Imagination Technologies published the first performance data and demonstrations for its E-Series GPU IP and introduced Neural Super Resolution, an AI-accelerated upscaling solution. The updates align with Imagination’s GPU-first strategy of providing chip designers with one architecture and software stack that can run graphics, compute and AI workloads.
As edge devices handle more demanding applications, chip designers have added new acceleration blocks, software stacks and data paths optimised for different workloads. This has led to fragmentation that increases integration effort and can leave fixed-function capacity unused when workload demand changes. Different edge AI processors have begun to converge on a similar feature set that balances performance and efficiency with flexibility. The GPU offers parallel acceleration, a software ecosystem, general-purpose flexibility and graphics functionality as a starting point for such architectures.
E-Series, first announced last year, integrates programmable AI acceleration alongside the GPU’s rendering pipelines. The result is one processor and one software stack that can run workloads including generative AI, neural rendering and gaming, either independently or alongside CPUs and other accelerators.
Programmable Acceleration for On-Device AI
AI models and their operations continue to evolve after a chip’s architecture is defined. The general-purpose nature of E-Series, its GPU-based programming model and use of industry-standard APIs allow developers to implement new operators throughout a device’s lifecycle. These are accelerated through the GPU’s tightly integrated Matrix Accelerator with support for high- and low-precision operations, including BF16, FP4 and MX data formats.
The processor reaches 4.7x faster prefill performance for a typical edge language model than the previous D-Series generation. For Qwen 3.5 4B, a quad-core E-Series GPU at 1.5GHz can achieve time to first token of 0.2 seconds and decode throughput of 150 tokens per second on a representative workload. An optimised backend for Llama.cpp is being upstreamed, with native PyTorch and ONNX Runtime to follow.
Compute Performance
E-Series improves performance across foundational compute operations used in AI, computer vision, signal processing and general-purpose computing. It runs Conv2D kernels 4.4x faster and GEMM kernels 4.8x faster than a D-Series equivalent. GPU utilisation levels of 89% are possible for matrix multiplication workloads.
Neural Rendering
Imagination’s Neural Super Resolution (NSR) technology applies E-Series matrix acceleration to the graphics pipeline using industry-standard extensions. The temporal upscaling solution is optimised for Imagination GPUs and combines a single-pass approach with a proprietary self-compression solution that removes approximately 65% of the network’s weights. The model completes a typical 540p to 1080p upscale operation within as little as 2.3ms latency on a single-core E-Series GPU at 1GHz.
When compared to native rendering, NSR increases frame rates while halving memory bandwidth consumption and delivers quality visually close to ground truth. NSR will be available as a library within the PowerVR SDK and through integrations for Unreal Engine and Godot.
Graphics and Gaming
Graphics remain part of E-Series. The architecture includes support for DirectX 12 FL11_0. In real-world gaming workloads, E-Series delivers performance improvements of up to 54% compared with the previous generation equivalent. The quad-core configuration can scale to over 60fps for selected AAA desktop titles. Performance per watt improves by up to 39%.
Supporting Quotes
Jon Peddie said: “The boundaries between traditional processor categories are blurring. CPUs are using vector extensions to boost AI performance, while NPUs are becoming more general purpose. As functionality converges, Imagination is betting that the GPU is best placed to succeed; after all, it has the advantages of parallelism, an established software ecosystem and unlike any other processor it can run graphics in addition to AI and compute workloads.”
Markus Mosen, CEO of Imagination Technologies, said: “Edge workloads have diversified massively in the last decade, resulting in processor fragmentation, software complexity and underutilised silicon. Imagination’s GPU roadmap is designed to address such challenges. By consolidating graphics, compute and AI functionality onto one programmable architecture, we are providing chip designers with a processor worth building on.”
Heng Zhang, General Manager, XiangDiXian Computing Technology, said: “Imagination shares our ambition of bringing high-performance rendering to a broader range of devices. They have been our partner across multiple generations, thanks to their forward-looking GPU roadmap and a pragmatic approach to GPU design that prioritises real-world performance and user experience over peak TOPS and FLOPS.”
Clay John, Technical Director at W4 Games [Godot Engine], said: “Developers usually have to make a trade-off between visual fidelity and performance. Hardware-optimised upscaling solutions shift that balance and provide more headroom to improve visual fidelity at the same or better levels of performance.”
Availability
E-Series configurations are available for applications scaling from smartphones up to high performance systems. Multiple lead partners have already licensed the technology, with first silicon expected to tape out later this year. Further details on E-Series and performance results are available at www.imaginationtech.com.
Qwen 3.5 4B dense model, w4a16, on a quad-core EXD-64-2048 at 1.5 GHz; context length of 1,500 tokens and an output length of 1,500 tokens.
