Imagination Technologies Reimagines GPUs With Ray Tracing
Imagination Technologies has revealed a ray tracing GPU with high power efficiency and low cost.
Imagination Technologies recently launched IMG DXT, a scalable ray tracing graphics processing unit (GPU) said to unlock cutting-edge graphics for mobile devices. Ray tracing is a technology that creates realistic 3D scenes where light mimics how it behaves in the real world. This technology, however, is traditionally computationally expensive because a processor must track all the rays emitted from a light source and calculate how each ray will interact with a surface or object in a scene—whether it will be absorbed, scattered, reflected, or refracted.
High-level overview of the IMG DXT, including unified shading clusters (USCs), texture processing units (TPUs), raster/geometry blocks, and top-level components.
Now, with Imagination's ray acceleration cluster (RAC) technology in the new GPU, mobile phone manufacturers can integrate ray tracing into their system on chips (SoCs). Imagination claims its IMG DXT delivers high performance and power efficiency while being affordable with up to 40% lower area cost. Below are some of Imagination's technologies found in the IMG DXT GPU.
Traditional rendering uses rasterization, a process in which a structure is built using a mesh of triangles that are then shaded to resemble the structure's appearance. With this approach, the real world can only be approximated. Ray tracing replicates, on the other hand, how low light behaves in the real world.
Photons from a light source bounce around the surroundings and reach the viewer's eyes. As the light interacts with objects, it is obstructed, reflected, and refracted, creating realistic shadows and reflections.
Compared to rasterization, ray tracing uses a single algorithm that can create effects without tools like shadow maps or other lighting passes. However, ray tracing is more complex and requires much more effort to use efficiently. Imagination first demonstrated its ray tracing architecture in 2012.
Imagination's ray tracing architecture uses traditional rasterization for most of the scene and ray tracing for the reflections and shadows, greatly reducing the required computational performance and bandwidth.
This architecture uses rasterization for most of the scenes and uses ray tracing for reflections and shadows, thereby reducing the bandwidth and performance required. Additionally, it uses backward ray tracing, reversing the concept of rays originating from a light source bouncing around the objects and finally reaching the eyes. Here, rays emit from the viewer or camera into the scene. If the light from certain points does not reach the source, it doesn't need to be processed, further simplifying the calculations.
Fragment Shading Rate
The new GPU can scale from an area-efficient half-RAC configuration to a quad-RAC design to cover various applications, from simple shadowing in hybrid implementations to premium array-traced graphics for high-end games.
Thanks to something called fragment shading rate (FSR), the new IMG DXT frees performance headroom for developers. FSR decreases the number of fragments processed and increases graphics performance with negligible impact on visual quality. The FSR technique trades image quality for improved performance and reduced power consumption. Rather than using one shade per pixel, the program allows a shade over a region of multiple pixels, lowering the bandwidth and energy requirement while delivering high performance.
The concept of FSR.
This technique works very well with ray tracing because fewer rays need to be used, resulting in less processing over a large zone of pixels.
Texture Processing Unit
Beyond graphics improvement, the Imagination team found that many modern games spend significant time on post-processing algorithms for blur, depth, and additional effects. Most of these algorithms are bottlenecked by the throughput rate of the texture processing unit (TPU).
IMG DXT's TPU creates a so-called "fast path" for these post-processing effects. The TPU doubles the performance only when the hardware detects regular processing across a region. These regions reuse processing points with texture cache and 2D sampling of a single texture with no detail or perspective. It accelerates tasks such as post-processing effects in games and computational photography filter processing. It improves image quality with faster and more power-efficient noise-eliminating algorithms.
With all these new technologies, IMG DXT cores exhibit 20% better performance density per area than Imagination's last generation of GPUs, the IMG CXT, while reducing power consumption. The DXT-72-2304 configuration delivers 72 GTexels/s and 2.5 TFLOPs of FP32 arithmetic performance, with a 50% higher maximum single-core performance than the CXT GPUs.
Shader Processing Unit (SPU)
IMG DXT features a new shader processing unit (SPU) design that consists of an arithmetic logic unit (ALU) for computational tasks and a texture processing unit (TPU) for shading pixels, geometry processing, and rasterization logic. With its new triple universal shader cluster (USC) design (3x ALU/TPU units), the GPU could deliver 50% more top-end compute and texture (ALU/TPU) performance per SPU.
The IMG DXT includes a new, higher-density SPU design with three USC/TPU blocks.
Besides hardware, the GPU has many developer tools to support ray-tracing capabilities across the stack. Moreover, it supports many compression technologies to maximize bandwidth efficiency, including Khronos ASTC HDR, which allows texture to have a high dynamic range between bright and dark areas while maintaining a high level of detail.
All images courtesy of Imagination Technologies