AI research needs major investment in infrastructure

Interest in AI research has exploded in recent years, and with it has come a need for computer power, not least for research that benefits from the use of machine learning and neural networks. Traditional supercomputers are not the optimal choice for these types of calculation: graphical processor units, or GPUs, are required, and lots of them. In addition, advanced user support must be provided, and a greater degree of interactivity with the systems than can be achieved with large HPC systems. (“HPC” is an abbreviation for “high-performance computing”, or what we call in everyday speech “supercomputers”.)

AI Infrastructure

During the autumn of 2018, Anders Ynnerman and a group of colleagues from KTH Royal Institute of Technology, LiU, Lund University and Umeå University were charged with the task of drawing up a report about Swedish requirements for AI infrastructure. The task included recommending how such infrastructure should be designed, and assessing how it would influence the overall system of research funding. The group requested input from several expert panels in various fields of research, and arranged a large national hearing to provide background information for its report.

The conclusion is unambiguous:
“There is an urgent and rapidly growing need for computational resources supporting AI workflows. Unless investments are made rapidly and with long-term plans for continuous build-up, the competitiveness of Swedish research in artificial intelligence and machine learning will be jeopardised.”

The good news is that the Knut and Alice Wallenberg Foundation has donated SEK 70 million for initiatives in AI infrastructure, as part of the large WASP AI programme (Wallenberg AI Autonomous Systems and Software Program). This foundation, together with the Swedish National Infrastructure for Computing (SNIC), is also the body that gave Anders Ynnerman and the group the task of drawing up the report.

Depends on cloud-based services

“The Wallenberg foundation has recognised the need, and we are extremely grateful for that. Seventy million kronor is a significant first step in the right direction, but the requirements have exploded and we must find ways to ensure long-term and continuous investment in infrastructure for AI research”, says Anders Ynnerman.

Swedish research currently depends on the cloud-based services provided by, for example, Amazon, or on each research group purchasing its own graphics processors.

“Neither strategy is cost-effective. And we can’t use external cloud-based services when we are processing sensitive data.”

The solution that Anders and his co-authors envisage is to collect resources and invest in a common infrastructure in one location under the auspices of the SNIC. This would be a resource for all researchers in Sweden, not just those associated with WASP. SNIC offers an established organisation and methods for the distribution of computer resources. It has access to supplementary hardware, such as storage and networks, giving synergy effects, and it has a well-functioning helpdesk, to which AI support can be added.

However, the report concludes that SNIC’s resources will not be sufficient, and further long-term financing is required. It points out that no solutions for this are currently available.

The SNIC initiative

This is not the first time that Anders Ynnerman has united the Swedish research community in concerted action. The very formation of SNIC is a result of his work.

Anders Ynnerman, professor LiU Photo credit Göran Billeson“The situation at the end of the 20th century was worrying: there was no long-term plan or financing for supercomputers in Sweden, and all centres were competing with each other. This prevented the collaboration that Sweden needed”, recalls Anders Ynnerman during the 30th birthday celebrations of the National Supercomputer Centre, NSC.

While driving in the British countryside between Newcastle and Oxford, he decided that something had to be done about the dismal situation. As soon as he returned to Sweden, he took the initiative to building the collaboration within Swedish supercomputing that today is SNIC. This now has six supercomputer centres, of which the NSC at LiU is the largest. They are financed equally by the Swedish Research Council and the 10 Swedish universities that are members of the collaboration. Anders Ynnerman was the fourth director of NSC, from 1998 to 2002.

“Investments of this type have other consequences. For example, the purchase by NSC and LiU in 1999 of what was at the time an advanced supercomputer from Silicon Graphics was the basis on which both Visualization Center C and the Center for Medical Image Science and Visualization (CMIV) were built. This had major implications for the possibility of awakening the interest of the general public, and in particular young people, for science and technology; making virtual space travel to March and Pluto possible; and – not least – leading to huge improvements in Swedish medical care”, Anders Ynnerman points out.

So now it is once again high time to unite Swedish resources.
“Developments in technology are rapid, but the group feels confident in its assessment that what we need now is a common GPU resource, and that the need for it is urgent.”

Translated by George Farrants