Paper Title: The First 128-Level-Cell Flash-Based AI Chip

Paper Abstract: One of the hurdles for wide adoption of machine learning has been efficient and high-performance edge compute. Developers use very large, expensive, and power-intensive systems to create new models, but the final implementation needs to be deployed in a cost-effective, small, and low-power solution. This has commonly led to significantly smaller models than state-of-the-art ones leading the benchmarks in an edge environment. A 128-Level-Cell (MLC) flash-based Analog Matrix Processor (AMP) is well positioned to close this deployment gap by providing edge inference compute that is 10-to-100× more efficient than a traditional digital inference. It accomplishes this by leveraging denser circuitry via smaller memory elements with multi-level flash instead of several SRAM cells and denser communication. Deep Neural Networks are an ideal application for the AMP since they can use analog-aware training to overcome the inherent non-idealities present in analog signal chains, and can amortize the fixed costs inside the analog system over larger calculations. A complete SoC solution with firmware and software suppresses the low-power operation noise, and is scalable to the different new NVM technologies beyond conventional floating gates.

Paper Author: HUNG NGUYEN, Senior Fellow NVM, Mythic

Author Bio: Hung Q Nguyen Joined Mythic as Senior Fellow NVM at very early time. Hung brought with him 30 years of NVM industry experience, ranging from EEPROM, Thin-oxide Flash and SST Source-Side-Injection Flash development. Being instrumental for the future NVM technology Enablement at Mythic, Hung has held various responsibilities, including NVM development team, managing the analog custom team. Hung is taking part in an explosive growth of Mythic, an Analog Compute AI company located in both Redwood City, CA and Austin, TX in the last 5 years, from several employees to nearly 150 and potential first AI product family with Flash-based Analog Compute design. Hung held BSEE, MSEE, MBA from UC Berkeley, Santa Clara University. He also held more than 100 patents in NVM technology and design.