Intriguing. This looks like a good summary: Chapter 2 and 3 presents two key design ideas that guides the overall development of BIDMach: codesign, which means

Intriguing. This looks like a good summary: Chapter 2 and 3 presents two key design ideas that guides the overall development of BIDMach: codesign, which means choosing the right combination of hardware and designing algorithms to best leverage the hardware, and roofline design, which means quantifying the theoretical performance limits of each component of a ML algorithm and making sure to get close to them. When implementing a algorithm, we need its hardware mappings. This includes computation pattern required by the algorithm, memory/IO pattern and communication pattern when designing a distributed algorithm. To reach performance limits, careful and often iterative design and coding is needed. This is time-consuming. It would be problematic to do such optimization for every machine learning algorithm. Instead, we create an intermediate layer - a set of common computation and communication kernels/primitives, between BIDMach and the hardware. This includes BIDMat, butter y mixing and Kylix. These deal respectively with matrix algebra, model synchronization across the network, and algorithms on graphs. An important role of BIDMat is to manage memory with matrix caching, which we will describe shortly. So it's a really well-engineered approach Some bold claims: We can bring the performance of sample-based Bayesian inference up close to symbolic methods "Symbolic" here seems to refer to expectation propagation algorithms such as http://research.microsoft.com/en-us/projects/infernet/ http://research.microsoft.com/en-us/projects/infernet/