Documentation

Deployments of machine learning networks with auto-regressive critical paths, or recurrence, often poorly utilize AI accelerator hardware. Such networks, like those used in automatic speech recognition (ASR), must run with low latency and deterministic tail-latency for at-scale real-time applications. In this paper, the team at Myrtle.ai presents an overlay architecture for an inference engine which is then implemented on a Speedster7t FPGA. The team further highlights the benefits of the AI-optimized Speedster7t architecture for low-latency, real-time applications.

Industry has come a long way over in the last three hundred years. Machines were first introduced in the 1700s, mainly water and steam driven, introducing the Industrial Revolution in the late 1700s. Automation and computer technology would enter the picture in the late 1960's, paving the way for the eventual automation, artificial intelligence (AI) and networked solutions of today. Although it might appear that humans are no longer in the picture, Industry 5.0 is bringing us full circle by combining the precision and efficiency of robotic systems, driven largely by AI, with the ingenuity and real-time thought of the human mind — all leading to more optimal manufacturing environments.

Achronix's Speedcore eFPGA IP allows companies to embed a programmable logic fabric in their ASICs, delivering to end users the capability to modify or upgrade the functionality of an ASIC after being deployed in the field. This flexibility dramatically expands the solution space that can be served by the ASIC as it can be updated to support changing standards and algorithms. Timing closure is particularly challenging due to the fact that the eFPGA fabric may host any number of designs over the course of device operation. Each of those designs must work independently with the rest of the ASIC, and timing closure can only be said to have been met if all of the possible designs targeting the eFPGA fabric can meet timing.

5G, 5G Advanced, and 6G bring many technical and commercial challenges that need to be met if the promised benefits of this new cellular technology are to be truly achieved. Any solution in this space must deal with the evolving specifications — FPGA and eFPGA IP technology is critical to the successful deployment of these next-generation network technologies.

The past decade has seen massive growth in centralized computing, with data processing flowing to the cloud to take advantage of low-cost dedicated data centers. It was a trend that seemed at odds with the general trend in computing — a trend that started with the mainframe but moved progressively towards ambient intelligence and the internet of things (IoT). As we move into 2018, this centralization is reaching its limit. The volume of data that will be needed to drive the next wave of applications is beginning to force a change in direction.