Steering a revolution: Optimized automated driving with heterogeneous compute

The power of heterogeneous compute

Heterogeneous compute refers to the use of different processing units — such as Central Processing Units (CPU), Neural Processing Units (NPUs), Digital Signal Processors (DSPs), and Graphics Processing Units (GPU) and computer vision accelerators — in a system to perform specific tasks more efficiently. 

Our ADAS platform harnesses the power of heterogeneous compute to unlock:

• Enhanced performance and efficiency: ADAS systems require immense computational power to process data from multiple sensors and perform complex algorithms, and efficient handling of large data movement. By leveraging the strengths of different processing units, heterogeneous compute allows for workload distribution, optimizing power consumption and improving overall system performance in ADAS applications. Moreover, advanced data compression, AI compilers and optimized memory architecture enables highly optimized data handling while minimizing offloading to DDR.  
• Efficient sensor fusion: Sensor fusion lies at the heart of ADAS systems, combining data from various sensors to create a comprehensive understanding of the environment. ADAS systems can include both early fusion using transformer based AI architectures, and late fusion depending on the safety concepts.

• Navigating urban environments: With more advanced L2+ systems, handling challenging urban scenarios require new AI planners and end to end architectures. These novel architectures can range from sensor-to-trajectory type foundational models, to combined AI models with online planning algorithms like Monte Carlo Tree Search (MCTS), and path optimization as Quadratic Programming (QP) or Dynamic Programming (DP). These diverse architectures benefit significantly from the heterogeneous compute architecture and efficient memory and data compression schemes supported by the SoCs to meet required latency and throughput.
• Enhancing the overall experience: Leveraging the ability to combine ADAS and in-vehicle infotainment (IVI) systems is another important use case where foundational models and LLM agents can benefit multiple use cases hosted on the SoC that can benefit from both the ADAS sensors and in-cabin sensors. Besides heterogeneous compute, our SoCs are designed to handle mixed criticality use cases. 

• Real-time responsiveness: In the fast-paced world of ADAS, real-time responsiveness is paramount. Heterogeneous compute allows for parallel processing of multiple tasks simultaneously, supporting faster data processing and decision-making. Our ADAS platform leverages this capability to make critical decisions in real time, enhancing the safety and reliability of ADAS systems.
• Power and thermal efficiency: ADAS systems are often deployed in resource-constrained environments, where power consumption and thermal management are critical considerations. Our heterogeneous compute solution optimizes power consumption by using the most energy-efficient processing units for specific tasks. At Qualcomm Technologies, our Qualcomm Hexagon DSP is integrated into many of our Snapdragon processors and is specifically optimized for audio, voice and image processing tasks. This not only reduces power consumption but also helps manage thermal constraints, enhancing the longevity and reliability of ADAS systems.

We invite you to download our comprehensive whitepaper, Efficient Heterogeneous Compute with Parallel Processing for ADAS, as it dives even deeper into the transformative power of heterogeneous compute for ADAS systems, providing in-depth insights, real-world use cases and technical details on how our ADAS platform leverages heterogeneous compute to revolutionize the automotive industry. 

As the automotive industry continues to evolve, ADAS systems play a pivotal role in providing safer and more efficient driving experiences. Our heterogeneous compute solution is at the forefront of this revolution, empowering ADAS systems with enhanced performance, real-time responsiveness and power efficiency.