Project PHI: System Design for Pervasive Hierarchal Intelligence
This talk presents, Project PHI (Pervasive Hierarchical Intelligence) a holistic effort to provide a comprehensive solution for making immersive machine intelligence a reality.
Our guiding principle is to retain as much generality and automation while delivering large performance and efficiency gains through specialization and acceleration for a wide range of learning and intelligence workloads. As the first milestones of Project PHI, we have developed Tabla and DnnWeaver, which are open source and publically available (http://act-lab.org/artifacts/tabla/ and http://act-lab.org/artifacts/dnnweaver/). DnnWeaver is the very first open-source hardware acceleration framework for deep neural networks. Tabla is a cross-stack solution—spanning from programming language to the hardware—that rethinks the hardware/software abstraction by delving into the theory of machine learning. It leverages the insight that many learning algorithms can be solved using stochastic gradient descent that minimizes an objective function. The solver is fixed while the objective function changes with the learning algorithm. Therefore, Tabla uses stochastic optimization as the abstraction between hardware and software.
Consequently, programmers specify the learning algorithm by merely expressing the gradient of the objective function in our domain specific language. Tabla then automatically generates the synthesizable implementation of the accelerator for scale-out FPGA realization using a set of template designs. Real hardware measurements show orders of magnitude higher performance and power efficiency while the programmer only writes 60 lines of code. These encouraging results show that rethinking the hardware/software abstractions from an algorithmic perspective can open new dimensions in system design for Pervasive Hierarchical Intelligence.