Imagination E-Series GPU pairs AI upscaling with faster on-device model inference
The AMW Read
E-Series meaningfully extends an established GPU IP supplier into neural graphics and LLM inference, with potential segment-level effects that depend on independent SoC results.
Imagination E-Series GPU pairs AI upscaling with faster on-device model inference
Imagination Technologies unveiled performance results for its E-Series GPU IP, which integrates a Neural Core and low-precision matrix computing into a programmable graphics architecture. The company says one E-Series core can upscale a frame from 540p to 1080p in 2.3 milliseconds. It also reports 4.7 times the previous generation's LLM prefill performance, 4.4 times its convolution performance, and up to 39% better performance per watt from changes that increase data reuse. These are company-reported results, not measurements from a shipping device.
The design puts Imagination in the AI infrastructure contest over which processor should handle AI work inside consumer, automotive and robotic systems. Its case for the GPU rests on keeping graphics and neural operations close together, reducing transfers to external memory while preserving a software path through OpenCL, Vulkan and higher-level model tools. Support for MXFP8 and MXFP4 targets language-model inference as well as image processing. The strategic question is whether that shared hardware and software stack can reduce the cost of maintaining a separate accelerator without sacrificing performance on specific tasks.
Chip designers evaluating E-Series should test complete SoC configurations, including memory bandwidth, power and software integration, against a GPU-plus-NPU design. Imagination says the GPU can take on some work previously assigned to an NPU, but also describes CPUs and other processors retaining distinct roles. Its published gains establish a useful benchmark for evaluation; they do not establish end-device latency, battery life or customer adoption.
