运用GPU + NNA 实现设备端异构计算

2018年11月 www.imgtec.com A history of Imagination

Dreamcast Focuses / Apple Over 11B Acquired by Listed in London Console strategy on invest Acquire MIPS shipment Canyon Bridge mobile

1985 1994 1995-1998 1999 2002 2006 2013 2017 2017/11

Set up: Computer First Develops MBX 1B GPU Launch of Spin off Focus on 3D GPU affordable IP mobile shipment first NNA MIPS video with NEC DAB business GPU hardware radios model shipping

© Imagination Technologies Flexible and efficient AI Solutions 2 PowerVR Vision & AI portfolio Combination of hardware, software, and ecosystem integrations

Ecosystem Open source tools & machine learning frameworks

Drivers & SDKs & Debugging & Software Low-level high-level Compilers & analysis APIs APIs tuning tools tools

Neural Hardware Network GPU Dedicated Accelerator Processors

© Imagination Technologies Flexible and efficient AI Solutions 3 PowerVR GPU Efficient compute across a wide range of markets . Flexible range of solutions . XE/XM/XMP Series: Cost efficiency focussed . Performance/mm2 . XT Series: Power efficiency focussed . Performance/mW . Highly scalable and configurable . Adapt to range of market requirements . Feature set, performance characteristics . Compute versus Graphics balance . System focus . Optimised to work with PowerVR NNA

© Imagination Technologies Flexible and efficient AI Solutions 4 PowerVR Neural Network Accelerator (NNA)

Complete neural network inferencing solution Typical SoC

Image

sensors Video

ISP Codec

. Brand new architecture

Peripherals PowerVR PowerVR 2NX

HDMI NN Accelerator

I/ O . Provides the best performance for even the smallest Fabric area budget Wireless 2NX Driver DDR Main CPU GPU Audio . Complete solution for NN acceleration Flash . Offload up to 100% of NN processing

. Optimised for low bit depth . Minimises the system bandwidth load

© Imagination Technologies Flexible and efficient AI Solutions 5 Flexible AI offerings: PowerVR NNA Fastest performance with PowerVR NNA

System Memory Bus

PowerVR NNA

• Fastest performance • Lowest cost • Supported feature set

© Imagination Technologies Flexible and efficient AI Solutions 6 Flexible AI offerings: PowerVR GPU Fully flexibility with PowerVR GPU

PowerVR GPU ALU Cluster

TA Scheduler ISP

Compute Data Master ALU Cluster

MMU/System System Level Memory Cache Interface

System Memory Bus

PowerVR GPU

• Fast performance • Low cost • Fully flexibility

© Imagination Technologies Flexible and efficient AI Solutions 7 Flexible AI offerings: PowerVR NNA + GPU Fastest Performance and fully flexibility

PowerVR GPU On Chip Memory ALU Cluster

TA Scheduler ISP

Compute On Chip Memory Data Master ALU Cluster

MMU/System System Level Memory Cache Interface

System Memory Bus

PowerVR NNA PowerVR GPU On Chip Memory

• Reduces bandwidth • Fastest performance • Fast performance • Lowers latency • Lowest cost • Low cost • Shared by other cores • Supported feature set • Fully flexibility

© Imagination Technologies Flexible and efficient AI Solutions 8 Moving beyond hardware

. AI has no open-standard API

. Popular frameworks are desktop first

. Requires rich software environment to compliment hardware

© Imagination Technologies Flexible and efficient AI Solutions 9 Software for AI

www.imgtec.com PowerVR AI Software Overview Unified solution across NNA and GPU

Neural Network Framework Offline . PowerVR AI Toolkit PowerVR AI Toolkit Extract Analyse Map Debug . Simplifies deployment workflow NN Model Optimised Weights

. Easy debugging and network analysis NN Model Weights Application . User optimisation networks possible Device On Android NN HAL IMG DNN API

. Supported APIs: IMG DNN & Android NN NNA Driver GPU Driver

. Seamless operability between GPU & NNA H/W PowerVR NNA PowerVR GPU . Straight forward development

. IMG DNN API available on GPU today

© Imagination Technologies Flexible and efficient AI Solutions 11 IMG DNN API Cross core low-level interface for easy integration for applications or SDKs

Common interface Target selectable Model importable

Simplifies Maximises Easiest development flexibility deployment

© Imagination Technologies Flexible and efficient AI Solutions 12 Android Neural Network API support Google’s brand new interface for AI acceleration

Introduced in 8.1 Extended in P Available for NNA and GPU

© Imagination Technologies Flexible and efficient AI Solutions 13 Summary Imagination provides flexible, efficient solutions for dedicated AI processing

. Complimentary range of hardware offerings . Select the best combination to match any requirements

. Highly efficient Vision and AI IP . Efficient mapping of neural network models . Highly optimised inference execution

. Single common NN API . Integrate once and utilise all our hardware

. Android supported . Perfect for mobile developers

© Imagination Technologies Flexible and efficient AI Solutions 14 Thank You!

www.imgtec.com