You have GPU, I have DSA.
Editor in charge丨Li Sijia
When it comes to the field of AI chips, a recent major event is that on October 13, Horizon and Volkswagen’s software company CARIAD officially announced their cooperation. The two parties have established a joint venture, with CARIAD holding 60% and an investment of about 2.4 billion euros. It is expected to be completed in the first half of 2023.
This is Horizon’s heaviest collaboration to date. And an episode is that half a year ago, Horizon announced the first fixed-point cooperation of the third-generation product Journey 5 chip, which will be spent by BYD, and the boarding time will be 2023. However, what is interesting is that BYD will also put into production models equipped with NVIDIA’s DRIVE Orin computing platform.
And the recent article of the C-dimension of the commune “What Nvidia doesn’t want, Horizon does” also introduced the competition between Horizon and Nvidia in the application of AI chips in the automotive field. Therefore, in this article, we will focus on popular science. In the end, these two companies What is the competition, and what are the considerations behind the choice of car companies?
All are AI chips, what’s the difference?
When it comes to both parties, their products are AI chips for autonomous driving. But the difference is still huge.
The application scenarios of AI chips are mainly divided into cloud and terminal, as well as edge terminal . At present, there are more cloud applications and they are relatively mature. In addition, cloud applications can be divided into two types: training and reasoning, among which training accounts for a relatively high market size.
The overlord of cloud applications is Nvidia. In the application scenarios of cloud (server, digital center) and terminal (mobile phone, smart car and other mobile terminals), the computing methods of AI chips are fundamentally different.
For example, when the cloud processes a large batch of accumulated data that arrives at one time (expanding the batch size, batch size), it can “wait” for “enough” of the data to start processing. However, the on-board chip needs to process stream data, and the data arrives one after another with the driving (time); the calculation needs to be completed in real time, and the delay should be reduced as much as possible.
For interactivity, the terminal requirements are higher. The cloud task itself is limited to the virtual world, without considering the interaction with the real world. The terminal is in the real world, and each task needs to consider interactivity.
In addition, power consumption and cost also occupy a heavier weight in the consideration of AI chips in the car. It can be seen that for automotive terminals, in addition to ensuring computing energy efficiency, AI chips also require low power consumption, low latency, and low cost .
At present, there are three main types of AI chips, namely general-purpose GPUs, customizable FPGAs, and dedicated ASICs. Nvidia’s Orin chip is based on a general-purpose GPU. The Horizon Journey 5 chip is both an ASIC chip and a DSA (Domain Specific Architecture) chip. It is worth noting that the core GPU architecture of the Orin chip is still the cloud architecture.
However, for car companies, the stronger the compatibility of the chip, the better, because car companies can do it in one step, and there is no need to re-verify, using one chip can realize the hardware pre-embedding of future high-end intelligent driving , for the future. Get ready for advanced smart driving. This is also the reason why many brands are willing to choose NVIDIA Orin chips.
Horizon uses a combination of software and hardware methodology for specific scenarios to design chips, that is, DSA chips, which greatly improves the effective computing power of the chips. However, our question is that after the limit of the Journey 5 chip is reached, the higher-level autonomous driving needs will be replaced by the Journey 6. Then, who will pay for the comprehensive cost of this replacement?
In addition to computing power, there is also FPS
In addition to the difference between general and customized, there is currently a “computing power theory” in the industry. It seems that the higher the computing power of the chip, the better. Actually, this is biased. In fact, it depends on the FPS (frames per second).
FPS generally refers to the number of frames per second of animation or video. FPS is a measure of the amount of information used to save and display dynamic video. Simply put, the higher the FPS value, the better the performance. For all computing platforms used in autonomous driving, FPS is considered as one of the evaluation criteria to measure the operating efficiency of advanced algorithms.
For example, in terms of chip computing power, the computing power of single-chip NVIDIA Orin and Journey 5 are 254TOPS (trillion calculations per second) and 128TOPS respectively, the power is 45W and 30W, and the power consumption ratio is 4.6TOPS/W and 4.2TOPS respectively. /W. Journey 5 is only half the size of Nvidia Orin.
However, the FPS (frames per second) of Journey 5 is 1283. It is much higher than Orin’s 1001FPS (measured after adjustment based on the NVIDIA RTX3090 with the same ampere architecture). The smoothness of video transmission in Horizon Journey 5 is obviously better than that of Nvidia Orin. No wonder Luo Hengyao, head of Horizon BPU algorithm, said, “In terms of energy efficiency, we (Journey 5 compared to Orin) have improved by more than 6 times.”
In fact, it’s not just Horizon that settles accounts with FPS. When the self-developed FSD chip was released in 2017, Musk, the founder of Tesla, compared the previously applied NVIDIA Drive PX2. In terms of computing power, the FSD is 3 times that of the Drive PX2, but the FPS is 21 times that of the Drive PX2.
