A novel data processing accelerator intellectual property (IP) Radiation-Hardened-by-Design (RHBD) core for use in NASA future missions is proposed. The core is an artificial neural network accelerator based upon work done at Google, IBM, and others. The IP core is known as a tensor core and follows an architecture of matrix multipliers, accumulators, register files, and fast and abundant memory access. The tensor core will be developed to be Advanced Microcontroller Bus Architecture (AMBA) bus compliant and will feature an architectural approach to easily expand the data processing elements when more die area is available. The core will be developed on the trusted Global Foundries (GF) 32nm Silicon on Insulator (SOI) process. There is extensive development currently occurring at this process technology, including NASAs future High Performance Spaceflight Computing (HPSC) platform. The core is proposed as an effort to develop a data processing acceleration to decrease the down-link data bandwidth of future space missions. If more processing can be accomplished in situ, a given mission can be expected to require less data bandwidth, a problem that is becoming more critical with the ever increasing number of active missions. The IP core will be developed to be incorporated into other development at the 32nm process. The IP core will also be structured in such a way as to be incorporated into Micro-RDCs future Reticle Programmable System on Chip (RPSoC) platform. The RPSoC is an active future platform, under development with funding from NASA and the Air Force, for digital and mixed signal designs to lower the cost of development at 32nm and to decrease lead-time from design inception to product delivery. The tensor core will be featured on this platform as a data acceleration core. The core will have RHBD techniques throughout the FEOL and BEOL to ensure that no data will be corrupted within the artificial neural network configuration or the data path. Potential NASA Applications NASA will directly benefit from this SBIR effort by creating a RHBD intellectual property tensor core that can be integrated into designs at 32nm Silicon on Insulator (SOI). The IP core will include Advanced Microcontroller Bus Architecture (AMBA) bus compliance to provide the simplest possible incorporation of the IP core for NASA high data throughput requirements where an artificial neural network could aid in decreasing the total data overhead in deep space multiple data transmissions. Potential Non-NASA Applications Reconnaissance and communication platforms will benefit from this tensor core. Since missions exist in both DoD and commercial components, the tensor core will be marketed to commercial aerospace and other government entities. Reconnaissance and communications missions both have requirements of taking in and relaying data to the larger constellation or down to a ground station. The tensor core will alleviate much of that required data bandwidth by processing the data into a reduced set.
