Radiation-Tolerant, Scalable Neuromorphic Energy-Efficient Accelerator for Heterogeneous Processor Systems

The accelerator developed by Niobium Microsystems, Inc. (Niobium) is scalable in terms of parallelism and memory capacity, so that it can be targeted towards a variety of platforms, from small battery operated devices to large high-performance compute systems. It also has the ability to perform online learning when operating on neuromorphic workloads, drawing inspiration from the Hebbian learning paradigm. Additionally, the accelerator is designed as a memory-mapped peripheral of a larger heterogeneous System-on-Chip (SoC) and as such it can utilize external memory and implement arbitrarily sized Neural Networks (NNs) and even multiple NNs at the same time.Niobium is also prototyping several different approaches for incorporating radiation-tolerant features in the core by leveraging Niobiumrsquo;s novel digital circuit design flow, and incorporating magnetoresistive random-access memory (MRAM) where appropriate to harden the memory. As part of the Phase II effort, Niobium proposes to proceed with the implementation of the accelerator core with two additional innovations that are crucial to the NASA mission, but also have broader market potential in commercial and defense applications. Specifically, we plan to utilize Niobiumrsquo;s asynchronous circuit design techniques to (1) enable broad Dynamic Voltage Scaling (DVS) for enhanced protection against long-term radiation effects as well as potential improvements in energy efficiency, and (2) incorporate low-overhead radiation-tolerant circuits that protect against transient radiation effects, commonly referred to as Single-Event Transients (SETs), while minimizing the overhead in terms of power, performance and area. Lastly, as part of the implementation effort, Niobium intends to perform a quantitative tradeoff analysis between MRAM and conventional ECC-protected static random access memory (SRAM) with redundancy for the system-level cache of the accelerator.