Currently, data-intensive scientific applications require vast amounts of compute resources to deliver world-leading science. The climate emergency has made it clear that unlimited use of resources (e.g., energy) for scientific discovery is no longer acceptable. Future computing hardware promises to be much more energy efficient, but without better optimized software this cannot reach its full potential.

In this project we develop a generic AI-driven co-design methodology, using specialized Large Language Models (like ChatGPT), to effectively generate efficient code for emerging computing hardware. We will validate our methodology with two radio astronomy applications, with sustainability as the key performance indicator. 

Data-intensive science, or eScience, is now firmly established as the fourth science paradigm, as introduced by Tony Hey and Jim Gray. While this opens exciting new abilities to explore new science in vast amounts of data, this comes at very significant processing and energy cost. This cost is often overlooked and poorly understood or appreciated. In the current climate crisis we can no longer accept that world-leading science has an unknown, and more importantly, potentially unconstrained environmental impact. We argue that a careful and deliberate consideration needs to be made whether the scientific impact outweighs the potential environmental impact by the processing required. Optimization and a better mapping of software to hardware can shift this balance in our favour. 

A key part in the optimization process is the ability to efficiently co-design software, hardware, and science. However, such co-design is time-consuming and complex due to required manual optimisation and the iterative nature of the process. Furthermore, traditional co-design only takes hardware and software into account, missing an opportunity to have emerging technologies drive new scientific discovery. By developing and extending novel AI techniques like generative large language models (LLMs), we will significantly accelerate the co-design loop, informing and challenging the science community to leverage the evolving scientific potential generated by these emerging technologies. This project combines two distinct co-design concepts: parallel hardware/software co-design that is accelerated and supported by collaborative human/AI co-design. To illustrate this, we offer two examples from radio astronomy. 

First, the LOFAR correlator and beamformer, which have gone through several implementations, with differing word-sizes tailored to the hardware characteristics. On BlueGene/L and BlueGene/P we used double precision complex floating point arithmetic, due to the availability of the double hummer FPU, which was built specifically for that purpose. On Cobalt, the GPU-based correlator beamformer currently employed, we use single precision floating point arithmetic, since that is precise enough for the science cases, while doubling the computational performance. More recent GPUs have AI-optimized tensors cores that utilise even smaller word sizes. These are experimentally used and may be considered for future implementations of the LOFAR correlator and beamformer. Smaller word sizes reduce the bandwidth and storage requirements, which can be traded for spectral bandwidth, giving the researchers more scientific flexibility. 

Second is the LOFAR polyphase filter, which we made configurable in the number of frequency channels produced. This provided the scientists with a parameter that was not originally in the design requirements. This allowed a trade-off between temporal sensitivity and frequency resolution, enabling better Radio Frequency Interference (RFI) mitigation, while simultaneously supporting higher time resolution for observation modes that need it, such as pulsar search. Alternatively, smaller data products could be generated, requiring fewer resources to store and process, thus limiting environmental cost. These two examples illustrate the potential flexibility gain of software-science co-design. The potential energy savings of co-design also are large. For example, the LOFAR correlator became approximately 6.6x more power efficient thanks to migrating from a Blue Gene/P supercomputer to a GPU-based cluster (from 338 to 51 MWh/year), reducing the carbon footprint with approximately 141 tonnes of CO2/year. 

Generative AI, in particular large language models, have shown a remarkable ability to do mundane coding tasks. It is now no longer doubtful that computer programming will change dramatically in the coming decade, as demonstrated with tools like GitHub co-pilot. In this project we introduce AI-driven co-design: we will investigate the ability of AI to accelerate co-design tasks by estimating performance potential of emerging technologies for specific tasks and aiding in the often arduous porting of code. It should be noted that the cost of training or specializing such AI is not free of environmental cost, and this should be considered. 

• artificial intelligence
• co-creation
• large language models

Share on Facebook Share by Bluesky Share on LinkedIn Share by WhatsApp Share by Mastodon