GPU has 1.3Gray/s ray-tracing for games on phones

Called IMG DXT, and part of a planned ‘D-Series’ family, the GPUs “offers mobile device manufacturers the chance to integrate ray tracing into their SoCs, to match their design goals from premium to mainstream devices”, according to the company. The GPU scales from a “half RAC [ray acceleration cluster] configuration up to a four-RAC design, covering applications from simple shadowing in hybrid implementations to premium ray-traced graphics for AAA content”.

Development tools include Vulkan RT ray tracing support, and also supported from Vulkan is ‘fragment shading rate’ (FSR) – fragment shading draws colours from a scene onto the frame buffer.

FSR trades away some visual impact to increases graphics performance by decreasing the number of fragments processed, and “makes ray-traced effects more efficient as fewer rays are sent into the scene, meaning realistic lights and shadows are achieved within a smaller silicon area”, said Imagination.

Developers control how much FSR is used. At its highest setting, used for background in an image for example, shader code is only executed once for a 4×4 fragment zone, resulting in a near 94% reduction in workload with matching power and bandwidth benefits, the company said. Zones of 2×4, 4,2, 1×4, 4×1, 2×2, 2×1 and 1×2 are available for increasing detail, as well as the no FSR (so ‘1×1’) for full detail.

Imagination has announced three example configurations (optimised for 3nm, at 1GHz):

• IMG DXT-72-2304 ‘flagship mobile’
  50% higher maximum single-core performance compared to CXT Series and 50% boost in performance and compute due to triple USC (see text)
  2.25Tflop/s FP32
  72Gtexel/s
  648Mray/s – 1.3Gray/s tracing option
• IMG DXT-48-1536 ‘premium mobile’
  10% better performance density compared to CXT Series (20% with ray tracing)
  1.5Tclop/s FP32
  48Gtexel/s
  433 – 866Mray/s tracing option
• IMG DXT-8-256 ‘mainstream’
  256 Gflop/s FP32
  8Gtexel/s
  216 – 433Mray/s tracing option

“DXT cores offer 20% better performance density per area than the CXT GPUs, while reducing power requirements,” claimed imagination.

The 50% boost in flagship performance comes from moving from a two SPUs (shader processing units) in a ‘universal shader cluster’ (USC) to three SPUs per USC. The SPUs are blocks with an ALU (arithmetic logic unit) for computation and a TPU (texture processing unit) for pixel shading, geometry processing and rasterisation.

Another feature of DXT GPUs is 2D dual-rate texturing – adding a ‘fast path’ through the texture processing unit (TPU) for when game developers rely on heavy texture sampling post-processing to implement filter effects. A near doubling of bilinear filtering performance is claimed with reduced power consumption overall for a small area increase.

Other features include support for Khronos ASTC HDR compression, which aims to retain detail while offering greater dynamic range between bright and dark areas.

Imagination GPUs have an integrated ‘firmware processor’ responsible for high-level scheduling and priority of workloads. With DXT the company has switched the XT branch of its GPUs to the RISC-V architecture. “Our automotive ‘XS’ range of GPUs has been using RISC-V to enable functional safety, so this switch is low-risk, and benefits from RISC-V’s advances in tool flows and C-programmable functionality,” it said. “The RISC-V processor is also up to 40% faster than the previous proprietary processor, enabling even speedier event handling and more functionality.”

Beyond Desktop

DXT will also be available in what imagination calls a ‘beyond desktop’ configuration (right) that scales up to four cores and has additional IP blocks: up to eight neural network accelerators (NNAs), on-chip memory shared between NNAs and GPUs and Ethernet packet processing for multi-port multi-Gbit IEEE 802.3 switching and routing.