BIDMach: faster machine learning on a single machine

I got this from my colleague Yucheng Low. BIDMach is a new open source project (BSD3 license) which speeds up basic machine learning algorithms using GPU.

Some impressive performance numbers on Github. For example:

Criteo Dataset

Criteo released a medium-sized (12 GB) dataset with a single target (click, no click) with a very sparse set of features. This is representative of many click prediction tasks in industry.

System	Nodes/Cores	npasses	AUC	Time	Cost	Energy(KJ)
Spark	8/32	10	0.62	964s	$0.64	1500
Spark	32/128	10	0.62	400s	$1.00	2500
BIDMach	1	1	0.66	81s	$0.01	6
BIDMach	1	10	0.72	805s	$0.13	60