Roughly a year after beginning formal efforts to explore an AI for Science initiative the Department of Energy’s Advanced Scientific Computing Advisory Committee last week accepted a subcommittee report calling for a ten-year AI plan that loosely emulates the Exascale Computing Initiative. The sub-committee report was delivered by Tony Hey of the subcommittee at ASCAC’s two-day virtual meeting held last Thursday and Friday.
“The subcommittee believes that this ten-year AI for Science initiative should be funded at the same scale as the successful Exascale Computing Initiative (ECI) and Exascale Computing Project (ECP),” contends the report and even included a graphic presenting how such a program might be structured that looks much like the ECI/ECP program.
The report recommends this AI for Science initiative be structured around four major AI R&D themes:
- AI-enabled applications
- AI algorithms and foundational research
- AI software infrastructure
- New hardware technologies for AI
and makes six major recommendations:
- Creation of a 10-year AI for Science Initiative
- Structure of an SC AI for Science Initiative
- An Instrument to Edge Initiative
- Training, focusing, and retention of AI/ML workforce
- Inter-Agency collaboration
- International collaboration
“In addition, the subcommittee stresses the importance for all six of the Office of Science programs to work together on the issue of hardware-software-algorithm co-design and data analysis at their major user facilities. Finally, the subcommittee supports the recommendation of the ECP Transition report [10] that stresses the importance of ASCR’s long-term Applied Mathematics and Computer Science research programs,” concludes the report.
An interesting feature of the new world of virtual meetings such as the ASCAC meeting, which was on Zoom, is there are often comments in the chat window accompanying presentations. Hey, a longtime HPCer currently with affiliations at University of Washington and Rutherford Appleton Laboratory, noted any AI for Science program would necessarily be multi-disciplinary and wide-ranging in terms of product.
Online meeting participants (~150-plus) picked up on this with various chat area comments suggesting the timelines and goalposts for an AI could be less clear and perhaps more variable. The ECI/ECP program has very specific goals and delivery dates for exascale systems and an associated software ecosystem. The goals are less clear-cut for AI. Nevertheless, prospects for an AI for Science effort are enticing with opportunites for advances in basic sciences and in critical tools. One example of the latter is AI’s potential in managing data (data reduction and transport) at the sites of big instruments where massive amounts of raw data are produced and must then be transported to distant sites for analysis. The report called this out out in its instruments-to-the-edge recommendation.
The full report, along with other presentations made at the meeting, are available at the ASCAC website. The rush to develop and adopt AI technologies, of course, has been palpable including enthusiasm from the Trump Administration and the U.S. Congress. A national AI Initiative is part of National Defense Authorization Act (NDAA) whose details are now being hammered out by Senate and Congress leaders. The house version had called for $1.157 billion for the act. In August, the U.S. Administration announced $1 billion to establish NSF-led AI institutes and DoE-led Quantum Information Sciences Institutes, and just this week announced a U.S.-U.K. cooperation agreement on AI research and development.
To some extent, it’s becoming difficult to sort through the complementary and competing government proposals, but AI (and quantum) are clearly hot and seen as potential transformational. Efforts to infuse AI technology into traditional HPC and adapt it for use in science have been steamrolling for some time.
As the latest report notes, “From July to October in 2019, the Argonne, Oak Ridge, and Berkeley National Laboratories hosted a series of four AI for Science Town Hall meetings in Chicago, Oak Ridge, Berkeley, and Washington DC. The four meetings were attended by over 1300 scientists from the 17 DOE Labs, 39 companies, and over 90 universities. The goal of the Town Hall series was ‘to examine scientific opportunities in the areas of artificial intelligence, Big Data, and high-performance computing (HPC) in the next decade, and to capture the big ideas, grand challenges, and next steps to realizing these.’ The discussions at the meetings were captured in the final report of the AI for Science Town Hall meetings.” HPCwire covered the initial report release in March.
Although the AI for Science report has now been accepted by ASCAC, turning its suggestions into a full-fledged program is likely to be a long slog emphasized Barbara Helland, associate director of ASCAC.
Helland presented some budget request numbers noting the House marked-up version of NDAA includes more than the DoE request. Although a reconciliation with a Senate bill had been expected around Labor Day, most observers now say a vote on NDAA won’t occur until after the election. (Helland’s slides below, click to enlarge.)
Helland also reviewed the exascale program status – there was little new information although she broadly confirmed the Aurora system being delivered by Intel was delayed. No details were provided beyond comments that Intel, ANL, and DoE were discussing solutions.
With so much material presented during the meeting, it’s best to review the agenda and drill into presentations of interest.
Link to ASCAC meeting agenda, https://science.osti.gov/-/media/ascr/ascac/pdf/meetings/202009/ASCAC_Agenda_202009.pdf?la=en&hash=A19D1D12F4B9AA99390AE1703819A1F73FA233E1
Link to ASCAC meeting presentation slides, https://science.osti.gov/ascr/ascac/Meetings/202009