Designed to be scalable and flexible the GPU IP has been developed to process graphics and compute workloads in cockpit, infotainment and advanced driver assistance systems. It includes a number of innovations in distributed safety, eliminating the overhead of achieving ASIL-B functional safety on Imagination processors. It has already been licensed for use in the automotive market.
The GPU supports the graphics and compute requirements of an entire vehicle line-up with single core configurations starting at 0.25 TFLOPS and scaling up to 1.5 TFLOPS - 50% higher peak performance than Imagination’s previous generation automotive GPU IP.
Imagination’s multi-core technology allows for two, three and four core configurations, while the low bandwidth bus between cores and support for isolation suits the use of IMG DXS in chiplets. It also included features that were introduced into the PowerVR architecture in the D-Series, such as Pipelined Data Masters and 2D Dual-Rate Texturing, to boost performance efficiency by 20% over previous generations of automotive GPU IP and ensure that it outperforms competitive cores on real world benchmarks.
Commenting Anshel Sag, Principal Analyst at Moor Insights & Strategy, said, “The demand for pixel processing in cars has grown enormously over the last decade, with digital cockpits, dashboard wide displays and rear entertainment screens offering smartphone-like visual experiences.”
DXS has been optimised for the compute workloads needed for next generation vehicle intelligence, such as computer vision and data processing. By combining the performance of DXS with an additional FP16 pipeline and a set of new compute libraries and AI toolkits, users can, according to Imagination, achieve up to ten times higher performance for compute workloads compared to previous generations.
“The GPU is no longer just used for graphics in cars,” explained James Chapman, Chief Product Officer at Imagination. “The long-term nature of designing hardware in the era of the software-defined vehicles is making hardware designers choose flexible and programmable processors that can handle the AI workloads of today as well as adapt to whatever the future brings. Today, our automotive GPUs are as much in demand for their compute abilities as for their rendering.”
All functions that a vehicle performs are graded based on potential risk according to the ISO 26262 Automotive Safety Integrity Level (ASIL). Automotive manufacturers and their suppliers must ensure that components involved in that function meet the necessary standards for fault detection and handling, with ASIL-A having the least stringent requirements and ASIL-D the most comprehensive.
Its new “Distributed Safety Mechanisms” solution means DXS achieves ASIL-B functional safety at a fraction of the Performance, Power or Area (PPA) overhead of the two main existing methods: dual-core lockstep, which increases silicon area by 100%, and workload repetition, which halves processor performance for safety workloads.
Distributed Safety Mechanisms has a near-zero impact on GPU performance and 10% area cost. It does this by taking advantage of the inherent parallelism of today’s processors and the fact that no thread is ever fully utilised.
A patented mechanism combines these threads into pairs and injects safety tests in idle moments to identify faults within the timeframe set by the ASIL standard.
Developed alongside the hardware is a new set of compute libraries (imgBLAS, imgNN, imgFFT) to help software developers achieve up to 80% GPU utilisation. This will also accelerate common compute workloads such as computer vision and pre/post processing perception data for ADAS systems.
The libraries combine with new reference toolkits (oneAPI and TVM) to create a practical software stack based on open standards with which software developers can easily port their compute applications onto Imagination IP based hardware, and from there maximise their performance.
Imagination’s automotive GPUs support OpenGL ES, Vulkan, OpenGL and OpenCL. They run popular automotive operating systems such as QNX and Green Hills Software’s INTEGRITY RTOS, as well as Linux and Android.
Hardware-based virtualisation enables Imagination GPUs to run up to eight operating systems simultaneously, with full memory isolation for fully secure GPU multitasking.
Imagination partners with CoreAVI on the development of a safety critical driver, supporting OpenGL SC and Vulkan SC for Imagination GPUs.
