perception and light maps is the new direction of the domestic intelligent driving assistance industry. As the number of vehicle intelligent sensing hardware continues to reach new highs and become more and more diverse, the amount of data to be processed by the chip is greatly "expanded", and the data types are becoming more and more diverse. The high-definition maps required for urban area/high-speed NOA are expensive in terms of coverage, accuracy, and data freshness. Instead of using maps as the main input source, giving vehicles more autonomous cognitive capabilities for perceptual data is the route in line with "intelligent" development, and high-definition maps can gradually "take a back seat" and only serve as an input to assist the entire system.
Therefore, another feature of Horizon products represented by the Journey 5 chip is to support the industry's advanced bird's-eye perception-based autonomous driving algorithm (BEV) and its required Transformer model, which can convert the perception data into different forms of expression or recompile it into a unified algorithm language, in order to meet the cognitive needs of the autonomous driving algorithm for the collected data, so that the intelligent car can better "understand" the collected perception information, and have a more "common-sense" cognition of the real world. This is significant when realizing autonomous driving capabilities in complex road conditions.
Horizon Journey 5 can meet up to 16 camera perception with a single chip, and openly supports multiple sensor perception fusion such as lidar/millimeter-wave radar. Horizon's internal expectations for the Journey 5 chip are to implement the urban NOA function through two chips, and one of their partners, Qingzhou Smart Navigation, has released the urban NOA function based on a single Journey 5 chip on 3.27, which fully demonstrates the great potential of this chip with the support of high-order algorithms.
With the rapid popularization of intelligent terminals with high computing power, how does the relationship between people and cars change?
Yu Kai, founder of Horizon Company, gave his prediction at the 2023 Electric Vehicle 100 Conference: L2+ high-speed NOA will still be mainstream in 2025, and the single Horizon Journey 5 chip is enough to meet the demand. The "endgame" of intelligent driving in the ten-year latitude may be the continuous optimization of L2+++, the relationship between people and intelligent cars has gradually evolved into the relationship between people and horses, and the car is no longer just a completely dominated tool, but has a certain autonomy, and the mode of getting along between people and cars will also change.
Dr. Yu Yinan also said that the rapid dilution of the cost of intelligent driving hardware will greatly accelerate the popularity of intelligent cars, and the cost of today's smallest domain control module may be able to buy high-end intelligent driving chips with hundreds of TOPS level computing power at the end of 2024, and the threshold for functions such as L2 intelligent driving assistance and high-speed NOA will be lower and lower.
Cars will eventually become the first commercial, large-scale, and applicable smart terminal for individual consumers in human history. Finally, the accumulation of R&D experience and achievements in the field of intelligent driving will be transferred to other fields at a very low cost, accelerating the intelligent process of the entire society.
However, compared with the developer's vision of an ideal future, the research and development of intelligent driving technology still faces many challenges at the current stage. At present, the high-end intelligent driving assistance is suitable for those "old drivers" with rich driving experience. Because such drivers know how to judge whether the vehicle decision is reasonable according to the road conditions, they also need to pay attention to the driving state at all times in the case of complex road conditions, and be ready to take over or correct at any time.
If high-end intelligent driving wants to achieve further development, reach mass popularity, and have a high degree of user usage, then it must reach a level that exceeds the average driving ability of novice drivers in order to cover the bottom for any driver. Dr. Yu Yinan jokingly put forward a point of view at the scene that the future training "driving school" for intelligent driving assistance may become a hot spot and business opportunity, because intelligent driving assistance systems do need people to understand and use more correctly, understand the decision-making logic behind it, so as to make appropriate responses in different situations.
summary
My view on this is that when high-end intelligent driving assistance systems are really popularized, today's human "driving" ability may become like "can drive manual transmission", which sounds particularly difficult, and the skill that can "show off", driving a vehicle is no longer a necessary ability, but a supplementary skill. The mode of transportation in human society will undergo profound changes, and the threshold for "mobility" will become extremely low for most people, and the participation of ordinary people in it will also decline indefinitely.
As a normal person, I am excited and can't wait for this sci-fi scenario to come true, and as a driving enthusiast, I hope that day will come later