Importance of Horizon/Journey success
As AI/ADAS become increasingly important, the AI inference chip behind that has become very important.
In the world of AI training, there are basically 2 options for Chinese firms: Nvidia/Cuda & Huawei/Ascend. Now that former has been sanctioned, the latter is only real viable option. Other Chinese AI training type of chips are either in too small production or not powerful enough to train models quickly
However, the demand for AI inference chips are not as high, since the AI models are already trained & inference chip just needs to make decisions based on the parameters and its current situation. You also don’t need high chip-to-chip interconnect speed since most models prefer to just use 1 chip for everything. 4 Nvidia Orin-X chip of 254 TOPS together does not actually give you 1016 TOPS computation in actual practice
As such the entrance barrier for AI inference chips is a lot lower. While, Nvidia occupies the high end market (with Jetson previously, Orin now and Thor in the future), many other chipmakers found success in the middle. Most famous of which is TI’s TDA-4 (current version has 30 TOPS of AI & 100K DMIPS of computation). Another one is Renesas’ V3H (7.2 TOPS & 9.2K DMIPS) which is used transforming data from less intense sources like radars & mirrors. Since the entrance barrier is low, they can also be easily replaced.
Recently, an AI startup Axera has come out w/ M55H (8 TOPS + 7.6k DMIPS) for simple smart cars (1 main camera + 1 forward looking MMW Radar, aka 1V1R config) & M76H (60 TOPS + 35k DMIPS) for more complicated 6 camera solutions. See various scenarios that M55 & M76 can cover below
There is also Black Sesame which has come up with A1000 AI chip as well as C1200 SoC, but I will explore them in another piece.
The most promising Chinese AI firm is Horizon Robotics and Journey series AI chips. It was found in 2015 and released its first chips Journey 2 into cars in 2019. Since then, it has come out with Journey 3 (J3) and Journey 5 (J5). From road map here, you can see how far it has advanced in by the time J5 came out. A wide list of domestic automakers are using J3 for low-end ADAS solution and J5 for medium-end solution supporting Highway NOA. NOA means Navigate On autopilot, so cars supporting HW NOA can theoretically drive Highway portion with minimal driver input. In practice, different products currently have different HW NOA performance. The current holy grail is all scenario NOA, which means car can drive Highway and city portions with minimal driver input. As you can see, Horizon has many customers.
And based on 1st half 2023 figures, Horizon is the 2nd most popular ADAS chip supplier in China
The battle in China is between Horizon and Nvidia. Huawei actually provides a better all-inclusive solution with MDC-610 platform as shown below, but most Huawei competitors don’t want to have to use its product. So far, it is limited to just Huawei partnered cars
You can see a very advanced solution here with its own AI portion, CPU portion, ISP portion in an integrated SoC as well as high connectivity to DRAM, SSD MCU module and data module. As well as high speed connection via CAN to various inputs from Lidar, MMW radar, camera and such. Decision making will then send commands via CAN to various ECUs in the car. Different ECUs are in charge of different electrical control for physical action. Interesting part here is also the integrated nature with Cockpit/Infotainment system (see MMI - Man Machine Interface)
Basically, Huawei’s current domain control/computation architecture is already close to pinnacle of what you would want EV’s smart computing platform to look like. No one else’s platform look like this yet.
In Nvidia’s case, their chip only supports AI & CPU portion, so automaker needs to integrate in the Cockpit SoC, central MCU, CAN/ETH, different analog chips and the computation platform.
In Horizon’s case, J5 simply isn’t powerful enough. Freetech’s ADAS solution for FAW Hongqi shows why this is the case
Freetech integrated 3 J5 chips along with 2 TI TDA4VH chips & 2 Infineon TC397 MCUs to fully support sensor suite of 12 cameras, 5 MMW radar and 3 Lidars. Each J5 chip offers 128 TOPS & 26k DMIPS, so they do not provide enough CPU computation to make predictions and do sensor fusion work. They did the hard work of transforming raw Lidar/Camera/Radar data to objects. They passed this data to TDA4s to do the remaining work. 444 TOPS & 273k DMIPS is the theoretical combined AI & computation power here for processing all these sensors and make decisions. In reality, additional chips add to cost, creates redundancies & add complexity from integrating multiple company hardware to same software. Just think about trying to get Journey chips to work in sync with TI chips for the first time. That’s not simple.
As such, numerous ADAS startups now utilize purely domestic solutions in order to establish supply chain security and to sort out integration issues much earlier on. It is easier to get 2 small companies like Horizon & SemiDrive to integrate their chips in same computation platform than it is to get Horizon to integrate with TI and Infineon. Small startup-like Chinese chip firms speak the same languge, work in the same timezone and are just eager to give their all to secure contracts. If TI lose some ADAS business in China, they still have contracts elsewhere. For example, TzTek’s TADC-D52 integrates 2 J5s with 1 SemiDrive X9U CPU and 1 SemiDrive E3 MCU to provide the core computational component of a platform designed to achieve City NOA & high level ADAS. Neusoft has similar solutions with J5 and SemiDrive chips
Each X9SP provides about 100K DMIPS through its 14 A55-cores. Both J5 and X9 are fabbed with TSMC 16nm process, so nothing cutting edge. With enough work, you can get Highway NOA with this solution.
