Lightmatter’s Post

As the demand for compute surges to fulfill the promise of Generative AI, chips are hitting the thermodynamic limits of energy efficiency and heat dissipation. Nicholas Harris and Lightmatter believe silicon photonics is the solution to unlocking faster data transmission with lower energy consumption. Will light be the foundation of tomorrow’s data centers when modern electronic circuits reach the final process node?

With compute eventually plateauing out the key differentiator is going to be how gainful would inter-IC communications be in the AI compute landscape. #ai #interconnects

The world of AI computing is rapidly evolving, and the excitement is palpable. Dr. Paul Calleja, Director of Research Computing Services, predicts that this Dawn (UK’s fastest AI SuperComputer) of new digital worlds will enable complex problems to be simulated, tested and solved at an unprecedented speed. This is a new era for UK compute, and we can't wait to see what the future holds. #AI #computing #digitalworlds #UKcompute

Even though many people and companies are starting to combine quantum and AI into a single term, the two are very distinct technologies. https://lnkd.in/gAHeqJxx

### Introducing SpiNNaker2: The World's First Commercial Neuromorphic Supercomputer SpiNNcloud Systems has unveiled the SpiNNaker2, a groundbreaking neuromorphic supercomputer that leverages the operational principles of the human brain to deliver high-performance AI computing. This hybrid system is touted as the first commercially available solution of its kind. Developed on an architecture conceived by Steve Furber, one of the original creators of the Arm processor, SpiNNaker2 employs a multitude of low-power chips to optimize the processing of AI tasks and other workloads. The SpiNNaker2 server board features 48 chips, each containing 152 Arm cores, resulting in a total of 7,296 cores. These chips are augmented with various additional components, including distributed GPU-like units designed to accelerate the processing of neuromorphic, hybrid, and conventional AI models. Up to 90 SpiNNaker2 boards can be mounted in a single rack, and scaling is achieved by clustering multiple racks together. This setup is claimed to enable real-time emulation of at least 10 billion interconnected neurons. For machine learning tasks, the system's performance can reach up to 0.3 Exa operations per second (Eops), or 10^18 operations per second. In comparison, Intel's research-grade neuromorphic computer, Hala Point, supports up to 1.15 billion neurons and achieves up to 30 Peta operations per second (Pops). According to SpiNNcloud Systems, the SpiNNaker2 distinguishes itself from traditional GPU-based AI platforms through its versatility and efficiency. Its numerous asynchronous, low-power blocks facilitate more effective workload management, making it a superior choice for diverse AI applications. https://lnkd.in/dnvE_UBu

The most powerful AI processing supercomputer in the world is set to be built in Germany, and planned to become operational within a mere year. Crikey. The JUPITER supercomputer (if you're wondering, the Joint Undertaking Pioneer for Innovative and Transformative Exascale Research, which as backronyms go is spectacular) was commissioned by Europe's supercomputer consortium, EuroHPC JU, (via Techspot) in October 2023, and the first of three planned to be built will be contained in the new facility in Germany, with plans to be operational inside a year. JUPITER will be powered by multiple Nvidia GH200 Grace Hopper Superchips, will be put to work primarily on AI training and is set become the world's most powerful AI system, with a staggering 90 exaFLOPs of performance expected when set to work on training AI models. Built around 50 container modules installed over 2,300 square meters, the modular approach taken to house this monster machine is hoped to cut the delivery time of the project by 50 percent, and should mean that future upgrades will be easier to implement over time.

Check out this important read on why Canada is lagging behind in AI infrastructure and how capacity restraints is hurting our standing in a vital race to lead the world in AI innovation.

Finding a New Path for High-Speed Data Connections One of the primary ways researchers are feeding the compute demands of modern apps like generative AI is by using optical technologies to create faster data connections. https://lnkd.in/duYYQuip

Very proud to be part of the Cambridge Open Zettascale Lab team that delivered Dawn - the UK's current fastest #AI supercomputer - for the University of Cambridge and via the national AI Research Resource other UK researchers. You can learn more about the system and some of the things we have planned for it in this article #HPC

Christopher Edsall Yes, artificial intelligence (AI) can do many things, but NOT everything. With our intellectual property (IP), we are doing what AI can NOT do. Without metadata, especially multilingual metadata, NO multilingual data can be found/retrieved for intelligence analysis, Internet of Things (IoT), etc., even by the most advanced technologies, like AI, supercomputers, etc. https://lnkd.in/g-aJFnXR When we used the following questions to test ChatGPT, an icon of AI, the experiment results showed that it can NOT answer them, while we can, with our IP; for example, a pure English question like: "Who, in England of UK, has new US patents granted on February 20th, 2024?" or a Chinese-English multilingual question like: "Who, in the '湖北' province of mainland China, has new US patents granted on February 20th, 2024?" Our IP is a copyrighted multilingual metadata. It can distill real time (2024.02.20) information about US patent holding/holder, based on the multilingual census geographical locations of Canada, mainland China, Hong Kong, Macao, Taiwan, Middle East (Israel, Saudi Arabia, UAE) and Europe (Austria, Germany, Switzerland, UK). Does our IP ring a bell to you? Thanks.