DigiKey-emag- Edge AI&ML-Vol-10

Why and how to use Efinix FPGAs for AI/ML imaging – Part 1: getting started

TSMC 16 nanometer (nm) node.

Figure 1. What makes an XLR block unique is that it can be configured to function as either a logic cell with an LUT, a register and adder, or a routing matrix. Image source: Efinix

Both offerings are based around the Quantum architecture, which is unique in the FPGA world. The standard FPGA architecture is based on CLBs which, at the simplest level, contain a look-up table (LUT) and flip-flops. The CLBs implement logic equations that are then interconnected via routing. Efinix’s Quantum architecture moves away from distinct logic and routing blocks with the XLR block. What makes an XLR block unique is that it can be configured to function as a logic cell with an LUT, a register and adder, or a routing matrix. This approach offers a finer-grained architecture that provides routing flexibility, enabling implementations that are logic heavy or routing heavy to achieve the desired performance.

Written by: Adam Taylor

Editor’s Note: New approaches to FPGA architectures bring finer-grained control and greater flexibility to address the needs of machine learning (ML) and artificial intelligence (AI). Part 1 of this two-part series introduces one such architecture from Efinix and how to get started with it using a development board. Part 2 discusses interfacing the development board to external devices and peripherals, such as a camera. FPGAs play a critical role in many applications, from industrial control and safety to robotics, aerospace, and automotive. Thanks to the flexible nature of the programmable logic core and their wide interfacing capabilities, one growing use case for FPGAs is in image processing when ML inference is to be deployed. FPGAs are ideal for implementing solutions that have several high-speed camera interfaces. In addition, FPGAs also enable the implementation of dedicated processing pipelines in the logic, thereby removing

Efinix FPGA devices

Efinix currently offers two device ranges. It initially introduced the Trion family, which offers logic densities from 4000 (4K) to 120K logic elements (LEs), and is fabricated using an SMIC 40LL process. The newest line of devices, the Titanium family, offers logic densities from 35K to 1 million (1M) logic elements, and is fabricated on the very popular

bottlenecks that would be associated with CPU or GPU-based solutions. For many developers, however, their applications require more ML/ AI functionality and finer-grained control or routing and logic, beyond what classic FPGA architectures with combinatorial logic blocks (CLBs) can provide. Newer approaches to FPGA architectures address these issues. For example, Efinix’s Quantum architecture uses an eXchangeable Logic and Routing (XLR) block. This article discusses key features and attributes of the Efinix FPGA architecture, emphasizing its AI/ ML capabilities and introducing real-world implementations. It then discusses a development board and associated tools that developers can use to quickly get started on their next AI/ML imaging design.

Figure 2. The Titanium FPGA Ti180 comes in a variety of options depending on the bus width, I/O, and memory requirements. Image source: Efinix

we get technical

44

45

Powered by