And then came Journey 6, which I detailed here
Not only does it provide 560TOPS, but also 350K+ DMIPS CPU computations as well as integrated GPU and MCU on the same SoC. The only thing missing is integrated Cockpit functionality. They will likely work with SemiDrive, Rockchip or Unisoc to create integrated central computation platform capable of ADAS, cockpit functionality and network communication.
While J5 was certified to ASIL-B standard (doesn’t look like J3 got any), J6 will be certified to ASIL-D standard. That’s the highest possible level and should satisfy any customer.
J6 is capable of processing 5.3Gpps image processing, supporting 24 cameras + Lidars + radars connected through TB-level databus with just 130ns latency and supports PCle 4.0.
The level of computation provided here is enough to support full scope NOA. Short of L4 ADAS, this 1 J6 chip is all you need in computation platform. They were able to sign on heavy weights like BYD, Li Auto, VW & GAC right from the get go. Bosch is also signed on. Likely to integrate its ECU more closely with J6 chip. Bosch plays a huge role in China Auto Inc’s move toward drive-by-wire. It has huge facilities in China that is now sanction proof. Chinese automakers are basically working with Bosch Huayu like its a local company.
Either way, this level of heavyweights in initial product unveiling shows that Horizon has made it. It is likely the winner of battle for ADAS AI inference chips because it provides a chip that has the perfect spec to do everything you need for L3 ADAS. It can fully support 3 Lidars, 5 MMW Radars & 12+ Cameras. It is fully integrated in 1 chip and achieved highest auto grade standard. Even BYD Yangwang series can get all the computation power it needs with just 1 J6 SoC + the latest QCOM cockpit chip it has.
J6 will also achieve this at much lower price point than competition from Nvidia. Here is a list of chips and their prices based on recent report
While I think J5 is likely $150-200, other prices are within ballpark of what I expect. Nvidia’s chips are extraordinarily expensive. It also uses the most expensive Arm core available at the time of development for its CPU component (A78AE for Orin and Neoverse-V2 for Thor). As such, J6 will likely sell for less than 1 high-end Orin and probably 1/4 of Thor.
There is also the additional drawback of Nvidia Thor possibly being unavailable to Chinese customers due to recent AI chip sanctions on China, since even RTX 4090 was found to be too powerful. As such, J6 will inevitably be the best available chip for Chinese automakers at a reasonable price.
Also, you will notice that SemiDrive and Horizon chips on this list all use A53/A55 cores on 16nm process. A5x core series are weak cores used on phone SoC for low power scenarios. One may ask why they are using such weak cores, when Nvidia and QCOM are using latest ARM cores? Even TI TDA4VH charges more and use an older performance core A72.
That’s where success really matters. These Chinese design firms are new and don’t have much money or customer base. As such, they had to go form the value play and design CPUs with these low power cores.
That changed with J6. Due to establishing many powerful partners, Horizon designed J6 with 14 A78AE cores for CPU and Mali G78AE core for GPU. As such, J6 provides 350k+ DMIPS while J5 provided just 26K DMIPS. Better GPU provides the better frame rate and rendering performance of J6. Horizon also likely gained much experience in raising CPU performance from J5 to J6. It likely started off focused on AI computation and was able to hire people with CPU design experience after achieving success in J3/J5.
I think this shows the importance of Horizon Robotics and Journey series. Success breeds further success. Horizon started at a good time, so was able to capture all the domestic automotive partners that wanted a backup solution to Nvidia. The sanctions on AI chips encouraged domestic partners to get further into bed with Horizon. They responded by working closely with everyone and being very responsive.
Comment above show the difference between working with Horizon vs Mobileye, TI and Nvidia. Domestic automakers simply did not get software support from TI that it gets from Horizon. Mobileye would only do blackbox cooperation initially. I guess now Mobileye has lost the entire Chinese market due to its reluctance to play ball.
Horizon’s willingness to work with automakers secured it this many customers
150+ car models have picked Journey series chips. 4 million chips have already been delivered. That’s success for a little known company outside of China. It has also managed to lock down VW as I documented here.
I don’t see this as the end for Horizon. They have a foundation now to expand to other chip design areas. They should aim for a fully integrated ADAS/cockpit chip. They should get more into MCUs. They should bring their AI inference chip technology to robots and other fields. They should also try to get into AI training chips.
Every company in China wants a real domestic AI chip competitor to Huawei. Maybe Horizon Robotics will be that company.