Autonomous Driving SoC Research Report, 2024

Autonomous Driving SoC Research Report, 2024

Autonomous driving SoC research: for passenger cars in the price range of RMB100,000-200,000, a range of 50-100T high-compute SoCs will be mass-produced.
According to ResearchInChina’s statistical analysis of its database, in 2023 the penetration rate of L2+ NOA (including L2.5 highway NOA and L2.9 urban NOA, with hardware pre-embedded) in new cars in China stood at 9.67%, of which the penetration rate of L2.9 was 4.88%.

In 2023, the total sales of new cars supporting L2+ NOA (L2.5&L2.9, with hardware pre-embedded) were around 2.12 million units. Currently, the intelligent driving SoCs installed are led by Tesla FSD, NVIDIA ORIN-X /ORIN-N/Xavier, Horizon Robotics J5/J3, Huawei MDC Ascend 610, TI TDA4VM/ TI TDA4VH, and Black Sesame A1000.

50-100T high-compute autonomous driving SoCs will be quickly introduced into passenger cars in the price range of RMB100,000-200,000.

In China, passenger cars priced at RMB100,000-200,000 are still the main sales range. In current stage, the penetration rate of L2+ NOA (L2.5&L2.9, with hardware pre-embedded) in this price range is still very low, mainly due to costs.

Passenger cars in the price range of RMB100,000-200,000 are still mainly configured with entry-level L2 autonomous driving at this stage, but the installation of L2.5 autonomous driving is also increasing rapidly.

Qualcomm autonomous driving SoC -SA8650/SA8620

Qualcomm SA8650 targets passenger cars in the price range of RMB100,000-200,000. According to Qualcomm's consistent product thinking, cost-effective and rich ecosystems are its main product direction.

SA8650 boasts CPU compute of 230kDMIPs, comparable to the top-end Orin. Orin has multiple versions, and only the top-end Orin-X features CPU compute of 230kDMIPs. It is supposed that SA8650 is composed of 4 Cortex-X3 large cores and 4 A55 small cores. Limited by costs, Chinese chips cost less in CPU, with 8 A55 cores at most and general computing power of 26kDMIPs. CPU costs much more than AI, and must have advanced processes. X3 must correspond to 4nm. Orin uses 12 A78AE cores, all large cores, to be on par with SA8650.

SA8650 has AI compute of 100TOPS, higher than the main competitor NVIDIA ORIN-N, and also has strong graphics output capabilities and supports up to 4 screens. SA8650 can correspond to 12 cameras, namely, 8 8MP cameras and 4 4MP cameras. The power consumption of SA8650 is about 25-40W. If it exceeds 25W, it will be best to use a water cooling design. Currently, the design scheme of SA8650 is water cooling.

Qualcomm has not provided chips individually since the fourth generation. All of its chips are sold in the form of modules. A module is composed of a SoC, 4 power management ICs, and 2-4 LPDDR DRAMs.

By one estimate, Qualcomm SA8650 module includes one SA8650 SoC, four PMM850U power management ICs, two Micron LPDDR5 DRAMs, and one Micron UFSs. The model of Micron LPDDR5 is MT62F3G32D8DV-026 AAT:B, codenamed D8DHD, with 8 die in package, 12GB per die, and the speed of 7500Mb/s. The capacity of one UFS may be 256GB, or it may be three LPDDR5 DRAMs. There is also another chip, which may be a PCIe interface or a temperature protection chip.

The first-generation Qualcomm Ride (SA8540P) had few responders, but the second-generation Ride (SA8650P, SA8620P) has been accepted by quite a few European and American OEMs, including BMW, Mercedes-Benz, Audi, Porsche and Stellantis. It has also been accepted by many mainstream Tier1s such as Valeo, Continental Automotive, Bosch and Veoneer.

Among Chinese Tier1s, Desay SV, Joyson Electronic, EnjoyMove Technology, Haomo.ai, and Baidu have all been developing based on SA8650 for nearly a year; Hangsheng Electronics, Zongmu Technology, Autolink, MEGA, PATEO, and Freetech are introducing SA8650. A large number of OEMs are already interested in it.

Advanced Version of DJI Chengxing Platform

Based on Qualcomm Snapdragon Ride (SA8650P) and without relying on HD maps and LiDAR, it enables point-to-point NOA on urban roads and expressways, and cross-floor home-zone parking assist.

The 7V+100TOPS (Qualcomm SA8650P) version of Chengxing Platform solution enables advanced intelligent driving functions like urban NOA, which is inseparable from DJI's expertise in stereo vision, omnidirectional fisheye perception, and high-performance optimization. In addition, on the 100TOPS domain controller, DJI's latest achievements in stereo OCC, road topology model, prediction and planning model, end-to-end model optimization and so forth are also deployed.

Based on the 7V sensor configuration, it can be expanded to a 10V configuration, further enhancing the ability to cope with individual urban scenarios (such as ultra-wide and ultra-long intersections).

