» Current | 2019 | 2018 | 2017 | Subscribe

Linley Newsletter

Imagination Revs Its AI Engines

December 18, 2018

Author: Mike Demler

Imagination’s new PowerVR Series3NX deep-learning accelerators (DLAs) offer a broader performance range than the previous 2NX generation, enabling designers to scale their inference engines to 160 trillion operations per second (TOPS) in multicore configurations. The 3NX lineup also includes a new programmable neural-network processor core, which the company derived from its Rogue GPU architecture. In this 3NX-F configuration, the GPU serves as a programmable general-purpose (GPGPU) compute engine, complementing the logic used for multiply-accumulate and other neural-network operations.

The company also announced updates to its PowerVR Series9 GPUs. The entry-level Series9XEP and midrange XMP include enhancements relative to the first-generation Series9 products the company introduced last year. The Series9XTP adds a new high-performance GPU to the Series9 lineup, offering an upgrade from the previously top-of-the-line Series8XT.

The Series9 cores offer incremental design optimizations, but the area-efficiency savings are substantial. Mobile-processor developers targeting low-cost and midrange phones will find the gains in performance per square millimeter attractive. PowerVR GPUs have recently won designs at companies such as MediaTek, helping Imagination reestablish this product family as a strong competitor to Arm’s Mali lineup.

The smaller 3NX configurations are likely to be more attractive to designers than the supersize multicore models. To convert neural networks for embedded applications, developers typically quantize the weights to 8-bit integers, but some researchers have demonstrated models using parameters with as little as 1 bit. The PowerVR DLAs offer greater flexibility than competing IP, so designers can choose the optimum mix of weights in increments from 4- to 8-bit precision. Along with the new weight-compression technology, the variable-precision feature will yield smaller DLAs that consume less area and power than fixed-precision designs.

Subscribers can view the full article in the Microprocessor Report.

Subscribe to the Microprocessor Report and always get the full story!

Purchase the full article

Events

Linley Fall Processor Conference 2019
October 23 - 24, 2019
Hyatt Regency, Santa Clara, CA
Linley Spring Processor Conference 2019
Held April 10 - 11, 2019
Proceedings available
More Events »

Newsletter

Linley Newsletter
Analysis of new developments in microprocessors and other semiconductor products
Subscribe to our Newsletter »