IBM Research

So your team of semiconductor researchers designed and built a brand new chip that has the potential to be far more efficient for AI computations, but how do they know if it works as intended? That’s where IBM Research Hardware Engineer JohnDavid Lancaster comes in. 🦾 In this behind-the-scenes look at his lab in Yorktown Heights, Lancaster demonstrates how he and his team test and validate the “𝗽𝗼𝘄𝗲𝗿 𝗯𝘂𝗱𝗴𝗲𝘁” — a “wattage-to-inference” ratio — of IBM’s latest AIU (Artificial Intelligence Unit) chips, which were specifically designed and optimized for #AI applications. “𝘖𝘯𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘬𝘦𝘺 𝘳𝘦𝘢𝘴𝘰𝘯𝘴 𝘺𝘰𝘶 𝘸𝘰𝘶𝘭𝘥 𝘶𝘴𝘦 𝘰𝘶𝘳 𝘈𝘐𝘜 𝘤𝘢𝘳𝘥 𝘪𝘴 𝘵𝘩𝘦 𝘴𝘶𝘱𝘦𝘳 𝘭𝘰𝘸 𝘱𝘰𝘸𝘦𝘳 𝘧𝘰𝘳 𝘩𝘰𝘸 𝘮𝘢𝘯𝘺 𝘪𝘯𝘧𝘦𝘳𝘦𝘯𝘤𝘦𝘴 𝘺𝘰𝘶 𝘤𝘢𝘯 𝘤𝘰𝘮𝘱𝘶𝘵𝘦.” Lancaster shows how mapping the voltage of the entire card can ultimately help inform sustainable (low-power) #encoder and #decoder model design. 📚 - Read more about our AIU here: https://lnkd.in/gRge8QHh ▶️ - Subscribe and follow for more lab tours: https://lnkd.in/gjytYHxa ---- #IBM #Research #chips

Today, IBM announces a major breakthrough in co-packaged optics that will bring the speed of light to generative AI: https://lnkd.in/gyD94Aua IBM Researchers have invented a way for fiber optics to connect chips on a circuit board, which promises to improve energy efficiency, boost bandwidth and accelerate generative #AI computing development.   This co-packaged optics technology introduces an all-new blueprint for how we transmit information. By enabling chipmakers to add six times as many optical fibers at the edge of a chip (a measure called “beachfront density”) compared to the current state-of-the-art, IBM’s optical structures have the potential to massively boost the bandwidth between chips.    As IBM distinguished engineer John Knickerbocker said about his team’s development of this new technology:     “Even the most capable semiconductor components are only as fast as the connections between them.”  ---- #IBM #Reserach #Chips #Semiconductors #FiberOptics #DataCenter

In this week's edition, we discuss an optics breakthrough that could revolutionize how we train generative AI. We examine clearing hurdles on the path to 2-nanometer chip production, explore bringing light speed to generative AI, celebrate the AI Alliance’s first anniversary, and discuss bringing foundation models to the power grid. ➡ Read more for the latest updates and subscribe here:

Today IBM, Meta and +140 organizations around the world, celebrate the one-year anniversary of The AI Alliance! 🎉 The AI Alliance – made up of companies, startups, universities, research institutions, government organizations, and non-profit foundations – champions a shared goal: to develop #AI collaboratively, transparently, and with a focus on safety, ethics, and the greater good. 𝗜𝗺𝗽𝗮𝗰𝘁 𝗶𝗻 𝟮𝟬𝟮𝟰: • 23 countries • 93 active projects • 1,200 collaborators • 12 working groups • 5 published AI Alliance Guides to AI 𝗟𝗼𝗼𝗸𝗶𝗻𝗴 𝗮𝗵𝗲𝗮𝗱 𝗶𝗻 𝟮𝟬𝟮𝟱: Explore the AI Alliance 2025 initiatives – Trust and Safety Evaluation Initiative (TSEI) and the Open Trusted Data Initiative (OTDI) – here: https://lnkd.in/eBpN_3n8 ---- #IBM #Reserach #MachineLearning #Safety #Accessability

IBM open-sources a pair of prototype tools for detecting hallucinations in #RAG (Retrieval-Augmented Generation) applications and estimating model uncertainty. If feedback is positive, these capabilities could be added to the next Granite release: https://lnkd.in/gUKNq932 Try the LoRas on Hugging Face here: https://lnkd.in/gAcfJnPb ---- #IBM #Research #AI

