If you have studied GPU accelerated machine learning project, you must be familiar with NVIDIA’s CUDA architecture. Next, you can also view the price online to see the cost of obtaining a high-performance video card that supports parallel programming for this particular brand.
But what if you could run machine learning tasks on a GPU and only use OpenGL? This is exactly what [lnstadrum] has been committed to for a period of time, because it can let a device as thin as the original raspberry PI Zero run tasks such as image classification at a much faster speed than using CPU alone. The trick is to break down your computing tasks into tasks that can be performed using OpenGL shaders, which are commonly used to push video game graphics.
An example of x2 neural network amplification.
[lnstadrum] explained that OpenGL versions over the past decade or so actually include so-called computational shaders, which are designed to run arbitrary code. Unfortunately, on motherboards such as pizero, this is not an option. Since 2007, pizero has only met the opengl for embedded systems (gles) 2.0 standard.
It is much more difficult to construct neural network in such a way to make it compatible with these more restrictive platforms, but the final results show its more interesting applications. During the test, raspberry PI zero and several older Android smartphones can run the pre trained image classification model at a fairly high speed.
This is not just a thinking experiment, [lnstatrum] has released an image processing framework called beatup. Using these concepts, you can use them now. The C + + library has java and python bindings. According to the documentation, it should run on almost everything. The framework includes a simple tool called X2, which can perform AI image amplification on all devices from the integrated graphics card of the laptop to the raspberry PI; This is a good way to examine this fascinating machine learning application.
To be honest, we are a little behind at this point because beatup was first publicly released in April this year. So far, it may not have been noticed, but we think this project has great potential, and we hope that once the news comes out, even the bottom hardware can produce impressive results, and it can see more potential.