Celestial AI mentioned in lengthy technical report An All-Optical General-Purpose CPU and Optical Computer Architecture
posted on
Sep 24, 2024 10:42AM
https://arxiv.org/html/2403.00045v2
Abstract
Energy efficiency of electronic digital processors is primarily limited by the energy consumption of electronic communication and interconnects. The industry is almost unanimously pushing towards replacing both long-haul, as well as local chip interconnects, using optics to drastically increase efficiency. In this paper, we explore what comes after the successful migration to optical interconnects, as with this inefficiency solved, the main source of energy consumption will be electronic digital computing, memory and electro-optical conversion. Our approach attempts to address all these issues by introducing efficient all-optical digital computing and memory, which in turn eliminates the need for electro-optical conversions. Here, we demonstrate for the first time a scheme to enable general-purpose digital data processing in an integrated form and present our photonic integrated circuit (PIC) implementation. For this demonstration we implemented a URISC architecture capable of running any classical piece of software all-optically and present a comprehensive architectural framework for all-optical computing to go beyond.
Index Terms:Photonics, optical computer, all-optical switching, computer architecture, CPU, RISC.
Optical computers have often been seen as the next step in the future of computing ever since the 1960s[1]. They promise much higher energy efficiencies, while offering near latency free, high-performance computing (HPC), as well as easily scalable data bandwidth and parallelism. Ever since 1957 with von Neumann publishing the first related patent[2] and Bell Labs creating the first optical digital computing circuits[3, 4, 5, 6], the field has gone through multiple hype cycles and downturns[1]. While an all-optical, general-purpose computer hadn’t been achieved yet so far, the DOC-II[7] did come close, still requiring an electronic processing step each iteration. A lot of results from the research since has found its way into use on the optical communication side, with optical communication to enhance electronic computing proving itself to be the most useful and successful development.
There is a very sensible push in industry and research, that the focus should not be on the computation itself, but rather on the communication between compute elements[8]. Currently, the energy consumption of electronic communication happening on and off chip can reach over percent8080\%80 % of the total power consumption of an electronic processor[9, 10, 11]. By replacing these electronic wires with optical interconnects, one can expect an immediate efficiency gain of up to a factor of 666, due to waveguides being near lossless and highly efficient electro-optical modulation schemes. As an example, communicating a bit across an electronic chip can cost up to 600 fJ, whereas using electro-optical modulation would enable far lower energies, down to the 1-10 fJ domain[12], with the potential of reaching the attojoules in the coming years. The industry expects the majority of efficiency gains in the near future coming from such electro-optical hybrid computing approach, with both large companies, like Intel and Nvidia, as well as startups, like Lightmatter, Ayar Labs, Celestial AI and Black Semiconductor, actively developing schemes for optical interposers to connect electronic computing elements and memory. This development is highly beneficial not just for the field of computing, but also photonics in general.