Edge-AI targeted with NVM memory cell

Target applications include energy-efficient, smart-sensor nodes to support artificial intelligence on the Internet of Things, or ‘edge AI’.

The proof-of-concept chip has been validated for a wide variety of applications (machine learning, control, security). Designed by a Stanford team led by Professors Subhasish Mitra and H.-S. Philip Wong and realised in CEA-Leti’s cleanroom in Grenoble, France, the chip monolithically integrates two heterogeneous technologies: 18 kilobytes (KB) of on-chip RRAM on top of commercial 130nm silicon CMOS with a 16-bit general-purpose microcontroller core with eight kilobytes of SRAM.

The new chip delivers ten times better energy efficiency (at similar speed) versus standard embedded FLASH, thanks to its low operation energy, as well as ultra-fast and energy-efficient transitions from on mode to off mode and vice versa. To save energy, smart-sensor nodes must turn themselves off. Non-volatility, which enables memories to retain data when power is off, is thus becoming an essential on-chip memory characteristic for edge nodes. 

The design of 2.3 bits/cell RRAM enables higher memory density (NVM dense integration) yielding better application results: 2.3x better neural network inference accuracy, for example, compared to a one-bit/cell equivalent memory.

The technology was presented on February 19th at the International Solid-State Circuits Conference (ISSCC) 2019 in San Francisco in a paper titled, ‘A 43pJ/Cycle Non-Volatile Microcontroller with 4.7μs Shutdown/Wake-up Integrating 2.3-bit/Cell Resistive RAM and Resilience Techniques’.

But NVM technologies (RRAM and others) suffer from write failures. Such write failures have catastrophic impact at the application level and significantly diminish the usefulness of NVM such as RRAM. The CEA-Leti and Stanford team created a new technique called ENDURER that overcomes this major challenge. This gives the chip a ten year functional lifetime when continuously running inference with the Modified National Institute of Standards and Technology (MNIST) database, for example.

“The Stanford/CEA-Leti team demonstrated a complete chip that stores multiple bits per on-chip RRAM cell. Stored information is correctly processed when compared with previous demonstrations using standalone RRAM or a few cells in a RAM array,” said Thomas Ernst, Leti’s Chief Scientist for Silicon Components and Technologies. “This multi-bit storage improves the accuracy of neural network inference, a vital component of AI.” 

Mitra said the chip demonstrates several industry firsts for RRAM technology. These include new algorithms that achieve multiple bits-per-cell RRAM at the full memory level, new techniques that exploit RRAM features as well as application characteristics to demonstrate the effectiveness of multiple bits-per-cell RRAM at the computing system level, and new resilience techniques that achieve a useful lifetime for RRAM-based computing systems.

“This is only possible with a unique team with end-to-end expertise across technology, circuits, architecture, and applications,” he said. “The Stanford SystemX Alliance and the Carnot Chair of Excellence in NanoSystems at CEA-Leti enabled such a unique collaboration.”