In addition, low latency performance is also a key indicator to measure the performance of AI chips. Because, the delay problem is directly related to driver’s license safety. For example, in an emergency braking scenario, 100 milliseconds means a braking distance of nearly 1.7 to 3.3 meters.
According to Horizon, the 8M single current perception structured output delay of Journey 5 is less than 60 milliseconds, which is better than the delay performance of NVIDIA ORIN. The low-latency data for the Orin chip is unknown.
Therefore, in terms of FPS, low latency, and low cost, Horizon still has great opportunities for development in the automotive field. Currently, more than 70 models of more than 20 car companies have applied the Journey series chips, which is also the way Horizon has customized routes. The results of the “Qi Bing” harvest.
Under the fifth wave of computing, the core that determines the chip, of course, is the architecture. Because the architecture is the core technology at the bottom of the chip enterprise, the iteration cost is huge, and it is also the most valuable technology.
As the Taishan Beidou who proposed the DSA architecture, John Hennessy and David Patterson pointed out in their acceptance speeches when they won the Turing Award in 2017: “The next decade will be the golden age of computer architecture.” This is a good footnote.
Horizon’s journey series chips, from the perspective of its architecture, have experienced three generations of BPU (Brain Processing Unit, an efficient artificial intelligence processor architecture independently designed and developed by Horizon) architecture named Gauss, Bernoulli and Bayes.
From low to high, the Gaussian architecture mainly deals with perception, that is, image recognition, image recognition of cameras, radars, sensors, etc., using a 40nm process. The Bernoulli architecture incorporates deep learning and a 20nm process. The Bayesian architecture is to join the Bayesian network, combined with deep learning to improve the accuracy of AI, 16nm process.
Back to Nvidia. NVIDIA, founded in 1993, is powerful in that it invented the GPU (graphics processing unit) in 1999, and released the CUDA general parallel computing architecture in 2006, realizing the decoupling of software and hardware. Developers no longer have to use a difficult GPU-specific development language, but can use a general-purpose programming language to invoke GPU computing power.
Through the decoupling of software and hardware, the dedicated chip GPU for graphics processing is turned into a general-purpose chip suitable for large-scale parallel computing, and NVIDIA stands on the C-bit of the AI era. For Horizon and other chip companies, it is also the object of high mountains and desperate pursuit.
Nvidia’s Orin is based on the Ampere architecture , which is also the main product architecture driving Nvidia’s data center business growth by nearly 90%. In other words, Orin, which is oriented towards high-level autonomous driving scenarios, still uses a general-purpose architecture modified by magic. It’s like the Qualcomm 8155 chip, which was popular this year, was magically modified from consumer electronics chips.
The benefits of a common architecture are “easy to get started”, ready to use, and rich in tools. However, the disadvantage is that many hardware performances are wasted in the early stage and are temporarily unavailable.
So, why is Nvidia in the cloud still using a common architecture? Just look at the proportion of the auto business.
On May 26, Nvidia released its Q1 financial report for fiscal year 2023 (as of March 2022), with revenue this quarter of $8.29 billion, a year-on-year increase of 46%. Among them, the data center business driven by hyperscale computing, cloud and AI business accounted for 45.23%; the game business supported by graphics card accounted for 43.67%. What about the car business? Insignificant, accounting for 2.1%.
Horizon’s revenue has not been publicly disclosed. However, in 2020, “LatePost” reported that Horizon’s revenue in 2020 will be 200-230 million yuan, of which 70% will come from the supply of chips and other products to car manufacturers. Customers include Changan, FAW and Lili Automobile.
From the perspective of revenue, the gap between Horizon and Nvidia is not an order of magnitude, and it cannot threaten Nvidia’s status at all. It is the relationship between whales and dolphins in the ocean world. Therefore, Nvidia has little incentive to make ASIC chips for car companies. This also gives Horizon and other chip companies a huge opportunity.
Of course, Horizon is still in the early stage of development, that is, the stage of constantly spending money. According to the data of Qichacha, if the undisclosed financing amount is not counted, Horizon has raised more than 3.4 billion US dollars (about 24.33 billion yuan). The most recent one is from Chery Automobile.
Horizon’s current core business is only the end-to-end automotive field. In terms of the market, the Horizon Journey series of chips have currently shipped more than 1.5 million pieces. And through the dedicated chip specially designed for the algorithms and needs of autonomous driving scenarios, achieving greater efficiency in efficiency is only the first step for Horizon to catch up with NVIDIA.
As mentioned earlier, BYD is not the only car company deploying both Nvidia and Horizon. Li Li is equipped with Journey 5 on Li Li ONE and the subsequent Li Li L8 Pro in 2021, while NVIDIA Orin is used on Li Li L8 MAX and L9.
Therefore, it is more difficult for the horizon to build a user-friendly software system and a user ecology that can support continuous evolution and iteration, and this is the ace of the NVIDIA CUDA system. Horizon’s creations still have a long way to go. However, in any case, Horizon brings a DSA chip option to Chinese auto companies, which is an “excellent” thing.