MulticoreWare and Imagination have collaborated to enable GPU compute on TI’s TDA4VM processor, resulting in significant performance improvements for ADAS systems.
The partnership has successfully demonstrated over 100x performance gains by running a stereo block matching (StereoBM) algorithm on the GPU instead of the CPU. This was achieved on a high-resolution (3200x2000) image, showcasing the potential of GPU compute.
With this development, automotive customers can now utilize IMG BXS GPUs with OpenCL to maximize performance per watt on TDA4 SoCs for sensor processing workloads such as camera, radar, and lidar. This adds to the existing compute accelerators available on the platform.
"The automotive software revolution is driving the need for intelligent features on low-power SoCs," says Vish Rajalingam, representative from MulticoreWare. "Our experience with OpenCL and Edge AI across various AI accelerators, GPUs, and SoCs opens up incredible possibilities for the future of mobility."
MulticoreWare has optimized the StereoBM algorithm to efficiently leverage GPU cores. The use of IMG BXS-4-64 GPU intrinsics and adaptive memory handling has played a crucial role in achieving optimal performance, especially when dealing with higher resolution camera data.
"The graphics processor in a car is a flexible, programmable, and high-performance option for supporting a vehicle's ADAS workloads," explains Jake Kochnowicz, Vice President at Imagination.