Multi-core IMG Series4 offers most responsive NNA, says Imagination
Claimed to be the ultimate AI accelerator delivering industry-disruptive performance for advanced driving assistance system (ADAS) and autonomous driving, the IMG Series4 is the next-generation neural network accelerator (NNA) from Imagination Technologies. It targets automotive Tier 1s, OEMS and automotive SoC manufacturers.
It is based on a new multi-core architecture, designed to run full network inferencing while meeting functional safety requirements. It executes multiple operations in a single pass to maximise performance per watt and for energy efficiency, advises the company. Series4 delivers 600 tera operations per second (TOPS), low bandwidth and low latency for large neural network workloads, says Imagination.
The multi-core architecture allows for flexible allocation and synchronisation of workloads across the cores. Imagination’s software provides fine-grained control and increases flexibility through batching, splitting and scheduling of multiple workloads. This can be exploited across any number of cores. Available configurations are two, four, six or eight cores per cluster.
Series4 offers 12.5 TOPS per core at less than 1W. For example, an eight-cluster core can deliver 100 TOPS. Imagination says that a Series4 NNA achieves performance that is over 20x faster than an embedded GPU and 1000x faster than an embedded CPU for AI inference.
By combining cores into a cluster, all the cores can be dedicated to executing a single task, reducing latency, and therefore response time, by a corresponding factor. Additionally, Imagination’s Tensor Tiling (ITT) has been introduced to the Series4. The patent-pending technology addresses bandwidth efficiency by splitting input data tensors into multiple tiles for efficient data processing. ITT exploits local data dependencies to keep intermediate data in on-chip memory. This minimises data transfers to external memory, reducing bandwidth by up to 90 per cent. ITT is a scalable algorithm with major benefits on networks with large input data sizes.
For automotive safety, the Series4 includes IP-level safety and a design process that conforms to ISO 26262 to help customers to achieve certification. Hardware safety mechanisms protect the compiled network, the execution of the network and the data-processing pipeline.
The Series4 is designed for self-driving cars and robotaxis which require new levels of artificial intelligence (AI). Series4 has already been licensed and will be available in December 2020.