Neuromorphic Processors mimic the structure and function of the human brain. Instead of a typical Von Neumann architecture having a discrete compute and memory location, passing data between each, a neuromorphic processor co-locates memory and computes similar to the human brain. The transfer of information between memory and a typical arithmetic logic unit (ALU) is often referred to as a memory bottleneck and is more critical in higher performance compute platforms. Neuromorphic Processors eliminate the memory bottleneck, improving the overall computational efficiency of the system.
Bascom Hunter produces three types of Neuromorphic Processors. The first is the deployment of a neural network into an FPGA. Doing this allows us to take existing COTS hardware and help customers bring existing (or proposed) neural networks into this hardware, speeding up the computation and lowering the power requirements. The second is the deployment of a neural network into custom hardware – typically based on a 3rd party ASIC processor. We can take ASICs produced by the vendor of your choice and put it into a platform and mission specific form factor. For example, into VPX or small-sat form factors. The third type of Neuromorphic processor is intended for our future focused customers looking to engage on cutting edge photonic based neuromorphic processors.
FPGAs offer tangible improvements over traditional CPU and GPU processors for both sensor deployments and Machine Learning (ML) tasks. Bascom Hunter excels in partnering with military and intelligence partners to rapidly develop custom FPGA solutions that leverage SoC and neuromorphic processors that tackle some of the toughest challenges facing our nation today.
ASICs promise significant improvements over traditional CPU, GPU, and TPU processors, perfect for low Size, Weight, and Power (SWaP) Machine Learning (ML) applications. Like FPGAs, ASICs are exceptionally well-suited for machine learning tasks that require a lot of computational power, such as deep learning and neural network training. Bascom Hunter works collaboratively with 3rd party ASIC vendors to integrate their technology for Defense and Intelligence applications.
Our nation’s most demanding applications require performance and accuracy at the speed of light. Missions like Automatic Target Recognition (ATR), Hypersonic Vehicle Detection, and Autonomous Navigation all require real-time processing and inferencing on countless high bandwidth multi-phenomena sensors in parallel, rendering modern Machine Learning (ML) processors like CPUs, TPUs, and GPUs inert. Bascom Hunter specializes in leveraging Silicon Photonics to offload the most compute-intensive tasks in ML, like Matrix Vector Multiplication (MVM). By leveraging analog optical processing, our processors avoid the classic electronic bottlenecks to achieve classification latencies as low as 200 picoseconds; translating to 0.16 Terra Multiply-Accumulate operations per second.