Momenta signed an agreement with Qualcomm

Mainly based on NVIDIA ORIN solution, Momenta’s advanced intelligent driving has already been designated by OEMs like SAIC, BYD, GAC and GM, and has begun to be delivered.

On April 22, 2024, Qualcomm and Momenta officially announced power-efficient and scalable architecture of SA8620P (36TOPS) and SA8650P (100TOPS) to enable ADAS and AD functions ranging from Highway Navigation Pilot (HNP) and Urban Navigation Pilot (UNP).

Chip: SA8620P (36TOPS), SA8650P (100TOPS)
Perception configuration: 7V3R/7V1R
Functions: Highway Navigation Pilot (HNP, highway NOA), Memory Navigation Pilot (MNP, urban commute NOA), Learned Parking Navigation Pilot (LPNP), and Parking Navigation Pilot (PNP)
Target customers: The solution is aimed at mainstream passenger cars pried at RMB100,000 to 200,000, with the entry-level coming with highway NOA as standard, without the need to pay extra optional fees. It is known that Momenta's Qualcomm-based platform has been mass-produced and applied by the two OEMs Toyota and GM.

Black Sesame Autonomous Driving SoC - A1000 Series

On July 17, 2024, Dongfeng eπ 007 offered significant upgrades in intelligent driving functions, including advanced intelligent driving functions such as highway NOA and Long-range Automated Parking Assist (LAPA), as well as a number of practical functions and experience optimizations, for intelligent driving models.

The high-configuration model of Dongfeng eπ 007 packs Black Sesame Huashan® A1000 automotive high-performance autonomous driving chip, which fully supports L2+ advanced intelligent driving assistance. Dongfeng eπ 007 can easily cope with daily intelligent driving needs after OTA updates. Huashan A1000 allows eπ 007 to be equipped with a driving-parking integrated high-compute intelligent driving system. The highway NOA enables full-course path planning, and can automatically complete tasks of following, automatic lane change, and on-ramp and off-ramp on urban expressways and highways.

A single Black Sesame A1000 SoC can support complete driving and parking functions, providing OEMs with high-performance and cost-effective driving-parking integrated autonomous driving solutions. At present, Huashan-2 A1000 is in full mass production and has been adopted by multiple leading Chinese OEMs including FAW, Dongfeng, Geely and JAC, with production models including Lynk & Co 08, Hycan V09, Lynk & Co 07, Dongfeng eπ 007 and Dongfeng eπ 008.

Single-chip cockpit-driving cross-domain integration is becoming the focus of the market.

As EEA tends to be centralized, the integration of intelligent driving domain and cockpit domain has become a trend.
NVIDIA, Qualcomm, SemiDrive and Black Sesame Technologies among others were the first to release cockpit-driving integration chips; the likes of Baidu, iMotion, Bosch, ZF, Desay SV, Hangsheng Electronics and Z-ONE have also announced single-chip cockpit-driving integration domain control products and system solutions;
It is expected that during 2024-2025, single-chip cockpit-driving cross-domain integration solutions will be massively installed in vehicles.

Currently the industry has launched SoC products for cross-domain integration, such as Qualcomm 8775, Black Sesame C1296, SemiDrive X9CC, and NVIDIA Thor. Tier1s are developing system solutions based on these chips.

Continental has implemented software-defined vehicle architecture based on Qualcomm 8775. It adopts SoC+MCU architecture, with Qualcomm 8775 SoC and NXP S32G MCU. Continental builds a software system based on its CAEdge software framework to enable vehicle-cloud integrated cockpit-driving integration solutions.

Baidu is also developing Apollo Robo-Cabin, a cockpit-driving integrated computing platform based on a SOC to enable both basic intelligent driving functions such as AEB and LCC and cockpit capabilities. It adopts a cockpit-driving partition isolation design that isolates hardware, communication links and resource scheduling to ensure the independence of business, communication and operation.

In the future, the ultimate goal of this chip-software integrated solution is to realize the urban commuting + automated parking 2.0 and intelligent cockpit capabilities in terms of intelligent driving functions. Of course, seen from Baidu's layout plan for intelligent driving solutions, the underlying computing chip should at least be a medium-compute chip by then. The current mainstream 8295 is still a bit overstretched.

In the cockpit-driving integrated intelligent driving mode, based on its own OS SOA, Baidu Apollo provides complete open architecture with atomic AI capabilities, realizing multi-modal integration and creating human-machine cooperative intelligent driving, and full-vehicle intelligent experience of the intelligent cockpit. Baidu Apollo has deployed intelligent driving products, for example, all-scenario AVP, driving-parking integrated highway pilot product ANP2, and three-domain integrated urban NOA product ANP3.

