Imagination Adds Neural Super Resolution to E-Series GPUs

Imagination has added newly dated E-Series GPU IP pages for the EXT-8-256 and EXT-64-2048, both built around a common pitch: graphics, general compute and AI acceleration on one GPU architecture and software stack. The September 21 listings add Matrix Accelerator hardware, low-precision number formats and a Neural Super Resolution model to the capabilities Imagination is presenting to chip designers.

This is an IP announcement, not a launch of a discrete graphics card. Imagination licenses GPU designs and software technology for integration into other companies’ silicon. The product pages describe target use cases ranging from mainstream mobile and industrial devices on the EXT-8-256 to laptops, flagship mobile devices and data-center workloads on the larger EXT-64-2048; they do not announce a retail board, price, clock speed, customer chip or shipping PC.

Both EXT pages describe a Matrix Accelerator embedded alongside the arithmetic logic units and sharing the same register storage. Imagination says this arrangement provides “4x Matrix Acceleration” for compute and AI workloads. That is a vendor architecture claim, not an independently measured four-times speedup for arbitrary applications.

The supported low-precision formats span FP32 and FP16 through BF16, INT8, FP8, MXFP8, FP4 and MXFP4. Those formats matter because inference and matrix workloads can trade numerical precision for throughput, storage and bandwidth, but format support alone does not tell us how fast a finished SoC will run a particular model. Clock frequency, memory subsystem, core configuration, software and the licensed implementation all remain relevant.

Imagination says the same Matrix Accelerator can be used for neural graphics techniques including upsampling and denoising, and lists a new Neural Super Resolution model for efficient upscaling. That makes neural rendering part of the E-Series graphics story rather than limiting the matrix hardware to standalone compute or AI workloads.

The important limit is what has actually been demonstrated publicly. Imagination’s product material establishes that NSR exists as an E-Series capability, but it does not make NSR interchangeable with NVIDIA DLSS, AMD FSR or Intel XeSS. Those technologies differ in hardware assumptions, models, integration paths and image-reconstruction techniques, so a name-level comparison would overstate the available evidence.

The smaller EXT-8-256 page lists 256 FP32 FLOPs per clock, 512 FP16 FLOPs per clock and 4,096 DOT8 neural-network operations per clock. Imagination positions it for mainstream mobile, premium digital-TV interfaces, entry-level automotive HMI and industrial uses. The larger EXT-64-2048 scales those listed per-clock figures to 2,048 FP32 FLOPs, 4,096 FP16 FLOPs and 32,768 DOT8 neural-network operations, with data-center, laptop and higher-end mobile use cases.

Per-clock figures are architectural throughput specifications, not benchmark scores. Without a shipping implementation and its clock, memory configuration, power envelope and software workload, they should not be converted into a claimed real-world frame rate, TOPS figure at an assumed frequency or comparison against a retail GPU.

Imagination lists Vulkan 1.4, OpenGL 4.6 through Zink, OpenGL ES 3.x/2.0/1.1 plus extensions and OpenCL 3.0 FP for both EXT variants. Linux consumer, Linux X.org and Android are listed as supported operating-system environments. Ray tracing is optional, while HyperLane virtualization supports up to 16 virtual machines and multi-core scaling is available.

Those capabilities reinforce the “one processor, one software stack” positioning, but they should be read as GPU-IP capabilities available to licensees. A device maker still determines which licensed configuration, driver stack, operating system and exposed features reach an actual product.

The notable direction is convergence. Imagination is putting matrix-oriented AI hardware into the same licensable GPU architecture used for graphics and compute, then using that hardware for a graphics feature in Neural Super Resolution. That mirrors a broader industry move toward GPUs that mix conventional rendering with lower-precision matrix workloads, while leaving implementation choices to the silicon vendor.

For PC readers, the EXT-64-2048 laptop and data-center positioning makes the architecture relevant to future computing platforms, but the September 21 pages do not identify a laptop design win, desktop card, chip customer, release date or price. Until a licensee announces a product, the safe conclusion is that Imagination has expanded the capabilities it offers to silicon designers—not that a new consumer GPU is about to ship.