IBM Research

Analog chips promise lightning-fast inference for AI models, but their physical properties make model training difficult or impossible. To address this gap, Tayfun Gokmen and his team are working on new algorithms that will enable model training in these energy efficient devices. https://lnkd.in/eq22dTjS

This week, Natcast.org launched the #NSTC Membership program to bolster the American semiconductor industry through collaboration. In response, Mukesh Khare, GM, IBM Semiconductors and VP, Hybrid Cloud Research at IBM, shared: “Advances in semiconductor technology and manufacturing require collaboration. Through the public-private ecosystem IBM has fostered at the Albany NanoTech Complex in partnership with NY CREATES and other industry leaders, we have shown how working together can accelerate semiconductor research and development, as well as strengthen our supply chains. We support Natcast’s NSTC Membership program to connect even more industry leaders, startups, government agencies, and academic institutions. Together, these members can use NSTC resources to power the American semiconductor industry and strengthen the U.S. semiconductor workforce.” Learn more about the NSTC Membership program: https://lnkd.in/eKxz3VpP

In this week's newsletter, we explore better weather forecasting, AI inferencing breakthroughs, and quantum leaps. We introduce a new general-purpose AI model for weather and climate, we discuss a brain-inspired, energy-efficient AI chip and the latest expansion of the IBM Quantum Data Center in Poughkeepsie, and we look into creative strategies to reduce the memory footprint of LLMs. Find out more on the latest news and subscribe here:

IBM Research has brought LLMs to our prototype AIU NorthPole chips. In testing, we found that NorthPole can perform inference on a 3 billion parameter LLM with lower latency and higher energy efficiency than the GPUs commonly used for AI. The brain-inspired design uses on-chip memory to achieve unprecedented memory bandwidth, which is necessary to deliver AI applications quickly and efficiently. https://lnkd.in/evENEsy7

Despite all their capabilities, large language models can struggle with long input sequences due to their limited memory capacity. To address this, several teams at IBM Research are working on solutions that make it easy to upgrade LLMs with additional memory without having to retrain the models. https://lnkd.in/ecE6huDh

In collaboration with NASA, IBM has released an open-source foundation model that can be customized for a variety of weather and climate-related applications and served from a desktop computer. Potential applications include creating targeted forecasts from local weather data, predicting extreme weather events, improving the spatial resolution of global climate simulations, and improving the representation of physical processes in conventional weather and climate models. Learn more! Blog post: https://ibm.co/3TDul9a Research paper: https://ibm.co/3TAILXG Download the foundation model on Hugging Face: https://ibm.co/3XPmhol

In this week's newsletter, we explore AI training and boosting your fantasy football performance. We share some of the highlights from the PyTorch Conference 2024, we look at how IBM researchers are making it easier to train AI models, and explain how AI can help you win Fantasy Football Read more on the latest news and subscribe here:

Dario Gil of IBM at our AI Summit this month: “I tell my enterprise clients, take your valuable assets, your data, make sure more of it is represented in foundational models that you will own, it will turn out to be one of your most valuable assets.”

At this year's PyTorch conference, scientists from IBM Research are presenting new advances that will make it easier than ever for developers to train AI models on the open-source platform. One of these achievements is a training milestone, which they reached with fully sharded data parallel (FSDP), a strategy to enable faster model training on fewer GPUs. The other is a PyTorch-native data loader that will help developers coordinate and reconfigure training workloads as needed, eliminating a common bottleneck in the model training pipeline. https://lnkd.in/ehmxrs2z

On this day in 1984, IBM announced volume production of the 256,000-bit memory chip. This 256-kilobit memory chip allowed for the assembly of more than 4 million characters of information on a single circuit card, making it the densest computer memory package ever offered to IBM customers at the time. The 256K-bit chip could store four times as much data and occupied only about twice the space of the IBM 64K-bit chip it replaced. The chip was manufactured using optical lithography techniques that produced circuit patterns, the smallest of which were geometric features just 1.5 micrometers wide. Increasing the storage density of computer memory chips was a big step in reducing the cost of component manufacturing, leading to lower costs for computer users during the burgeoning PC era. Stay tuned for more moments in IBM's research history. #IBMHistory