The AI Architecture Golden Age
Wow. Blogs are hard things to keep up. Three years since my last post? But I have an excuse - consulting work has been keeping me (too) busy. Meanwhile, technology has moved on, and I’m revising my product concepts.
My focus hasn’t changed - it is to produce a low-cost AI-Vision platform. At this point there are several reasonable hardware/software platforms for putting AI-Vision at the edge. I’ll be taking a look at several of them in future blogs, and giving thoughts on pros and cons.
One thing I know from past experience - many of the 30+ well-funded startups for AI hardware will fade away. The harsh reality is that very few of them will ever achieve sufficient uptake to become sustainable.
These are the candidates I’m evaluating for AI Vision at the edge:
Xilinx - Obviously. Xilinx is a mature company, its offerings that support this space have appropriate capability including modest power requirements and industrial temperature range. The new Vitis software platform may offer easier support for various hardware targets. They have been investing a lot of R&D to compete in the new AI-enabled markets. Where they tend to fall short is on price, and the learning curve for their technology is steep. I’m currently looking at evaluating an Avnet Ultra96 (http://www.zedboard.org/product/ultra96-v2-development-board) with a PIcam HD camera.
Google - Their Coral TPU technology (https://coral.ai/products) has many good attributes including price and software support maturity. It can be paired with a good many inexpensive hosts. My worry is that they lose interest and abandon their little hardware hobby before it can gain traction. Using their USB-based accelerator with a Raspberry Pi host is interesting.
Intel - The Movidius acquisition allowed them entry into the AI market and they’ve invested quite a lot to try making it successful. But it’s a pretty strange and semi-closed architecture to use directly. The recent release of the OpenCV AI Kit modules (https://www.kickstarter.com/projects/opencv/opencv-ai-kit/description) is very interesting because of its linkage to OpenCV and the APIs that seem to reduce barriers to entry. Intel has been floundering ever since ARM’s lower power allowed it to capture the majority of the mobile and embedded markets, where the high volume opportunities are. Time will tell whether they can dig themselves out of their hole.
Nvidia - They have the training market practically locked up. They are trying to leverage this fact and enter the embedded markets with their Jetson Nano (https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-nano/) offering, but I believe they are attacking things from the wrong (floating point DNN) direction for the sake of compatibility. The Nano doesn’t fare well in the all-important power requirements department. Nvidia has enough stability through its graphics, HPC and AI training markets to support this experiment, for now. We’ll see whether support continues if they fork over $40B to buy ARM.