It is not too often that one’s technical work finds its way into the The New York Times. https://lnkd.in/e3fAxEen The NYT article is based on a scientific report, published earlier this year in PNAS by IBM Research and Yamagata University, describing the development and application of an AI system to assist archaeologists discovering geoglyphs in the Nazca region, a UNESCO World Heritage site.   https://lnkd.in/eyAHg7GH These geoglyphs are motifs found throughout the Nazca region, were created by modifying the ground surface approximately 2000 years ago. Because they have been exposed to weather and other impacts for all these years, they are very difficult to find, particularly the smaller figurative and relief-type ones, which are often only on the order of 10 m or so. While the report constitutes the first Nazca-wide survey searching for geoglyphs, which is now the basis for further understanding why these geoglyphs were created, I am equally excited about the massive benefits that IBM’s AI technology has provided to advance science by automating a very manual and non-scalable process of finding, documenting etc. historical artifacts. It took nearly a century to discover a total of 430 figurative geoglyphs. By contrast, our PNAS paper demonstrates that the development and deployment of a geospatial AI system has led to the discovery of 303 new figurative geoglyphs within only 6 months of field survey, nearly doubling the number of known figurative geoglyphs. But even better, we expect that the AI-analysis will lead to the discovery of ~500 additional geoglyphs, simply because the archeologists did not have the chance to inspect the AI-suggested sites. While there are many healthy and good discussions about the challenges of AI and its uses, this work illustrates quite beautifully and convincingly, how AI can assist and accelerate scientific discovery by at least an order of magnitude. Work like this one will help to motivate the acceleration of AI development, especially for science. An example is the development of the recent geospatial AI Foundation Models, which will provide additional benefits in the near future. https://lnkd.in/ecrjQRmx 

In the early 80s, a few curious IBM researchers were testing a new excimer laser to see if they could use it for circuit board etching. The laser technology was so accurate, the team hypothesized it could be used on organic material, without any collateral damage. To test the idea, the researchers used their laser on the turkey bones from their #Thanksgiving leftovers. Their intuition was correct: there was no damage beyond where the laser was shot. A few years later, their findings inspired a group of #ophthalmologists to use these lasers in corrective eye surgery, one method of which eventually became #LASIK. Don’t be so quick to throw away your leftovers — you never know what they might inspire! https://lnkd.in/gPNKHBuq ---- #IBM #Research #Hardware #Healthcare #Innovation

Supercharging Training using float8 and FSDP2 ⚡ Read our latest blog to find out how we achieve up to 50% throughput speedup while achieving loss & evaluation benchmark parity in training: https://hubs.la/Q02Zmfvb0

In this week's edition, we explore accelerating discoveries with new AI tools, powerful chips, and turkey. We introduce a new tool designed to speed up the discovery of new materials, discuss adapters intended to manage AI outputs, take a closer look at a Hardware Bring-Up Lab, and reveal how Thanksgiving turkey contributed to the development of LASIK. ➡ Read more for the latest updates and subscribe here:

On this day in 1966, IBM researchers announced a major breakthrough in chipmaking, drastically shrinking down the size of memory cores using ancient candle-making techniques. These new experimental cores were just slightly wider than a human hair and were nearly impossible to produce using the traditional pressing technique used to form memory cores at that time. 𝗧𝗵𝗲 𝗽𝗿𝗼𝗰𝗲𝘀𝘀   Fabrication began with a nylon "wick" dipped into a bath of varnish and magnetic powder, the magnetic taper was then passed through a drying oven and the process was repeated until the desired thickness was achieved. In a similar process, a layer of non-magnetic material was added for added strength.  The magnetic taper was subsequently cut into sections and frozen in a block of wax. At this stage, the nylon "wicks" were removed, leaving behind hollow tubes made of magnetic ferrite material. The wax block was then sliced into sections, where thousands of cores could be cut. Finally, these slices were fired in an oven, which vaporized the wax and sintered the magnetic materials. This led to the development of smaller, packed memory cores, which in turn increased computer memory capacity as well as device processing speeds. 𝗧𝗵𝗲 𝗽𝗵𝗼𝘁𝗼𝘀 Pictured are miniature experimental memory cores created using the techniques of ancient candlemakers, followed by images of standard IBM magnetic-core memory planes. Stay tuned for more highlights from IBM Research's history. ---- #IBM #Research #Compute #IBMHistory