Qualcomm’s Zeroth AI API is finally here in SDK form for developers to hack at. If you recall SemiAccurate’s first look at Zeroth, we said it would happen like this and it just did.
In the beginning Zeroth was often confused with the initial demos on the platform, scene detect and handwriting recognition. If you talked to the Qualcomm personnel at the shows, it was quite clear that these demos were built using Zeroth, not Zeroth itself. Technically called Qualcomm Zeroth Machine Intelligence Platform, it is a set of APIs that have a portion of their functionality running on the mobile devices you know and love.
The idea is simple enough, train your neural net/AI/algorithm-name-of-the-moment on a cluster in the usual way and package up the ‘learning’ in a discrete data file. This data set is distributed with your app to the mobile devices where the realtime portion runs on the Zeroth engine in a smartphone or tablet. In short the heavy lifting is not done in your hand, just the immediate task at hand, it is a great way to split the job up to optimize both accuracy of learning and energy use in the field.
Since Zeroth is an SDK and a generalized toolkit, users can train their program however they want using most popular deep learning models. Qualcomm lists Caffe and CudaConvNet but says it is agnostic and can support nearly any framework that makes sense. The underlying engine is also now tuned to support the various hardware blocks in the 820, CPU, GPU, DSP, and others, so computation can be done on the most efficient hardware for the task at hand.
Interestingly Zeroth does not force you to use specific hardware so if you want to augment your program with Zeroth capabilities but it needs the GPU for example, you could run the AI portion on another less optimal unit instead. This is why it is an SDK and tools rather than an overarching monolithic program, developers can use it in the most optimal way for their program even if it isn’t the best way forward from Qualcomm’s point of view.
This hardware affinity was demonstrated to SemiAccurate last November running on the 820, the previous demos were all running in software on previous Snapdragon SoCs. Even in software form the performance was adequate, it will be interesting to see how much faster it is once optimized. More importantly it should be vastly more power efficient in this form so using it on a device won’t kill battery life.
If you are a Qualcomm licensee you can get the SDK, tools, and engine for Zeroth now. If you are not a licensee and want the SDK, tools, and engine now, you probably should become a licensee, either that or wait until they open it up to the public in the near-ish future. Given the power and potential efficiency of Zeroth, it could do a lot to bring AI and machine learning to the mobile space in a way that actually benefits users. Time to get coding people.S|A