Meet Dawn, one of the most powerful AI supercomputer in the UK. With more than a thousand top-end Intel graphics processing units (GPUs) operating inside its server stacks, Dawn enables scientists within the University of Cambridge and across the UK to make advances in critical research fields such as clean energy, personalised medicine and climate.
Dawn has been created via a highly innovative long-term co-design partnership between the University of Cambridge, UK Research & Innovation, the UK Atomic Energy Authority and global tech leaders Intel and Dell Technologies. This partnership brings highly valuable technology first-mover status and inward investment into the UK technology sector.
Powering a cutting-edge sea ice forecasting system
One of the projects that’s been running on Dawn almost since the start is IceNet, a cutting-edge AI sea ice forecasting system developed by an international team and led by the British Antarctic Survey (BAS) and The Alan Turing Institute.
IceNet has been trained on observational data to forecast the next three months of daily sea ice concentration maps.
The project advances the range of accurate sea ice forecasts, outperforming a state-of-the-art dynamical model in seasonal forecasts of summer sea ice, particularly for extreme sea ice events. This step-change in sea ice forecasting ability is bringing us closer to conservation tools that mitigate risks associated with rapid sea ice loss.
Running the IceNet pipeline on Dawn has allowed the team to complete training significantly faster, experiencing higher data throughput with easier scaling.
The team’s use of a reproducible pipeline allowed the IceNet infrastructure to be easily ported to Dawn and allowed easy trialling on the top-end Intel GPUs operating inside its server stacks.
Dr Scott Hosking, head of the BAS AI Lab and Director for the Environment and Sustainability Grand Challenge at The Alan Turing Institute, leads the project.
He says: “AI models learn from the data, and you need big compute for that and that’s where Dawn comes in.
“With a traditional physics-based numerical model, you have to run the entire model from scratch each time to come up with an answer, this can take hours or days to run.
"But with AI, once you’ve trained the model you can make predictions within a matter of seconds, running it again and again to forecast multiple possible scenarios.
“And the cool thing is that once we’ve trained the model on Dawn we can stick it on our laptop and we can be in the field, on the ship, and we can make forecasts on the fly. That’s a fantastic capability.”
Scott explains that where Dawn really comes in is forecasting sea ice despite the uncertainty created by our changing climate.
“The climate is continually changing, we’re on an upwards trajectory with global warming with the Arctic warming faster than anywhere right now,” he says.
“One of the challenges with AI models is that we train them on past data and of course the past is not representative of the future, so as more data becomes available, we will need to regularly retrain the model. This is where Dawn is such a powerful tool and capability to have, to quickly retrain the model with that new data.
“But one of the novel aspects with IceNet is we also trained the model on future climate simulation data, so that means that IceNet has seen a future world, if you like, and we then fine-tune the model on the satellite observations."
And all this is already being put to real-world use, with IceNet’s forecasts being integrated into tools for wildlife protection, community empowerment, and marine navigation.
“Ships are currently one of the biggest carbon emitters, and ploughing an icebreaker through the ice demands a lot of fuel. With our forecasts we can help shipping plot the best route to produce the least carbon”, says Scott.
“Another example is wildlife who depend on the sea ice to survive. Understanding and being able to forecast sea ice can tell us when migrations can take place and allow conservationists to develop mitigation plans.”