At Auto China 2024, Black Sesame Technologies exhibited the mass-produced chips in the Wudang series, C1236 and C1296. A single C1236 integrates sensor access for NOA domain control, algorithm acceleration, line-speed data forwarding, 4K display, etc. Black Sesame Technologies and FAW Hongqi jointly launched a single-chip intelligent vehicle control project based on the C1200 family. The new cooperative solution is based on the C1200 family and will cover intelligent driving, vehicle data exchange and control functions.

Black Sesame Technologies officially announced CoreFusion, a C1296-based cockpit-driving integrated software open platform jointly created with Nesinext.

Will SoCs self-developed by OEMs become mainstream?

Background of OEMs strengthening independent development of SoCs
Cost control: The unit price of the mainstream ORIN-X SoC is at least USD250, that is, more than USD500 for two. The price of Thor is at least USD500 per unit. OEMs develop SoCs in house, hoping to achieve cockpit-driving integration with a single chip and replace the current NVIDIA ORIN + Qualcomm 8295 combination to cut down costs. Typical examples include NIO NX9031, and Xpeng and NVIDIA's customized 750T version of Thor. However self-developed SoCs need to achieve scale effect, and with shipments of at least 1 million units, they will be cost-effective.

Independent and controllable: in the context of US sanctions, relying on only Orin, Thor or Qualcomm 8295/8255 brings a bigger supply chain risk. OEMs prefer to introduce more domestic alternatives into high-level autonomous driving. A typical example is SiEngine, which is supported by Geely and has developed AD1000 intelligent driving SoC and SE1000 cockpit SoC.

Product customization and AI-oriented development: the SoC designs of NVIDIA, Qualcomm and other companies are oriented to general needs. IP, circuit design, tool chain and other links all target the general needs of customers, resulting in high complexity in chip design. Yet OEMs can design chips just based on their own needs and do not need to open them to the outside world, so the chip design is less complex, and highly integrated with their own intelligent driving algorithms, and even cloud AI chips. Typical examples include Tesla FSD and cloud Dojo chips. Li Auto is also self-developing intelligent driving SoCs and cloud AI chips, and plans to tape out in 2024 and start production as early as 2026.

Challenges faced by OEMs in self-developing SoCs
-In the current market environment, it is quite challenging for OEMs to independently develop chips. On the one hand, they have to compete with professional chip design companies in development speed, product definition capabilities, human resources and management capabilities; on the other hand, it is difficult for a single OEM to purchase tens of millions of chips, so whether they can make continuous investments is a huge challenge.

The annual shipments of an automotive SoC are lower than one million units, making it difficult to support the continuous investments in chip R&D. If OEMs do not have sufficient shipments, it will be difficult for them to recover the costs, and lowering the chip cost of vehicles will be even a false hope.

Independent chip manufacturing is more suitable for OEMs or OEM alliances with large shipments (SiEngine has developed multiple leading customers like Geely, FAW, Zeekr, and Volvo; Xpeng has chosen to form an alliance with Volkswagen) or OEMs with high unit prices of vehicles (NIO and Li Auto target high-end users).

For OEMs, chip + operating system must be developed in pair to bring its performance advantages into full play.

In September 2023, NIO introduced SkyOS, an all-domain vehicle operating system for intelligent electric vehicles. It is the underlying operating system for NIO cars. According to NIO's plan, the full-function mass production of SkyOS will be realized on NT3 platform-based models, which will be equipped with the self-developed intelligent driving SoC Shenji NX9031.

It is speculated that Shenji NX9031 should be a 5nm cross-domain integration chip mainly for autonomous driving. To save costs, it is expected to be manufactured by Samsung, just like Tesla FSD, as OEM by TSMC is too expensive.

Transistors: 50+ billion
CPU, ARM IP, Cortex-A78AE, 32-core CPU configuration, and both large and small core configuration, performance of 615KDMIPS, far higher than NVIDIA ORIN (228KDMIP, 12-core CPU)
ISP, developed in-house by NIO, high dynamic range; 26 bits (bit width is mainly subject to the dynamic range of ADC, the higher the bit width of ISP the higher the frame rate can be, and the bit width of most ISPs are only 10 or 12 bits); image bandwidth of 6.5GPixel/s, far higher than NVIDIA ORIN (1.85Gpixel/s); processing delay of <5ms
NPU, the core of intelligent driving SoC, has computing power equivalent to 4 ORIN-X SoCs; the self-developed reasoning acceleration unit NPU (NPU TPP compute, total power performance: <4800) can run various AI algorithms flexibly and efficiently.
Memory: LPDDR5X, bandwidth of 8533Mbps, performance higher than LPDDR5
Bus: ARM, NOC bus IP, building a data transmission channel between CPU, BPU, GPU and MCU
Functional safety: ASIL-D level, just adding an MCU island, generally 2 to 4 Cortex-R52 cores for lockstep



Please Note: PDF E-mail from Publisher purchase option allows up to 10 users and does not allow printing or editing. This functionality will require a Global Site License.