Significant Performance Gains Reinforce Imagination’s GPU-First Strategy

Imagination Technologies’ first performance figures for its E-Series GPU IP turn its “GPU-first” strategy from an architectural proposition into a measurable product claim. The company is arguing that edge systems do not need separate, lightly utilised blocks for graphics, general compute and artificial intelligence. Instead, one programmable GPU architecture, supported by one software stack, can execute all three classes of workload and dynamically allocate silicon as demand changes.

Technically, E-Series extends the parallel graphics pipeline with a tightly integrated Matrix Accelerator. It supports high- and low-precision formats including BF16, FP4 and MX, which are important because edge inference typically trades numerical precision for throughput, memory efficiency and lower energy use. Unlike a fixed-function neural processor, the GPU programming model can accommodate new operators after a chip has been manufactured. That matters when model architectures evolve faster than semiconductor design cycles.

Imagination reports substantial generation-on-generation gains. Against an equivalent D-Series design, E-Series is claimed to deliver 4.7 times faster prefill for a representative edge language model, 4.4 times faster Conv2D kernels and 4.8 times faster GEMM kernels. Matrix multiplication can reach 89% GPU utilisation. On a quad-core EXD-64-2048 running at 1.5GHz, the company says a quantised Qwen 3.5 4B model achieves 0.2-second time to first token and 150-token-per-second decoding, using a 1,500-token input and output.

Those numbers address two different user experiences. Prefill performance determines how quickly a model processes its prompt, while decode throughput determines how rapidly it produces the answer. A low first-token delay makes an assistant feel responsive; sustained decoding supports richer local generation. However, these remain vendor results on specified workloads, not independent comparisons across complete systems. Memory type, thermal limits, compiler maturity and model implementation will strongly affect shipping-device performance.

E-Series also applies AI acceleration inside the rendering pipeline. Imagination’s Neural Super Resolution uses temporal data to reconstruct a 1080p image from a 540p render. Its single-pass network uses proprietary compression that removes about 65% of weights, completing the upscale in as little as 2.3 milliseconds on one E-Series core at 1GHz. The company says the technique roughly halves memory bandwidth versus native rendering while producing imagery close to ground truth. By rendering fewer source pixels, a device can redirect power and compute budget toward frame rate or visual complexity.

Conventional graphics has not been sidelined. Imagination claims up to 54% better gaming performance than the prior-generation equivalent, more than 60 frames per second in selected AAA PC titles on a quad-core configuration, and up to 39% higher performance per watt. DirectX 12 Feature Level 11_0 expands compatibility, while planned PowerVR SDK, Unreal Engine and Godot integrations reduce adoption friction for developers.

Why does this matter? Edge-chip designers face a worsening integration problem. Every accelerator adds interfaces, verification work, memory traffic, drivers and toolchains. Consolidation can reduce engineering cost and die-area waste, while a programmable resource is less likely to sit idle when workload mixes change. It may also keep sensitive data on-device, reduce cloud latency and make capable AI available where connectivity is limited.

The strategy nevertheless involves trade-offs. A unified GPU may not beat a purpose-built NPU on every operation or power envelope, and software compatibility matters as much as peak arithmetic. Imagination’s upstream Llama.cpp backend, with PyTorch and ONNX Runtime support planned, is therefore strategically important: hardware without familiar deployment paths rarely wins developer commitment.

Bottom line: The announcement is best viewed as validation, not victory. Multiple partners have licensed E-Series and first silicon is expected to tape out later in 2026, but production chips must still confirm performance, efficiency and scalability. If they do, Imagination will offer semiconductor vendors a credible alternative to increasingly fragmented heterogeneous designs—and demonstrate that the GPU can become the organising processor for edge intelligence, not merely the engine that draws the screen. That would strengthen competition in licensable GPU IP while giving device makers greater architectural choice as AI, graphics and compute steadily converge at the edge.

Contact Imagination

Also Read:

Worldwide Design IP Revenue Grew 12.4% in 2017

AI Based Software Designing AI Based Hardware – Autonomous Automotive SoC Platform

Webinar Preview: Alexa, can you help me build a better SoC?