Rainer - Fotolia
Graphcore opens Southeast Asia hub in Singapore
The AI chipmaker is looking to expand its sales teams and hire more engineers to support customer deployments in Singapore and the rest of Southeast Asia
Artificial intelligence (AI) chipmaker Graphcore is setting its sights on the Southeast Asian market with the opening of a regional hub in Singapore.
The hub is expected to support the British company’s local teams and customers across ASEAN’s fast-growing economies, including Vietnam, Malaysia, Thailand, Indonesia and Singapore.
Speaking to Computer Weekly in an exclusive interview, Philip Tan, Graphcore’s Southeast Asia country manager, said plans are afoot to expand its sales teams and hire more engineers to support customer deployments.
The move into Southeast Asia follows the success that the company has seen in other key Asian markets such as China, Japan and South Korea.
Tan said that in South Korea, telco giant Korea Telecom is already using Graphcore’s systems to power its internet datacentre with high-performance AI compute, while a large Chinese company had reportedly replaced its graphics processors with Graphcore’s intelligence processing unit (IPU).
Graphcore’s massively parallel IPU was designed from the ground-up to crunch AI workloads faster and more efficiently. Built using TSMC’s seven-nanometre process, the company’s latest-generation IPU touts 1,472 cores, with the ability to run nearly 9,000 independent parallel program threads.
The company has also been making advances in its software development kits, which are now updated every three months, accelerating its IPU’s performance along with support for more machine learning models.
Natural language processing
Earlier this year, Graphcore put its IPU-POD64 system to the test in the MLPerf AI benchmark, taking just over nine minutes to train the Bert natural language processing (NLP) model in the open division. In a blog post on the test results, Graphcore claimed that its systems deliver “significantly better” performance-per-dollar than rival Nvidia’s offering.
In Southeast Asia, Tan said Graphcore has seen interest in its IPUs from financial institutions, after researchers from the Oxford-Man Institute of Quantitative Finance published a paper on the use of the chips to speed up training of advanced price prediction models.
“When that paper was published, we received a lot of unsolicited calls from hedge funds and stockbroking firms, because you can easily understand the kind of value in terms of dollars and cents if this proves to work for them,” said Tan.
While those firms may not be the largest customers, Tan said they are ready to build their own AI capabilities and would be among the first “low-hanging fruits” for Graphcore to pursue in the region, along with universities and research institutions.
Graphcore has also seen an uptick in local partners as the company expands its footprint across the globe. These include Singapore’s Cxrus Solutions and NetWeb, both of which are “very capable” in high performance computing, said Tan.
However, getting independent software vendors (ISVs) onboard might not be as easy, as many of them tend to be hardware agnostic and prefer to peddle specialised software tailored to specific AI use cases.
“It’s going to be a very interesting market because the AI market really needs the ISVs,” said Tan, adding that Graphcore has hired a director of developer relations and is working on partner programmes targeted at ISVs.
Read more about AI in APAC
- University of Melbourne has teamed up with public and private sector organisations to build an AI application that can predict traffic congestion up to three hours ahead.
- Singapore-based bank DBS is looking to hire 150 developers and engineers skilled in AI and blockchain through its hackathon programme.
- Alibaba’s AliceMind algorithm recorded a higher accuracy rate than humans in answering questions related to images in the global VQA challenge.
- Transport agency in New South Wales taps machine learning to restore public confidence in the state’s transportation network amid the pandemic.
Graphcore IPUs are also available through Graphcloud, a cloud-based, machine-learning platform running on IPU-POD systems hosted by US-based Cirrascale, a specialist provider of cloud-based AI infrastructure.
Tan did not discount the prospect of making Graphcore IPUs available through global hyperscale cloud suppliers, such as the likes of Microsoft, a Graphcore investor, if there is demand for the technology on hyperscale platforms.
Meanwhile, he said there has been good demand for Graphcloud, including from customers in Singapore that are trialling the service.
Asked if Graphcore has been affected by the global semiconductor crunch, Tan said: “We haven’t seen any effects from the supply chain and we don’t see any potential risks in the near future.
“Supply chain management has always been an advantage for Graphcore, and we have a very close friendship with our foundry partner TSMC,” he said.
“Graphcore is also a close innovation technology partner for TSMC exploring future process nodes. Our chip capacity can support our customers over the globe in the very long run.”