Recogni announces development of specialized AI hardware optimized for autonomous vehicle inference workloads.
As the automotive industry transitions to autonomous vehicles via sensor-fusion from cameras, radar, and LiDAR, the computational demands are enormous. Current AV computing systems consume considerable energy, which translates into reduced vehicle range for electrically powered vehicles. Recogni, a San Jose, Calif.-based startup, recently came out of stealth mode to address this challenge with its approach to perception processing for identifying bikes and pedestrians in real-time, claiming to use less power than alternatives such as Intel's Mobileye and NVIDIA's Drive Xavier. According to CEO RK Anand, "These vehicles need datacenter-class performance while consuming minuscule amounts of power. Leveraging our background in machine learning, computer vision, silicon, and system design, we are engineering a fundamentally new system that benefits the auto industry with very high efficiency at the lowest power consumption." Recogni claims its technology will handle the computational and inference work for an entire car using less than 10 W while processing data in real time.
The Recogni integrated module comprises three passively cooled image sensors, an external depth sensor, and a custom chip connected via Ethernet cable to external LiDAR or radar sensors. The custom chip can perform more than 1,000 teraoperations per second—capturing and analyzing up to three uncompressed 8- to 12-Mpixel streams at 60 frames per second. It recognizes objects, fuses depth-sensor information, and provides intelligence to the central system within 16 ms for urban settings and 8 ms for highway settings, achieving a claimed 70% compute efficiency in typical vision applications. The system can identify small objects such as traffic lights from over 200 m away in real time and determine their color—red, yellow, or green—capabilities that LiDAR and radar cannot match, according to the company.
The efficiency of Recogni's module stems from its reliance on passive cooling, eliminating power-consuming fans, and the onboard chip's close physical proximity to the three cameras, which reduces energy spent transferring sensory data. In total, the system consumes about 8 W of power. The founding team includes CEO RK Anand, a founding engineer at Juniper Networks; co-founder and Chief Business Officer Ashwini Choudhary; and Chief Technology Officer Eugene Feinberg, a former Cisco Systems engineer. All three previously worked together at a San Jose-based camera tech startup called mPerpetuo.
Recogni has received $25 million in Series A financing led by GreatPoint Ventures, with participation from Toyota AI Ventures, BMW i Ventures, Faurecia, Fluxunit (the VC arm of lighting and photonics company OSRAM), and DNS Capital. Choudhary announced the company is "currently in discussion with multiple auto manufacturers to provide them a full suite of enabling technology, from modules to software." The company initially targets level 2 autonomous vehicles as defined by the Society of Automotive Engineers—those equipped with advanced driver-assistance systems like Cadillac's Supercruise, NVIDIA's Drive Autopilot, and Volvo's Pilot Assist.
Recogni's roadmap progresses from level 2 to level 3 vehicles, where the driver is no longer required to monitor the environment, then to level 4 vehicles that can largely drive themselves without constant human intervention, and eventually to fully autonomous level 5 vehicles. The company predicts that by the 2024 timeframe, AI systems will make robotic taxis feasible from both cost and capability perspectives, with personal self-driving cars following within another year or two.