Researchers from multiple universities have been working hard to address the challenge of designing large-scale DNN chiplet accelerators, focusing on optimizing monetary cost (MC), performance, and energy efficiency. This complexity arises from the numerous parameters, such as network-on-chip (NoC) communication, core positions, and different DNN attributes. As such, it is crucial to explore a vast design space for effective solutions.
Introducing Gemini – the architecture and mapping co-exploration framework for DNN chiplet accelerators. It employs a novel encoding method to define low-power (LP) spatial mapping schemes, allowing for an exhaustive exploration of hidden optimization opportunities. The framework utilizes a dynamic programming-based graph partition algorithm and a Simulated-Annealing-based (SA-based) approach for optimization. Through this, Gemini’s mapping component uses the SA algorithm with five operators tailored to efficiently explore the LP spatial mapping space. These operators include modifying partition attributes, swapping cores within computational groups (CG), and adjusting DRAM-related attributes.
The evaluation process of Gemini involves assessing MC, energy consumption, and delay through an Evaluator module. This leads to the optimization of data transmission, intra-core dataflow, and D2D link communication, contributing to enhanced performance and energy efficiency. On the architecture aspect, Gemini provides a highly configurable hardware template, enabling precise evaluations for performance, energy, and MC.
Experiments conducted by the researchers showcase that the explored architecture and mapping scheme outperforms existing state-of-the-art (SOTA) designs like Simba with Tangram mapping. Gemini also achieves significant improvements with only a marginal increase in MC, demonstrating its effectiveness in co-exploring the architecture and mapping space.
We are excited to share with you the discoveries made by the researchers from Tsinghua University – the novel Gemini framework offers a comprehensive solution to the intricate challenges of designing DNN chiplet accelerators. The experiments not only validate Gemini’s effectiveness but also shed light on the potential benefits of chiplet technology in architecture design.
Overall, Gemini stands out as a valuable tool for researchers and practitioners aiming to design high-performance and energy-efficient DNN accelerators. Don’t miss out on this incredible opportunity to enhance your work – follow us on Twitter, join our 35k+ ML SubReddit, 41k+ Facebook Community, Discord Channel, and LinkedIn Group. Be sure to also check out the Paper to learn more about all the great work the researchers have done.