Imagination Technologies Publishes E-Series GPU Performance Data and Introduces Neural Super Resolution
Imagination Technologies published the first performance data and demonstrations for its E-Series GPU IP and introduced Neural Super Resolution, an AI-accelerated upscaling solution. The updates align with Imagination’s GPU-first strategy of providing chip designers with one architecture and software stack that can run graphics, compute and AI workloads.
As edge devices handle more demanding applications, chip designers have added new acceleration blocks, software stacks and data paths optimised for different workloads. This has led to fragmentation that increases integration effort and can leave fixed-function capacity unused when workload demand changes. Different edge AI processors have begun to converge on a similar feature set that balances performance and efficiency with flexibility. The GPU offers parallel acceleration, a software ecosystem, general-purpose flexibility and graphics functionality as a starting point for such architectures.
E-Series, first announced last year, integrates programmable AI acceleration alongside the GPU’s rendering pipelines. The result is one processor and one software stack that can run workloads including generative AI, neural rendering and gaming, either independently or alongside CPUs and other accelerators.
Programmable Acceleration for On-Device AI
AI models and their operations continue to evolve after a chip’s architecture is defined. The general-purpose nature of E-Series, its GPU-based programming model and use of industry-standard APIs allow developers to implement new operators throughout a device’s lifecycle. These are accelerated through the GPU’s tightly integrated Matrix Accelerator with support for high- and low-precision operations, including BF16, FP4 and MX data formats.
The processor reaches 4.7x faster prefill performance for a typical edge language model than the previous D-Series generation. For Qwen 3.5 4B, a quad-core E-Series GPU at 1.5GHz can achieve time to first token of 0.2 seconds and decode throughput of 150 tokens per second on a representative workload. An optimised backend for Llama.cpp is being upstreamed, with native PyTorch and ONNX Runtime to follow.
Compute Performance
E-Series improves performance across foundational compute operations used in AI, computer vision, signal processing and general-purpose computing. It runs Conv2D kernels 4.4x faster and GEMM kernels 4.8x faster than a D-Series equivalent. GPU utilisation levels of 89% are possible for matrix multiplication workloads.
Neural Rendering
Imagination’s Neural Super Resolution (NSR) technology applies E-Series matrix acceleration to the graphics pipeline using industry-standard extensions. The temporal upscaling solution is optimised for Imagination GPUs and combines a single-pass approach with a proprietary self-compression solution that removes approximately 65% of the network’s weights. The model completes a typical 540p to 1080p upscale operation within as little as 2.3ms latency on a single-core E-Series GPU at 1GHz.
When compared to native rendering, NSR increases frame rates while halving memory bandwidth consumption and delivers quality visually close to ground truth. NSR will be available as a library within the PowerVR SDK and through integrations for Unreal Engine and Godot.
Graphics and Gaming
Graphics remain part of E-Series. The architecture includes support for DirectX 12 FL11_0. In real-world gaming workloads, E-Series delivers performance improvements of up to 54% compared with the previous generation equivalent. The quad-core configuration can scale to over 60fps for selected AAA desktop titles. Performance per watt improves by up to 39%.
Supporting Quotes
Jon Peddie said: “The boundaries between traditional processor categories are blurring. CPUs are using vector extensions to boost AI performance, while NPUs are becoming more general purpose. As functionality converges, Imagination is betting that the GPU is best placed to succeed; after all, it has the advantages of parallelism, an established software ecosystem and unlike any other processor it can run graphics in addition to AI and compute workloads.”
Markus Mosen, CEO of Imagination Technologies, said: “Edge workloads have diversified massively in the last decade, resulting in processor fragmentation, software complexity and underutilised silicon. Imagination’s GPU roadmap is designed to address such challenges. By consolidating graphics, compute and AI functionality onto one programmable architecture, we are providing chip designers with a processor worth building on.”
Heng Zhang, General Manager, XiangDiXian Computing Technology, said: “Imagination shares our ambition of bringing high-performance rendering to a broader range of devices. They have been our partner across multiple generations, thanks to their forward-looking GPU roadmap and a pragmatic approach to GPU design that prioritises real-world performance and user experience over peak TOPS and FLOPS.”
Clay John, Technical Director at W4 Games [Godot Engine], said: “Developers usually have to make a trade-off between visual fidelity and performance. Hardware-optimised upscaling solutions shift that balance and provide more headroom to improve visual fidelity at the same or better levels of performance.”
Availability
E-Series configurations are available for applications scaling from smartphones up to high performance systems. Multiple lead partners have already licensed the technology, with first silicon expected to tape out later this year. Further details on E-Series and performance results are available at www.imaginationtech.com.
Qwen 3.5 4B dense model, w4a16, on a quad-core EXD-64-2048 at 1.5 GHz; context length of 1,500 tokens and an output length of 1,500 tokens.
