Imagination adds hardware ray racing to new GPU flagship

Update: August 6, 2023

Aimed at mobiles as well as automotive, cloud and desktop use, it has been dubbed IMG CXT, and includes the first outing of the company’s ‘Photon’ ray tracing architecture.

“By adding Photon hardware ray tracing, IMG CXT represents a jump in GPU IP, delivering incredible performance for games and other graphical use cases,” according to Imagination. “Photon brings desktop-quality visuals to mobile and embedded applications, and has already been licensed for multiple markets.”

Ray tracing imitates how light behaves in the real world, allowing 3D scenes to be created that are “near indistinguishable from real life”, calmed the company. “Due to its computational complexity, ray tracing has only recently come to desktop PCs and gaming consoles.”

An example CTX implementation, the CXT-48-1536 RT3 core, has three instances of Imagination’s ‘ray acceleration cluster’ (RAC in top diagram), offering up to a total of 1.3Gray/s. “This delivers photorealistic ray traced shadows, reflections, global illumination and ambient occlusion with high frame rates in a mobile power budget,” said Imagination.

RACs off-load ray tracing computation from the GPU’s shader cores. An RAC consists of: ray store, ray task scheduler and coherency gatherer, and is closely coupled to two of Imagination’s 128bit-wide ‘unified shading clusters’.

For high bandwidth, the ray store keeps ray data structures on-chip during processing, rather than have them shipped on and off chip to DRAM.

The ray task scheduler is the part that off-loads the shader clusters, using dedicated hardware to deploy and track ray workloads.

The coherency gatherer analyses rays in flight and, said Imagination: “bundles rays from across the scene into coherent groups enabling them to be processed with much greater efficiency.”

Throughput per RAC is up to 433Mray/s with up to 16Gboxtest/s. VulcanRT ray query and ray pipeline are supported.

Rasterised graphic performance is also up with CTX, with 50% more compute, texturing and geometry performance than the company’s previous GPU generation.

When scaled for use in data centres and PCs, the GPUs could generating over 7.8GRay/s and up to 9TFlop/s of FP32 (32bit floating point) for raster use, said the company, adding: “IMG CXT is suited for premium automotive human-machine interfaces. It enables the photo-realistic representation of vehicles for surround view and dense urban areas for sat-navs, as well as mixed-reality head-up displays, digital speedometers and other information displays.”