Imagine asking a super-intelligent chatbot questions about cancer, climate change, or complex physics—and getting answers that could move research forward faster than ever. That’s exactly what’s happening right now at Argonne National Laboratory (ANL) in the U.S.
They’ve just started training a massive new artificial intelligence model called AuroraGPT, and it could become a powerful tool for scientists around the world.
What Is AuroraGPT?
AuroraGPT is being designed as a kind of scientific version of ChatGPT, but way more advanced and specifically built for researchers. It’s being trained on a massive amount of scientific information, including:
- Research papers
- Scientific data
- Computer code
- Experiment results
This AI will live on Aurora, one of the world’s fastest supercomputers, located at ANL. The system uses Intel’s powerful Ponte Vecchio GPUs to process all this data.
A Chatbot for Science
The end goal? A chatbot interface that scientists can use to ask questions and get smart, data-driven answers.
Let’s say a researcher is working on a new drug or studying climate change patterns. Instead of digging through countless documents, they could ask AuroraGPT and get instant, science-backed insights.
It could save time, reduce costs, and speed up discoveries in fields like:
- Biology
- Cancer research
- Renewable energy
- Physics
- Earth sciences
Just Getting Started
Right now, the training has only begun. The model will eventually include one trillion parameters—that’s AI-speak for how many factors it considers when generating answers. The more parameters, the smarter and more accurate the AI can be.
At the moment, training is happening on just 256 parts—or nodes—of the Aurora supercomputer. But the plan is to scale it up to all 10,000+ nodes, which will take months.
To handle such a huge amount of data and memory, they’re using Microsoft’s Megatron/DeepSpeed system. This software breaks the training into smaller, manageable chunks and runs them in parallel across different GPUs. It’s like training a giant brain by dividing the workload among thousands of smaller parts.
How Powerful Is Aurora?
Intel says its GPUs in the Aurora system actually outperformed Nvidia’s A100 chips in earlier tests on another supercomputer called Theta. That’s a big deal, since most popular AI models—including OpenAI’s GPT-4—are trained on Nvidia’s hardware.
With this kind of performance, Intel and ANL hope to achieve linear scaling—meaning the AI gets smarter and faster as they add more computing power.
AuroraGPT could soon become a key player in global scientific research, offering fast, reliable help for the world’s toughest problems.
Source:
Comments