Could there be a more affordable FPGA solution for Edge AI applications?

We need to accept that we have entered a different era now. With the global chip crisis, we started to turn to the chips available in the market, not the chips needed for designs. This also resulted in the emergence of versions of a project with more than one chip. Besides these, new solutions were being produced even in the crisis environment. We can give an example of deep learning application in FPGAs, which has become popular in recent years. In fact, when we think about it, an FPGA capable of parallel processing can be a perfect solution to neural networks, which have a highly parallel processing load. Based on this point, manufacturers have offered many solutions such as Xilinx Vitis AI, Deep Learning Applications with Vivado HLS, and TensorFlow-Lite. In addition to these, another manufacturer went its own way and offered a solution: Efinix FPGAs.

Efinix was able to stand out from other FPGAs with its shorter lead times and affordable price. But in our minds, we have a question: Can it offer a solution as good as other FPGAs? Let's answer this by reviewing the answer together.

Traditional FPGAs have dedicated routing lines. When synthesized based on these routing lines, the synthesis time may be very long and it may not even be synthesizable because routing resources are insufficient. Efinix is moving to what it calls a Quantum architecture to solve this problem. Here, logic and routing cells are no longer separate, a new structure emerges known as an XLR (eXchangeable Logic and Routing) cell. In other words, it can offer you solutions according to the synthesis needs instead of complex and dedicated blocks. As a result, it helps you achieve higher resource utilization which still meets timing.

XLR cell
Figure 1 XLR Cell - Image source 

Trion FPGAs, the first quantum series of Efinix, has already started to consolidate its place in projects. Trion FPGAs, which have example projects, especially in Vision applications, can give you the base design you need. So can Trion FPGAs be used for Deep Learning applications? Definitely yes! While Efinix is publishing the Edge Vision SoC Framework, they have not only released a design that will fit edge vision applications but also a framework that will appeal to edge AI applications.

Figure 2 Edge Vision SoC Framework - Image source

The Edge Vision SoC framework uses Quantum accelerators to facilitate hardware/software partitioning and achieve the desired performance. Within this framework, Efinix provides example designs for specific functions, such as video processing, AI object detection, machine learning, multi-camera fusion, etc.

Complete system solution example
Figure 3 Concept design for NNs - Image source

Efinix has created a RISC-V based, open infrastructure for you to develop your accelerator function without having to create everything from scratch. This accelerator socket has specific inputs and outputs to the accelerator function, RISC-V processor, DMA controller, and other processing blocks.

Figure 4 Risc-V based Accelerator design - Image Source

Efinix was not content with what it had and took itself to the next level, the Titaniums, the second quantum series. Titanium FPGAs are starting to appear with high-speed I/Os, high-speed interfaces such as PCI Express and SERDES. In addition to these fast interfaces, Titanium FPGAs become a very suitable option for Edge AI applications with highly configurable embedded memory blocks and dedicated highspeed DSP blocks.

Figure 5 Deep Learning Design Flow - modified Image Source

When the Titanium series and Quantum Acceleration come together, we can see how phasing out software bottlenecks to hardware accelerators improves overall system performance in edge AI applications. If you think it would be nice to see an application on hardware, you can take a look at the link.

Considering what I have said, I think we have found the answer to the question I asked at the beginning of my article. The answer is: Yes! Efinix FPGAs can offer as good a solution as any other FPGA. But to be honest, it may take some getting used to. Someone who hasn't used pure RTL might struggle with this IDE, or soft-core Risc-V can be challenging for you at first. Efinix takes you out of the superficial and pushes you to dig deeper. Sometimes it's good to see the actual process, it can help you make sense of the problems more efficiently. If you want to look further, you can check out the detailed documentation on the Efinix website: https://www.efinixinc.com/

Finally, I would like to say that technology is a cumulative field. I believe that we will achieve our goals more easily when we work together. To the good times,

Bahar Ateş