Accelerating AI Computing to Fuel ADAS Evolution

High-performance Computing is Essential
 

Changing lanes and heeding traffic lights. What millions do routinely during their commutes illustrates some of the critical bottlenecks in current AD/ADAS computing. Besides changing lanes, mere mortals can easily identify, analyze, and anticipate pedestrian intent, as well as determine the phases of a traffic light from a fairly long distance, day or night, and in real-time.

For an ADAS application to function successfully in an individual vehicle, it must first identify small objects that could be stationary or moving, and up to 250m away. Most modern vehicles already have multiple high-resolution cameras, up to 8MP to capture images. However, analyzing the volume of data acquired from multiple cameras and other sensors, and acting responsibly in real-time demand both exceptionally high-performance and low-latency computing with optimal power efficiency. 

Scaling to many cars on a busy street, systems must also be able to work seamlessly in real-time with other vehicles and sensors in proximity, and at Big-Data-like volume that adheres to stringent privacy/security requirements. How must ADAS computing evolve to execute this at least as well as humans do today and at scale? The answer may lie in purpose-built, power-efficient SoCs developed not by traditional automotive manufacturers, but by a new set of agile startups.

AD/ADAS Compute Is Positioned for Growth
 

Computing is estimated to represent more than ten percent of the annual market opportunity for ADAS by 2025 (Source: Bank of America). This is driven by the ADAS market, which is estimated to be USD 96.7B by 2030, growing at over 30% CAGR (Source: UBS Global Research).

To change lanes or heed traffic lights, an inferencing system requires exceptional computational efficiency. For example, it must identify a euro standard pallet (20cm x 20cm x 40cm) from 160m distance in any lighting condition. Doing this successfully requires at least 8 MP camera resolution and a 120dB dynamic range image sensor operating at 40 fps. The processing power required to infer such classes of objects will require a few hundred TOPs while operating at 25ms cadence.

ADAS computing systems include both Systems on a Chip (SoCs) and specialized AI Accelerator Chips, to leverage deep neural network algorithms as efficiently as possible. Beyond optimizing performance per watt, or performance per SoC footprint, these new devices must also speed up Deep Learning for perception and sensor fusion tasks.

High-performance, low-latency ADAS computing depends on processing power delivered by these new types of SoCs. AI Accelerator SoCs will determine the overall system performance, and hence command a major share in profit margins of the ADCU system. The transition from decentralized to centralized architecture is further driving the growth of AD/ADAS compute.

AI Semiconductor Landscape
 

Traditional automotive semiconductor manufacturers are lagging in their AI development as they must simultaneously contend with incremental near-term ADAS features for L2. This means opportunities for agile, innovative startups or spinouts to define and develop compute systems and chipsets for L3 and beyond.

Currently NVIDIA and Mobileye target this performance gap. Both leverage their leadership in GPUs and Smart Camera respectively. Combined with software toolchains and future roadmaps, both companies will raise the bar with subsequent releases.

Despite their promising roadmaps, there is still room for a giant leap in PPACAT (Performance Power Area Cost Accuracy Tools). This can only be accelerated by newer companies fostering healthy competition. At the end of the day, developing AI SoCs is a game of efficiency, safety, and yes, profitability.

It is a near-greenfield opportunity for start-ups and new entrants. In the last few years, over 45 startups have emerged in the specialized AI chip space, including non-automotive applications as well. Over $3B venture capital investments went into this sector in the last four years, comprising of over 20 startups. These include AIStorm, AlphaICs, Ambiq, Baecsense, Blaize, Cambricon (IPO), Cerebras Systems, DeepVision AI, EDGENeural AI, Efinix, Flex Logix Technologies, Graphcore, Groq, Hailo, Horizon Robotics, Kneron, Mythic, Nuvia(acquired), Recogni, SambaNova, SiMa.ai, Syntiant, Tenstrorrent, Untether AI, Vastai Tech, Wave Computing, Xceler Systems, etc. (source: Pitchbook, 2021).

Of these startups, Blaize, Hailo, Horizon Robotics, and Recogni are more focused on AD/ADAS applications. Horizon Robotics and Recogni stand out with their higher targets for performance and power efficiency.

AD/ADAS Computing is an Open Opportunity

Startups, “intrapreneurs” of traditional companies, and spinouts within the Compute, Data Center, and AI/ML accelerator SoCs segments are poised to re-define and develop the computing gap left by traditional automotive semiconductor vendors. AI accelerator SoCs can deliver high-performance, low-latency, power-efficient computing in ADAS applications, as well as high value for its automotive OEMs. These smaller, agile companies provide the innovative spark to move the industry collectively forward, and companies like Continental can assist by leveraging its scale and ecosystem partners.

Once proven, AI-enabled vehicles can give drivers and passengers peace of mind, an important emotional hurdle for transitioning to autonomous driving. Purpose-built, power-efficient SoCs and AI accelerators will reliably enable higher levels of ADAS and true autonomous driving, fostering safer, more efficient vehicles all over the world.