Inventing what's next in AI hardware

Welcome back to the Circuit Breaker, where you can find the best recaps on the latest innovations in AI, quantum computing, semiconductors, and more, from across IBM Research and beyond.

Week of August 11 - August 15

• How the IBM Research AI Hardware Center is building tomorrow’s processors
• Introducing IBM's new library for writing generative programs

Take a look inside the team inventing what's next in AI-first hardware

In 2019 when the IBM Research AI Hardware center was founded, IBM executives had to explain to clients why AI mattered. That's no longer the case. Now the big questions involve how AI models can be deployed as quickly, effectively, and efficiently as possible. Even before the Center's founding, this has been its mission.

🧮 Starting in 2015, before the AI Hardware Center even existed, IBM Research scientists were devising ways to perform deep learning in 16-bit precision. In the next few years, they showed it was possible to train a deep neural network in 8-, 4-, and even 2-bit precision.

📉 At the same time, transistor scaling was slowing down, and it was becoming clear that conventional CPUs and GPUs wouldn't be well-suited for the unique computational demands of AI. Specialized chips would be needed, along with a software stack that could squeeze every last bit of performance out of them.

🤝 With large initial investments from IBM, SUNY Polytechnic Institute, and the state of New York, the IBM Research Center set out to design tomorrow's AI hardware. And this year its first chip will be deployed in IBM z17.

Read more about this team and its journey here.

When generative AI isn't limited to what's in the prompt

📝 Generative AI has exploded in popularity and ability. In a few short years, we’ve seen tools evolve from simple artistic filters and data extraction to sophisticated systems capable of crafting helpful information and content for users in both personal and professional settings.

🍾 By the way we interact with generative AI remains a major bottleneck. Using today’s models, you end up in a game of trial-and-error phrasing, where different requests for the same information can produce disparate results. This unpredictability makes them difficult to deploy in enterprise-grade workflows, where repeatability and precision are essential for mission-critical applications. On top of that, there’s nothing efficient about typing out long requests in plain English into a chat window.

💻 This is where generative computing comes in. As part of the broader shift towards context engineering, IBM Research is pioneering a new idea that treats LLMs as programmable computing elements, just like any other piece of programmable software. This approach introduces structured abstractions, development tools, and runtime environments to replace brittle prompts with robust, maintainable systems. The goal is to bring the rigor of software engineering to generative AI.

🍄 There's mushroom for growth still. One of the first tools IBM is open-sourcing in this space is Mellea, named after the humble honey fungus. The open-source Mellea library offers developers tools to start building structured generative programs that are consistent, scalable, and efficient. It’s a first step towards helping AI builders move beyond ad-hoc prompting and toward reliable, context-aware workflows that can be maintained and reused across applications.

Read more about generative computing here.

Demystifying microelectronics through new displays and vantage points

Launch your tech ambition and register for IBM TechXchange 2025

AI-Controlled Robots Play Hide and Seek in Space

Learning to track an LLM's Chain of Thought (CoT)

Moderna and IBM use quantum computing to model mRNA

IBM tops charts in the 2025 Innovation Guide for Generative AI Consulting and Implementation Services paper by Gartner

Gain access to cutting-edge quantum capabilities with IBM Quantum Credits

Highlighting new publications from IBM researchers that we liked the sound of:

• NGQA: A Nutritional Graph Question Answering Benchmark for Personalized Health-aware Nutritional Reasoning
• Think Again! The Effect of Test-Time Compute on Preferences, Opinions, and Beliefs of Large Language Models
• ConceptAttention: Diffusion Transformers Learn Highly Interpretable Features

If you liked this, please consider following IBM Research on LinkedIn. And if you want to go even deeper, subscribe to our monthly newsletter, Future Forward for more on the latest news on breakthroughs in AI, quantum computing and hybrid cloud.