Vermont Patent of the Month – March 2023
Artificial Intelligence, once just a thing of Sci-Fi movies, is now in our everyday lives. From AI assistants like Siri to robotic vacuum cleaners choosing their own pathway and avoiding obstacles. They also operate in ways we never see, with banks implementing AI-based fraud detection and physicians using AI-based diagnostic tools to more accurately read MRIs.
This type of technology is only possible when an AI system can be trained with an extensive database. These Artificial Neural Networks (ANN) received training data and uses individual neurons to assign weights to input which is eventually translated to a “right” or “wrong” output. This output then informs the next training set and the process is repeated until the ANN is right almost every time.
Unfortunately, ANNs require significant memory and a machine’s limited bandwidth can quickly bottleneck an AI system’s capabilities. Green Mountain Semiconductor Inc. (GMS) has developed a new method of processing data, using Random Access Memory (RAM) to reduce this bottleneck and allow AIs to run faster than ever.
Their solution builds on top of existing in-memory data processing concepts, where simple binary addition and multiplication needed to represent single Nodes (or Neurons) of an ANN are in the memory chip. With the recursive neural network designed by GMS, an AI system can issue one command from the processor to the memory die and all Nodes of a given ANN can be represented and all calculations can be done in the memory die at once before returning the output of the given ANN back to the processor as a single result.
This method eliminates all communication between the memory and the processor, with the exception of the writing of input data into the ANN and the reading of the inference result output of the ANN. In this way, die-to-die communication could be reduced to a bare minimum, and the issues of memory chip proximity to the processor and the bandwidth of the processor to memory connection and all the power and time lost in that interaction, are eliminated.
GMS’s method is a game-changer because it eliminates the need for Dynamic RAM altogether, and the need for processors to access data from memory, which significantly improves performance and reduces power consumption. The technology can potentially lead to the creation of smarter, more efficient AI systems that can handle complex tasks faster and more accurately.
GMS, founded in 2014, began with a bold but simple vision which was to utilize ingenuity, unique expertise and an abundance of passion for future things to innovate on a truly break-through level to create technologies that continue to transform the way we experience, discover and change the world.
Are you developing new technology for an existing application? Did you know your development work could be eligible for the R&D Tax Credit and you can receive up to 14% back on your expenses? Even if your development isn’t successful your work may still qualify for R&D credits (i.e. you don’t need to have a patent to qualify). To find out more, please contact a Swanson Reed R&D Specialist today or check out our free online eligibility test.
Who We Are:
Swanson Reed is one of the U.S.’ largest Specialist R&D tax advisory firms. We manage all facets of the R&D tax credit program, from claim preparation and audit compliance to claim disputes.
Swanson Reed regularly hosts free webinars and provides free IRS CE and CPE credits for CPAs. For more information please visit us at www.swansonreed.com/webinars or contact your usual Swanson Reed representative.