TL;DR:
- Tsinghua University unveils ‘Gemini’ to enhance chiplet-based DNN accelerators.
- Focus on balancing cost, performance, and energy efficiency.
- Gemini employs innovative encoding for low-power spatial mapping.
- Utilizes dynamic programming and Simulated-Annealing-based optimization.
- SA algorithm with five operators refines spatial mapping space.
- Dynamic optimization of data transmission and intra-core dataflow.
- Achieves superior performance with minimal cost increase.
- Promises to revolutionize the deep neural network accelerator landscape.
Main AI News:
Researchers from Tsinghua University have introduced ‘Gemini,’ a groundbreaking AI-driven approach aimed at enhancing the performance and energy efficiency of chiplet-based deep neural network accelerators. This innovation comes in response to the pressing need to strike an optimal balance between monetary cost (MC), performance, and energy efficiency in existing DNN accelerators.
The intricacies of this challenge revolve around a multitude of factors, including network-on-chip (NoC) communication, core placement, and diverse DNN attributes. To tackle this complexity effectively, researchers are exploring an expansive design landscape to uncover innovative solutions.
Gemini, the brainchild of these research efforts, presents a novel approach that redefines low-power (LP) spatial mapping schemes using a unique encoding method. This method allows for a comprehensive exploration of hidden optimization possibilities. The framework leverages dynamic programming-based graph partition algorithms and a Simulated-Annealing-based (SA-based) approach to achieve optimization.
Within Gemini, the mapping component harnesses the power of the SA algorithm, incorporating five operators tailored to efficiently explore the LP spatial mapping space. These operators encompass the modification of partition attributes, core swapping within computational groups (CG), and adjustments to DRAM-related attributes. The framework dynamically optimizes data transmission, intra-core dataflow, and D2D link communication, ultimately leading to enhanced performance and energy efficiency. An Evaluator module plays a crucial role in the evaluation process, assessing MC, energy consumption, and delay.
The architectural aspect of Gemini provides a highly configurable hardware template, facilitating precise evaluations for performance, energy efficiency, and MC. Experiments conducted with this framework demonstrate its superiority over existing state-of-the-art (SOTA) designs, such as Simba with Tangram mapping. Remarkably, Gemini achieves significant enhancements while incurring only a marginal increase in MC, underscoring its effectiveness in co-exploring the architecture and mapping space. This innovation holds immense promise in revolutionizing the landscape of deep neural network accelerators and their optimization.
Conclusion:
The introduction of ‘Gemini’ by Tsinghua University signifies a significant advancement in the field of chiplet-based DNN accelerators. By addressing the delicate balance between cost, performance, and energy efficiency, Gemini opens up new avenues for optimizing deep neural network accelerators. Its innovative encoding method and dynamic optimization techniques demonstrate the potential to disrupt the market by delivering superior performance without a substantial cost increase. This development could lead to a transformative shift in the landscape of deep neural network accelerators, with wide-reaching implications for industries relying on AI technologies.
