A few days ago, at the Horizon Journey 5 Technology Open Day, Horizon co-founder &; CTO Huang Chang said, "If L2-level assisted driving can be within 10TOPS, L2+ needs dozens of TOPS, L3 needs more than 1000TOPS, and L4 needs 1000 TOPS. To move towards true automatic driving and driverless driving, L4 or even L5, the computing power is at least supported by the effective computing power of several thousand top. "
Then, let's talk about the protagonist of this activity. Horizon is specially built for advanced autonomous driving applications-Journey 5, which will also be the first 100 TOPS computing AI chip in China to realize pre-assembly and mass production.
Computing power is not the only index to measure chip performance.
Nowadays, when we talk about the performance of a driver assistance chip, the first thing that comes to mind is its TOPS (computing power). The higher the TOPS value, the higher the computing power of the chip. However, from the horizon, TOPS is not the only criterion to judge the performance of the chip. They think that the most reasonable standard to truly evaluate an AI chip should be FPS/Watt or FPS/$, that is, the power consumption or chip cost paid to achieve such AI processing performance must be processed by algorithms.
In order to explain the above equation in detail, on the left is the "FPS/Watt or FPS/$" just mentioned, which will be affected by three factors. First of all, the hardware architecture brings "tops/watt &；; Tops/$ "depends on the continuous evolution of technology and manufacturing process, that is, the manufacturing process we often say is from 28nm, 16nm, 7nm, and then to 5 nm; Secondly, the processing speed of the algorithm based on the unit effective computing power of "FPS (frames per second)" obeys the new Moore's Law in the field of algorithms, and the number of calculations and complexity required to complete the same AI task every 9- 14 months will be reduced by half. Finally, the software architecture is "utilization". When the aforementioned hardware architecture and algorithm architecture are locked, the software architecture can be continuously improved, and the compiler can continuously schedule the same algorithm on the same chip through compilation, disassembly, reorganization, deployment and operation, thus improving performance.
Therefore, thanks to the horizon's understanding of chip performance, Journey 5 improved the maximum computing performance from 1283FPS when it was released in July, 2002/kloc-0 to1fps today, achieving a nearly 20% improvement in the performance of Journey 5.
Further digging, the powerful performance of Journey 5 is the third generation BPU (Brain Processing Unit) Bayesian architecture built by Horizon for advanced autonomous driving. Its biggest characteristics are high performance, low delay and low energy consumption.
At the beginning of designing this BPU Bayesian, Horizon focused on the latest neural network architecture and the algorithm concerned in the autopilot scene, insisted on the high parallelism of software and hardware, and used storage and calculation as much as possible, so that it could provide enough computing power density and energy efficiency ratio under very flexible conditions. At the same time, BPU could complete an AI reasoning task in a shorter time through compilation and optimization.
Here, we can have a look at the performance comparison between Journey 5 and Orin on the typical classification model of NVIDIA, but let me explain in advance that this set of data is not measured by Orin, but based on the indirect calculation of Xavier, the last generation of autonomous driving chip in NVIDIA.
Similarly, for the object detection of EfficientNet, which is closer to the automatic driving scene, the performance advantage of Journey 5 leads the whole with the same accuracy, reaching nearly three times the advantage of Orin. Because the power consumption of Journey 5 is very low, only 30W, it still has lower power consumption under the condition of achieving higher performance, so its energy efficiency ratio is nearly 9 times that of Orin.
Fully link the product matrix to make the product easier to use.
Beyond the chip, Horizon also hopes that through its rich, mature and perfect set of system tools, car companies and eco-partners can develop and deliver their related products in a more efficient way.
To this end, Horizon has launched the corresponding vehicle intelligent development platform, including AI chip and software stack, Tiangong Wu Kai AI tool chain and Eddie AI development platform. Among them, Tiangong Wu Kai AI tool chain can provide a very efficient automation link to help customers realize the quantitative process with minimum precision loss. Aidi AI development tool platform is a driver engine facing the trend of software 2.0. Based on all kinds of chip tools provided before Horizon, we make corresponding algorithms, models and applications, and use such an end-to-end tool to realize subsequent iterations and solve problems efficiently.
Lv Peng, Senior Director of Horizon Intelligent Driving Product Planning and Marketing, was quoted as saying, "From the top-level reference algorithm to the following application middleware, basic middleware, operating system and hardware reference platform, and then to the chip, tool chain and AI development platform, a whole set of development environment actually helped our customer greatly reduce the difficulty, time and cost of his development on Horizon chip platform, which brought about the improvement of the whole development efficiency. Our partners can receive more projects, get more models, and better serve good car companies in terms of efficiency improvement. This is also the concept of Horizon, hoping to help our partners and car companies reduce the difficulty and cost of development. "
In fact, they did. Horizon mentioned an example of an intelligent robot in the activity. With the help of tools related to Horizon Chip Tool Chain, the intelligent robots completed the function migration in less than two months, and quickly realized the requirements of various mass production indicators of the system, and finally helped them to continuously deliver mass production on the chip, saving a lot of time and cost.
Chip security cannot be ignored.
The last thing to say must be safety, especially in the field of automatic driving, and the safety of the chip will become more important. The performance of Journey 5 in the security field is very outstanding. Horizon can be said to be the AI chip company with the most complete security certification system in China. Journey 5' s complete chip scheme meets ASIL-B product certification standards and will receive a full set of certification. Horizon is the first AI chip company in China to obtain ISO 26262 functional safety process certification, and Zhengtu 5 is also the first automotive chip in China that is completely based on ISO 26262 process.
Write it at the end
In April this year, this Journey 5 from the horizon has completed the urban road test on the real vehicle, and this year Q4 will put the SOP on the production vehicle. In addition, up to now, Journey 5 has won the designated production car projects of BYD, FAW Hongqi and self-propelled tourists. So what are your expectations for its future performance in the field of driver assistance? Welcome to click "+Follow" above to leave a message for discussion with me.