Neural Computers with Folded Computation, Memory, and I/O

Introduction to Neuromorphic Computing

Neuromorphic computing is a brain-inspired non-Von Neumann computing system that addresses the challenges posed by the Moore’s law memory wall phenomenon. It has the capability to enhance performance while maintaining power efficiency. Neuromorphic chip architecture requirements vary depending on the specific application and the level of complexity. The chip employs a heterogeneous many-core architecture composed of numerous replicated Functional Cores. Neuromorphic systems distribute computation across small elements, similar to neurons, using methods guided by anatomical and functional neural maps from electron microscopy and neural connection studies. The search utilized keywords and search terms including PIM, chiplet-based architectures, deep learning, deep neural networks (DNNs), memory subsystem, dataflow-awareness, communication optimization, thermal considerations, resistive ReRAM, network-on-chip (NoC), latency, power, accuracy, heterogeneous architecture, machine learning workloads, computational units, memory technologies, Von Neumann bottleneck, performance optimization, energy efficiency, and scalability.

In-Memory Computing (IMC) Approach

The first level of folding computation, memory, and I/O into one learned model exploits a memory device’s state dynamics to perform computational tasks in the memory itself, similar to how the brain’s memory and processing are co-located. The second level draws on the brain’s synaptic network structures. A common feature of these approaches is to take inspiration from the brain computation, by co-locating memory and processing (in-memory computing-IMC approach), and then overcoming the von-Neumann architecture. Hardware artificial neural networks (ANN) can implement IMC computing and provides significant potential for enhancing both the performance and energy efficiency of neural network (NN) inference. The architecture is designed to address the challenges of improving the efficiency of CNN inferencing by integrating multiple processing-in-memory (PIM) cores using resistive random-access memory (ReRAM) technology on a single chip.

Recent Research and Development

Recent research has shown promising results in the area of neural computers with folded computation, memory, and I/O. For example, the Netflix AI team has open-sourced VOID, an AI model that erases objects from videos, including physics and all. Google AI has introduced the WebMCP to enable direct and structured website interactions for new AI agents. Zhipu AI has released GLM-4.7-Flash, a 30B-A3B MoE model for efficient local coding and agents. NVIDIA AI researchers have released NitroGen, an open vision action foundation model for generalist gaming agents. Meta AI has released Omnilingual ASR, a suite of open-source multilingual speech recognition models for 1600+ languages. Google AI has introduced Gemini 2.5 ‘Computer Use’ (Preview), a browser-control model to power AI agents to interact with user interfaces.

Applications and Future Directions

The potential applications of neural computers with folded computation, memory, and I/O are vast and varied. They include artificial intelligence, pattern recognition, and sensory processing. Future research directions include exploring new architectures and technologies to further improve the performance and energy efficiency of these computers. The development of standards and protocols for these computers, such as the Model Context Protocol (MCP), will also be crucial for their widespread adoption. As the field continues to evolve, we can expect to see significant advancements in the development of neural computers with folded computation, memory, and I/O.

How this compares

Watch: Technical Walkthrough