Skip to main content

3D-PFLOPS

This is an AI-generated image of our 3D-PFLOPS framework. We first used Gemini Flash 2.0 to generate it, and it used ~1.94 to 2.9 watt-hours of energy and around 1,207 tokens; then, for sharpening, we used ChatGPT, which estimated that it used in the range of 50-200 tokens and consumed 1-10 m watt-hours. Now imagine millions/billions of users are generating millions of AI-generated images and videos? What would that cost be?

The 3D-PFLOPS: 3D Integrated Parallel Fabrics enabling Layered Opto-electronic Processors for Near-memory AI Computing project develops a new heterogeneous computer architecture that combines electronics and photonics to make artificial intelligence systems dramatically more energy-efficient. By keeping computation close to memory and routing each phase of an AI workload to the hardware best suited for it, the design overcomes the data-movement, power, and thermal bottlenecks that limit today’s AI accelerators.

  • PI: Dr Omiya Hassan
  • Co-PIs: Dr Karthik Srinivasan and Dr Purab Ranjan Sutradhar
  • Awarded by: US National Science Foundation (NSF)
  • Award Number: 2615713 | Award Details

Project Goals

Electro-Optical application-specific integrated circuit (EOASIC)
Lead: Dr Hassan & Dr Srinivasan

Design a simulation framework for CMOS and photonics processing elements (PE) as training and inference engines.

We are currently hiring graduate research assistants (GRA) for the photonics section of the proposal. Reach out to the designated leads. (US citizen preferred)

Thermal modeling of 3D Heterogeneous AI Engines
Lead: Dr Hassan, Dr Srinivasan & Dr Sutradhar

Verifying the physical feasibility and scalability of the 3D-PFLOPS architecture and designing an architecture-specific thermal model.

Near-Memory Multiprocessing System
Lead: Dr Sutradhar & Dr Hassan

Develop a Gem5X-based system-level simulation framework which integrates all three dies into a near-memory heterogeneous computing architecture.

We are currently hiring graduate research assistants (GRA) for the near-memory systems section of the proposal. Reach out to the designated leads.

We are open to collaborations; please reach out to the PI of this project or designated project goal leads for more information.
Our Sponsorship and Collaborators

Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the U.S. National Science Foundation.