V-PCMixer: incorporating 3D vertical phase-change memory based analog in-memory computing in mixer architecture

Area of Science:

• Computer Engineering
• Materials Science
• Artificial Intelligence

Background:

• Analog in-memory computing (AIMC) offers a solution to von Neumann architecture limitations by enabling massive parallel computation.
• Phase-change memory (PCM) is a promising non-volatile analog device with high technological maturity.
• MLP-based Mixers present a simpler, high-performance alternative to transformer architectures.

Purpose of the Study:

• To propose the V-PCMixer, an efficient AIMC architecture tailored for MLP-based Mixers.
• To leverage 3D Vertical Phase-change Memory (V-PCM) for high-areal-density synaptic weights.
• To demonstrate system-technology co-optimization by integrating computational flows into a 3D structure.

Main Methods:

• Utilized a 3D V-PCM device with an efficient fabrication process and cross-point array configuration.
• Employed improved etch and deposition techniques to ensure low variance and read noise.
• Leveraged the inherent matrix transposition capability of the 3D V-PCM array for the V-PCMixer.

Main Results:

• The V-PCMixer architecture achieved approximately 90% accuracy on Cifar-10 classification.
• A minor accuracy drop of 3-4%p was observed when converting the trained model to analog.
• The 3D V-PCM structure facilitated reduced computational complexity and load.

Conclusions:

• The V-PCMixer demonstrates a viable and efficient AIMC architecture for MLP-based Mixers.
• System-technology co-optimization through 3D integration offers advantages over 2D array extensions.
• This work highlights the potential of 3D V-PCM for advanced AI hardware acceleration.