Description: Advantages of Analog in-memory computation (AIMC) for rad hardened electronics Fully-digital accelerators for machine learning and deep learning algorithms usually have very high power consumption and therefore very low power efficiency mainly because of the frequent transfer of network parameters from main memory to the processor, this efficiency limitation is also known as memory bottleneck. These unnecessary weight transfers can be avoided through a concept known as in-memory computation in which we perform a computation in analog within the same memory which stores network parameters. This will give us two key advantages resulting in higher system efficiency: (i) power wasted previously to transfer weights will be saved and (ii) now we can perform the necessary computations with higher parallelism, i.e. one operation per each memory device. Moreover, there are lots of reasons of why analog in-memory computing system may tolerate radiation better than their digital counterparts in harsh environments like space. Here we will go over some of the key differences between these two systems which will make the AIMC hardware a more favorable one. It should be noted that the focus here is on the computing module like an accelerator and excludes parts like controller or memory which may be the same in both analog and digital systems. The three main advantages of our architecture for rad hardened electronics are: Von-Neumann architecture is sensitive to radiation as it separates memory and processor, multiplying the risk of failure. In AIMC systems we are merging memory and computing together. Digital systems have limited redundancies. The failure of a single transistor in a processor can results in whole system failure. In AIMC we have 1000x more redundancy. Neural networks are very tolerant to imprecise weights, and missing links. We are implementing neural network in silicon rather than emulating them digitally. This transfers the redundancy directly to our hardware.

Benefits: The Mentium neuromorphic chip will have many applications for high-throughput low-SWaP neural net computing, especially in embedded systems. For Aerospace missions, this includes communication, vision processing and other sensor modalities for object classification and scene interpretation, decision making encoded through deep learning, and control applications. Research towards radiation tolerance will enable this processor advance to be used in space applications, from LEO to lunar sustainability to martian and outer-planet exploration.

There is a vast commercial market for space electronics. Many commercial and non-NASA agencies will find benefit in the same application as NASA would, even if the main market her would in LEO application (launch systems and satellites). Image enhancement object detection sensors data analysis communication optimization navigation fault-detection and reaction Autonomous flying/docking/landing