Talk:Dataflow architecture
 | This article has not yet been rated on Wikipedia's content assessment scale. It is of interest to the following WikiProjects: | ||||||||||
| |||||||||||
Creating a new article for the dataflow used in NPUs
[edit]I have added a paragraph to this article.
- Maurice Wilkes wrote in 1995 that "Data flow stands apart as being the most radical of all approaches to parallelism and the one that has been least successful. ... If any practical machine based on data flow ideas and offering real power ever emerges, it will be very different from what the originators of the concept had in mind."[1]
Many new neural processing units say they are based on the idea of dataflow, and some are mentioned in the existing lead of this article.
- ... successfully implemented in specialized hardware such as in ... artificial intelligence (as: polymorphic dataflow[2] Convolution Engine,[3] structure-driven,[4] dataflow scheduling[5]).
I think it would be good to have a new article that explains these new implementations without the reader having to first read about the original unsuccessful concept. Perhaps it would be useful to create a stub article which can be expanded. But what should it be called? Many sources about these new implementations use the word "spatial". The article on spatial computing is about a different concept, although I have temporarily added this paragraph there.
- In computing, the word "spatial" has also been used to refer to the unrelated concept of moving data between processing elements that are arranged in a physical space. In 1992, "spatial machines" were suggested as an approach to parallel computation.[6] In 2013, a programming standard was proposed for "spatial computing".[7] Computer scientists at ETH Zurich have proposed a "spatial computer" model for energy-efficient parallel computation.[8] AMD describe AMD XDNA as a "spatial dataflow NPU architecture", and the University of Illinois is developing a compiler framework for "spatial dataflow accelerators".[9][10]
I suggest following the sources from AMD and the University of Illinois to call the new article spatial dataflow. Are there any comments? JonH (talk) 02:20, 27 December 2024 (UTC)
- For a good explanation of "spatial dataflow", see Dataflow Architecture: Pure, Hybrid, and Spatial by Nathan Beckmann (2019). JonH (talk) 22:25, 12 January 2025 (UTC)
References
- ^ M. V. Wilkes, Computing Perspectives, Morgan Kaufmann, 1995, ISBN 1-55860-317-4, page 79.
- ^ Maxfield, Max (24 December 2020). "Say Hello to Deep Vision's Polymorphic Dataflow Architecture". Electronic Engineering Journal. Techfocus media.
- ^ "Kinara (formerly Deep Vision)". Kinara. 2022. Retrieved 2022-12-11.
- ^ "Hailo". Hailo. Retrieved 2022-12-11.
- ^ Lie, Sean (29 August 2022). Cerebras Architecture Deep Dive: First Look Inside the HW/SW Co-Design for Deep Learning. Cerebras (Report).
- ^ Yosee Feldman and Ehud Shapiro, CACM, 35(10), pages 61 to 73, 1992.
- ^ HPCWire, OpenSPL Consortium Unveils New Programming Standard for Spatial Computing.
- ^ Lukas Gianinazzi, et al, 2023.
- ^ AMD, AMD XDNA Architecture.
- ^ NSF, CAREER: An Agile Compiler Framework for Spatial Dataflow Accelerators.