Imagination Technologies has released RTL for its newest GPU, the DXTP. Compared with a similarly configured older DXT design, the DXTP raises power efficiency by approximately 17% according to the company. Imagination also touts the new GPU’s suitability for computing (e.g., AI processing), supporting the DXTP with OpenCL libraries and the OneAPI and LiteRT (nee TensorFlow Lite) AI frameworks. Two customers have already selected the DXTP, one (presumably Zelos) for an automotive design and another for a mobile device.
Three Groups of Two Versus Two Groups of Three
The DXTP improves efficiency by raising performance more than power. Gains come from increasing cache size and SoC-bus bandwidth. Imagination also reorganized its main GPU building block, the scalable processing unit (SPU). A DXTP SPU integrates two texture units and two shader cores, whereas the DXT SPU had three of each. (The new GPU now resembles the DXT’s predecessor, the CXT, which had two texture and shader blocks per SPU.)
To achieve the same raw performance as a two-SPU DXT, the DXTP, therefore, requires three SPUs. Having 50% more SPUs transacting with the new memory hierarchy, the DXTP raises sustained throughput. Further improving real-world throughput, Imagination redid how the new GPU sets up operations, accelerating it by 16×. The speedup will be most noticeable on smaller computational or graphics tasks where setup previously dominated execution time.
Imagination offers two DXTP configurations, and the company’s initial customers each chose a different one. Integrating three SPUs, the mellifluously named DXTP 48-1536 delivers a peak throughput of 48 GPixels/second and 1.536 FP32 TFLOPS, 3.072 FP16 TFLOPS, or 6 INT8 TOPS. The four-SPU DXTP 64-2048 has proportionally higher peak throughput. Neither config has ray-tracing hardware, but customers can request it. The first chips employing DXTP could ship at the end of next year.
Capitalizing on AI Mania
Nvidia GPUs dominate data-center AI; through software support, Imagination is laying the foundation for GPU-based AI processing in licensees’ designs for edge devices. Although these designs may also integrate a dedicated AI accelerator (NPU), an Imagination GPU provides additional capability and a programming model similar to Nvidia’s. For example, some edge-AI NPUs aren’t customer-programmable and don’t support floating-point data.
Like other D-series GPUs, the DXTP can perform multiple graphics and computing tasks simultaneously. Each SPU can juggle tasks in parallel and can host eight virtual environments. Alternatively, the GPUs support preemptive multitasking on a single SPU and pinning an SPU to a process. Multitasking and virtualization have various applications and should help developers use GPU resources for AI functions without blocking graphics processing.
The hole in Imagination’s AI story, however, is performance estimates for well-known models. Imperfect like the Dhrystone CPU benchmark, ResNet-50 is the standard for convolutional neural networks (CNNs). No model has such status for transformer networks, but tokens per second and time to first token for a model like Llama 8B are common metrics.
Bottom Line
Imagination is among the oldest suppliers of semiconductor designs (IP) and one of the few offering GPUs. In recent years, the company has offered safety-certifiable GPUs for the auto market and DirectX 11 GPUs for cloud gaming. The DXTP doesn’t have such a market-defining feature but instead is a general Linux/Android GPU providing a generational performance uplift. Imagination once dominated handset graphics, and landing a mobile-device customer for the DXTP could mark the company’s resurgence.