SBIR/STTR Award attributes
Green Mountain Semiconductor Inc. is proposing a design and method for evaluation pursuant to the solicitation H6.22-1123. The hardware proposal is a radiation hard neural processor with embedded MRAM in an FD-SOI process; however, much of the innovation is put into the evaluation of error tolerance and design solutions to statistical radiation effects. Most fault tolerant design solutions imply a negative impact on device performance. Size, power and speed can all be degraded from additional design solutions to prevent faults from radiation. nbsp;In order to create the most efficient neural processing hardware, an evaluation of the impact to the design, and more specifically, what subsystems of the design cause the highest fault rate, must be done. This can be accomplished in software using statistical models in hardware verification simulation. By building a technology specific utility for fault injection, critical subsystems can be evaluated and corrective design solutions (i.e. redundancy, dice latches, error correction codes, device selection and other methods) can be used to reduce the fault rate optimally. This technique can be taken further in scope to evaluate the fault tolerance of trained neural networks during inference. For future work, it will be possible to train networks while applying statistical fault injection to allow the network to become naturally resistant to radiation induced fails.
