% IMPORTANT: The following is UTF-8 encoded.  This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.

@ARTICLE{Leroux:1050455,
      author       = {Leroux, Nathan and Manea, Paul and Sudarshan, Chirag and
                      Finkbeiner, Jan and Siegel, Sebastian and Strachan, John
                      Paul and Neftci, Emre},
      title        = {{A}nalog in-memory computing attention mechanism for fast
                      and energy-efficient large language models},
      journal      = {Nature computational science},
      volume       = {5},
      number       = {9},
      issn         = {2662-8457},
      address      = {London},
      publisher    = {Nature Research},
      reportid     = {FZJ-2026-00225},
      pages        = {813 - 824},
      year         = {2025},
      abstract     = {Transformer networks, driven by self-attention, are central
                      to large languagemodels. In generative transformers,
                      self-attention uses cache memoryto store token projections,
                      avoiding recomputation at each time step.However, graphics
                      processing unit (GPU)-stored projections must be loadedinto
                      static random-access memory for each new generation step,
                      causinglatency and energy bottlenecks. Here we present a
                      custom self-attentionin-memory computing architecture based
                      on emerging charge-basedmemories called gain cells, which
                      can be efficiently written to store newtokens during
                      sequence generation and enable parallel analog
                      dot-productcomputation required for self-attention. However,
                      the analog gain-cellcircuits introduce non-idealities and
                      constraints preventing the directmapping of pre-trained
                      models. To circumvent this problem, we design
                      aninitialization algorithm achieving text-processing
                      performance comparableto GPT-2 without training from
                      scratch. Our architecture reduces attentionlatency and
                      energy consumption by up to two and four orders of
                      magnitude,respectively, compared with GPUs, marking a
                      substantial step towardultrafast, low-power generative
                      transformers},
      cin          = {PGI-14 / PGI-15},
      ddc          = {004},
      cid          = {I:(DE-Juel1)PGI-14-20210412 / I:(DE-Juel1)PGI-15-20210701},
      pnm          = {5234 - Emerging NC Architectures (POF4-523)},
      pid          = {G:(DE-HGF)POF4-5234},
      typ          = {PUB:(DE-HGF)16},
      doi          = {10.1038/s43588-025-00854-1},
      url          = {https://juser.fz-juelich.de/record/1050455},
}