DARPA’s MOCHA Program Aims to Revolutionize Software Compilers for Heterogeneous Computing
The U.S. Defense Advanced Research Projects Agency (DARPA) is calling on the computer industry to create innovative software compilers designed to manage the complexities of heterogeneous computing environments, aiming to enhance performance and efficiency. This initiative, detailed in DARPA’s broad agency announcement (HR001124S0035), is known as the Machine Learning and Optimization-guided Compilers for Heterogeneous Architectures (MOCHA) program.
The MOCHA project is focused on developing a new generation of compilers optimized for a range of computational components, including central processing units (CPUs), graphics processing units (GPUs), and various accelerators. These advanced compilers are intended to address the challenges posed by modern heterogeneous architectures, which blend different types of processors and accelerators.
Traditionally, enhancing hardware performance relied on increasing microprocessor clock speeds. However, this approach hit limitations in the mid-2000s, leading to the adoption of multithreading techniques. Despite this shift, multithreading also faces constraints, prompting the use of specialized co-processors and accelerators to boost performance. Unfortunately, conventional compilers struggle to efficiently generate machine code for these diverse hardware setups.
