AXIOM is a project aimed at building a hardware substrate and software layers to target cyber-physical systems (CPS) that interact with users in real-time running applications such as face recognition. The paper should be of interest to any practitioners in the area of CPS who are interested in the computing resources to support such applications with a particular focus on energy efficiency. The paper describes the designs and plans for a three-year multi-university project to build AXIOM.
The most interesting aspect of this is the design choices made for AXIOM. The authors have settled on a hardware substrate that combines a field-programmable gate array (FPGA) and low-power ARM cores into a programmable system on a chip (SoC). This is supported by experimental data that shows excellent energy efficiency on the matrix multiplication micro-benchmark, for instance. Another interesting design decision is the task-oriented approach to extracting parallelism for a single ARM+FPGA board. The rationale seems to be easy extraction of such parallelism from sequential code using the Mercurium compiler. Finally, the authors settle on a distributed shared memory (DSM) framework to distribute work across multiple boards. The rationale is that it is easier to start from symmetric multi-processor (SMP) programs and move to DSM rather than starting from a message-passing paradigm that would appear to require a lot of engineering effort for parallelizing targeted cyber-physical workloads using message passing.
AXIOM is an ambitious project. It is worth watching to see how the authors deal with some of the trickier issues that they will encounter, such as security and fault tolerance, that do not seem to have played a major role in their design decisions thus far. Overall, this is an interesting paper and a project definitely worth following for cyber-physical researchers.