AI Singapore builds up on-premise setup for compute-heavy tasks

AI Singapore, the country’s national artificial intelligence (AI) programme, is setting up new on-premise computing infrastructure to handle intensive tasks that cost too much to run on the cloud.

The new setup will make use of Dell’s PowerEdge servers and PowerSwitch switches that are tasked to run compute-intense tasks as well as those involving sensitive data that are not run on public clouds.

The move means AI Singapore will expand its on-premise setup to complement the cloud infrastructure that it uses to run its experiments to develop new AI capabilities.

Set up in 2017, the organisation aims to run tasks that are not easily available on commercial solutions sold in the market. It has tied up with academic institutions in Singapore to develop, for example, a way to build Asian language understanding with travel site Expedia.

Another project it has worked on is a way to identify signs of diabetics by analysing images and pictures of a patient’s wound.

Such compute-intensive tasks could cost less to run on premise than on the cloud, said Laurence Liew, director of AI industry innovation at AI Singapore.

While some tasks that require bursts of compute power may be best run on the cloud, those that need consistently robust high-performance computing (HPC) may be better run on premise, he told reporters in a media briefing in Singapore today.

Some compute-intensive tasks can cost as much as S$5,000 to S$10,000 a month on the public cloud, he explained. “In two weeks, I can buy a server.”

The move towards a hybrid and multi-cloud setup is consistent with many organisations, which have found that costs quickly build up after initially rushing to the cloud to cut costs.

Romain Bottier, HPC and AI senior solution architect for South Asia at Dell Technologies, said the company has various ready-to-market AI-related solutions that are delivered on its hardware and software platforms.

For example, customers do not have to spend time developing models for every AI-related task, but use ones that can be generated based on similar models in the past, he added.