About two years ago, Google announced the availability of TensorFlow processing units in its cloud.
They are dedicated microcontrollers built for training and running Machine Learning models. TPU are available within Gcloud as an execution platform for ML (of course, optimized for TensorFlow).
During the summer, they unveiled the edge equivalent of these TPU, which are named… Edge-TPU 🙂
These are very specific ASIC designed to execute ML models on an edge device, i.e. a small device close to the sensors gathering the data. This allows for a fast decision, without the need to send a truckload of data back up to the cloud.
But wait for it… Microsoft did just uncover a device called DataBox Edge. I know, the main purpose of this device is to provide a storage gateway to help you use Azure storage locally, and move the data between the device and Azure, hence the name. Bear with me, the path is a bit convoluted, and I would like you to enjoy every turn of it.
Databox Edge is also equipped with what has been called IoT Edge. This nifty piece of technology will enable you to run Azure-based workloads on an edge device, such as Azure Functions, Azure ML, Azure Stream Analytics etc. IoT Edge has been out in the open for about a year now, to be deployed onto compatible devices.
And, and that’s where we hit the Edge-TPU spot, also included in Databox Edge is a shiny new Microsoft hardware, called Brainwave. The name kind of gives away the purpose, especially after I guided you through the maze. Anyway, this chip is designed to run AI models on an edge device, and do it with impressive performance and efficiency.
I know, at this point, you would point out at the fact that it might again be a case of “We did it first!” from Google.
I’d like to focus a big difference between the two approaches. For once, I could not say which would win in the long term. In theory I prefer the approach from Microsoft, but that does not mean it will prevail (or that they would not change tactics and build something more like Edge-TPU).
The difference is that Google built an ASIC, whereas Microsoft used Intel FPGA to deploy its Brainwave architecture.
OK, this needs some explaining. First the names :
ASIC means Application Specific Integrated Circuit.
FPGA means Field Programmable Gate Array.
You see where this is going?
An ASIC is a very specific chip, designed to do only one thing, but optimized to its core. I should be able to execute one kind of job, but do it perfectly.
One the other hand, an FPGA is reprogrammable after its deployment, to be able to adapt to future needs. Its performance is close to an ASIC, but not quite equal.
To complete the panorama, going from specific to general use, we would then add GPU (Graphical Processing Units, as in your graphics cards) and then CPUs (ye good ol’ Pentium).
Microsoft took the path of versatility, whereas Google focused on a particular use.
As I mentioned, I’m not sure who has the best strategy, and whether there will even be a fight, but I am very curious to see both chips in the wild!