TeAAL and HiFiber Tutorial (MICRO 2025)

[Koji Inoue]

みなさま

依頼を受けましたので、SWoPP-ML をお借りして、MICRO 2025

で開催されるチュートリアルの案内を流させて頂きます。

よろしくお願いします！

TeAAL and HiFiber: Precise and Concise Descriptions of (Sparse) Tensor Algebra Accelerators

https://teaal.csail.mit.edu/tutorials/2025.micro-teaal/index.html

There has been an explosion of research into specialized accelerators for a variety of domains including for deep learning, graph algorithms, FFT, cryptography, and more. However, most papers do not contain enough information to reproduce their results. Instead of slogging through pages of descriptions and squinting at figures, what if you could understand the salient characteristics of each accelerator by reading through about 30 lines of code? Our tutorial “TeAAL and HiFiber: Precise and Concise Descriptions of (Sparse) Tensor Algebra Accelerators” will teach you just that. This tutorial will be held in Seoul on Saturday, October 18, 2025 1-5pm, co-located with MICRO 2025.

Through the tutorial, you will learn:
- A set of fundamental abstractions (cascades of Einsums, fibertrees, etc.) that can be used to describe domain-specific accelerators
- A declarative specification language (TeAAL - https://github.com/FPSG-UIUC/teaal-compiler) for writing an accelerator design in terms of the above abstractions
- A format-agnostic loop nest representation (HiFiber - https://github.com/Fibertree-Project/fibertree) for visualizing and executing kernel implementations
- A set of analysis techniques (e.g., for performance modeling) that can be applied to kernels expressed in terms of the above abstractions

As a part of the tutorial, we are providing an “accelerator zoo”—a set of recent accelerator proposals, their TeAAL specifications, and a compiler that will automatically generate the corresponding HiFiber code (from the TeAAL specification): https://github.com/FPSG-UIUC/accelerator-zoo.

--

いのうえ