First announced in 2022, Imagination’s D-Series has witnessed the popularisation of GenerativeAI, first in the cloud and then in consumer devices, like smartphones. In the cloud, the GPU has become the de facto accelerator for AI thanks to its programmability, scalability and ability to process AI workloads quickly. This preference is expected to extend to consumer devices as well. The focus for emerging GPUs has therefore extended to focus not only on their ability to deliver gaming and advanced graphics effects, like ray tracing, but also on their ability to operate as an effective AI accelerator.
To achieve ever more demanding use cases and computing requirements, SoCs in modern consumer devices often feature a range of heterogenous computing processors – DSPs, NPUs, CPUs – that can also execute AI workloads (some with greater programmability or performance than others). Out of all those processors, only the GPU can handle graphics and AI with a performance, power and area budget that’s competitive. If a chip designer can simplify their system by running AI tasks on a processor, they were going to need anyway to run the user interface, it proves a highly optimized and efficient product.
Pixel processing will always be a priority for Imagination – this is a fundamental of what GPUs do and why they are chosen to go into chips. However, in D-Series and beyond we are rapidly accelerating the compute-focused features in our GPUs to enable system designers to take advantage of their clear advantages in AI acceleration.
D-Series consists of four products, fifteen off-the-shelf configurations and a range of market-leading innovations, starting with the concept of scalable ray tracing for mobile devices in DXT. With DXT, hardware designers can still deliver great ray traced effects but now at half the area that Imagination’s C-Series GPUs required. It’s a substantial evolution of the Photon architecture, still the most sophisticated ray tracing solution on the market, sitting at Level 4 on the Ray Tracing Levels System. On top of its ray tracing capabilities, DXT boosts performance density by 20% over its predecessor and focuses heavily on performance efficiency, introducing a set of features that would also be adopted by subsequent D-Series products, including DXTP, such as a new and improved firmware processor, 2D Dual-Rate Texturing, Pipelined Data Masters and Fragment Shading Rate. DXT is successfully licensed for use in the mobile market.
Our desktop and data centre customers received our first GPU IP designed specifically for them, the high performance DXD, at the end of 2023. Reaching up to 18 TOPS INT8 when clocked at 1.5GHz, DXD extends Imagination’s API coverage to include DirectX for the first time (specifically for D-Series, FL11_0). It has been licensed by ambitious companies who are creating their own graphics cards for workplace productivity and cloud gaming. In this second market, HyperLane, our hardware-based virtualisation technology that enables up to eight operating systems to run, fully securely, on a single GPU, comes into its own. It means that when deployed in a server, each DXD GPU can support eight gamers at a time.
Automotive is one of Imagination’s most significant markets and one where we have the leading share when it comes to GPU IP for in-vehicle infotainment. As cockpit and infotainment screens grow in size, number and complexity, more powerful GPUs are needed to process all the extra pixels; and at the same time these GPUs need to conform to automotive safety standards when deployed in support of safety-critical functions, like displaying speed information to the driver. However, adhering to these standards has traditionally come with a 2xPPA penalty.
Imagination’s D-Series GPU for the automotive market, DXS, removes the PPA penalty associated with achieving ASIL-B functional safety. Its innovation in distributed functional safety, including the introduction of the patented safety pairs concept, means that companies no longer need to lay down two cores, or run workloads twice, or rely on expensive software test libraries to check that the processing logic is operating correctly.
Furthermore, as car autonomy increases, more companies are deploying Imagination GPUs as programmable AI accelerators within the vehicle’s autonomy system. To support this, DXS ships with extra features that speed up compute workloads and a new set of highly optimised OpenCL compute libraries that boost realised GPU performance up to 4x compared to open-source equivalents. Taking all these optimisations together, DXS offers up to ten times greater compute performance than Imagination’s previous offering into the automotive market (up to twenty times for safety-critical compute workloads).
DXTP combines the compute performance and efficiency improvements delivered in DXS, the graphics performance and efficiency features introduced with DXT and additional microarchitectural adjustments that together boost energy efficiency by up to 20% (vs DXT) and enable DXTP to tackle more complex game scenes and increase its stamina (i.e. its ability to sustain performance for longer).
Find out more about what’s new with DXTP in this blog.
It is the ultimate D-Series product, ideal for delivering advanced graphics and compute acceleration into smartphones and other power-constrained devices, and it is fully backed up by Imagination’s software solutions and feature-rich developer tooling. It can handle a large language model as easily as an intricate game scene, and to even do both at the same time.
This flexibility is what now separates the Imagination GPU from other accelerators. DXTP is a single solution, with an extensive software ecosystem, that can be programmed to handle a huge variety of workloads, from user interfaces to neural networks. It is a hard-working piece of silicon whose parallelism ensures great performance and whose tile-based computing pipeline ensures efficiency. As the D-Series portfolio is fully populated, and we look to the future, Imagination will continue to enable customers with performance efficient GPUs that deliver pixel-perfect graphics and advanced intelligence to all devices, from smartwatches right up to servers.