EXT-64-2048 GPU

E-Series GPU IP features a deeply integrated Matrix Accelerator and support for low-precision number formats, ensuring fast and flexible acceleration for cutting-edge gaming and AI workloads.

Imagination E-Series GPU IP unifies graphics, compute and AI on one processor, with one software stack.

Matrix acceleration units deep inside the GPU’s shading clusters deliver efficient acceleration for popular but compute-heavy edge workloads like computer vision, generative AI and neural rendering.

The cost of developing advanced chips is enormous. Every mm² of silicon has to deliver value throughout the project’s lifecycle.

A general-purpose, GPU-based accelerator that can render a user interface, support an overloaded NPU and be trusted to run new models as they appear adds considerable value to a modern, edge AI system.

The Architecture of an E-Series Core

Understand how one GPU IP can handle graphics, compute and AI tasks. The single-core EXD-64-2048 GPU shown on the left delivers 64GTexel/s, 2 TFLOPS FP32 and 32 TOPS INT8 @ 1GHz.

• RISC-V Firmware Processor
  • The firmware processor is in charge of scheduling workloads across the GPU, communicating with third party accelerators and handling errors. It operates independently of the CPU delivering superior performance.
• Unified Shading Cluster (USC)
  • The USC is the programmable compute engine at the heart of every Imagination GPU. Inside each USC, an ultra-wide 128-lane ALU delivers massive parallel processing throughput, enabling efficient execution of graphics shaders, neural networks and general-purpose compute workloads on the same architecture.
• Matrix Accelerator
  • Matmul acceleration with support for high and low precision number formats (FP32, FP16, BF16, INT8, FP8, MXFP8, FP4, MXFP4) is embedded deeply within the Imagination GPU’s ALU’s, sharing the same register storage for maximum data efficiency and power consumption.
• Multi-core Scaling
  • Customers can scale E-Series up to cloud levels of performance using Imagination’s  decentralised, chiplet-friendly multi-core technology.
  • A quad-core E-Series GPU running at 1.6GHz reaches nearly 200 TOPS INT8.

Headline features

• Matrix Accelerator
• Burst Processors
• Low precision number format support
• Neural Super Resolution
• Compute SDK
• Up to 16 VMs using HyperLane
• Multi-core scaling available
• Ray tracing optional

API Support

• Vulkan 1.4
• OpenGL 4.6 via Zink
• OpenGL ES 3.x/2.0/1.1 + Extensions
• OpenCL 3.0 FP

OS Support

• Linux Consumer
• Linux X.org
• Android

Bus Interface

• AXI
• ACE-Lite

Compression

• PVRIC5 Lossless and Lossy Framebuffer Compression
• ETC and ASTC LDR and HDR Support

Performance

• 2,048 FP32 FLOPs/Clock
• 4,096 FP16 FLOPs/Clock
• 32,768 DOT8 NN Ops/Clock

Pixels per Clock

• 64 ppc Bilinear 3D Textured
• Up to 128 ppc 2D Dual Rate Mode

4x Matrix Acceleration

Significant performance gains for compute and AI workloads using new hardware that is embedded alongside the arithmetic logic units, sharing the same register storage.

Low Precision Number Format Support

Capable of running edge-friendly AI models using FP32, FP16, BF16, INT8, FP8, MXFP8, FP4, MXFP4 data formats.

Power Efficiency Boost

The new Burst Processors pipeline guarantees the execution of blocks of operations and stores intermediary data inside the local registers, reducing unnecessary data traffic and boosting battery life.

Neural Rendering

The new Matrix Accelerator can be applied to graphics workloads in support of popular neural techniques like upsampling and denoising. A new Neural Super Resolution model is available for efficient upscaling.