Australia has no next-gen HPC investment plan and clouds can't fill the gap Australia needs an exascale computer system, and a refresh of its current HPC fleet, but lacks a plan or the budget for either – and can't expect cloud providers or quantum computers to offer a suitable substitute for sovereign capacity. That's the thrust of a brief [PDF] published today by the Australian Academy of Science. …
NCI and CSIRO could definitely benefit from access to Exaflopping HPC capacity. The US ORNL's Frontier's been out for 2 years already and that it remains the only offical Exaflopper on Top500, for this long(!), is quite unacceptable IMHO. Australia's Pawsey Centre Setonix is like a mini-Frontier (3rd gen EPYC 64C + MI250x), but 27 PF/s is 40x too doggone slow. Canada's even worse off though, with the Robert and Underhill twins at 8 PF/s (need at least 100x speedup).
At $10 per person ($300 million) Exaflopping machines are just about commodity hardware today for governing bodies, and any country with 30 million people should either already have one, or have plans to get one within the next 2 years in my opinion. AMD's MI300A will make this an even easier decision (or Nvidia's GraceHoppers if you insist). Falling further behind Russia's Yandex should not be an option!
"and any country with 30 million people should either already have one, or have plans to get one within the next 2 years in my opinion"
Oz needs to identify potential areas of global expertise where (a) there's already some advanced Australian capability, and (b) there's the potential to make money. Only when there's a national strategy to support (through policy not subsidy) selected areas, and when there's signs of success, that's the time to think about spaffing taxpayers money on exascale computing. On all indices and studies, Australia is recognised as lagging global economies on all knowledge-led sectors, other than IT and information services - yet the comment in this very forum is that IT is not a good career choice in Australia, that employers would rather hire cheaper migrant labour. How's an exascale computer going to help that?
The country could significantly develop its space sector, but bumbles along with little by way of effective government support. Nanotech is an area where Oz has a foothold, but needs a lot more support to become a global centre. It could specialise in certain niches in biomedical sciences (allowing that pharma is already dominated by other countries). The country has massive potential for secure renewable energy that could attract energy intensive industries. Had it not been so disorganised, Oz would not have just missed the boat in spectacular fashion for new semiconductor fab locations given geopolitical concerns.
Australia: So much potential, so much confidence, but little organisation or vision. And that's why building exascale computing now isn't going to achieve a lot.
This has been a known problem in Australian research circles for a long time, with infrastructure further crippled by a decade of the LNP desire to punish universities for being idealogically unsound. Any money they appeared to give (usually medical) research was robbed from another part of the pre-existing pie. Also there's no way for researchers to contribute their funds to a HPC pool so theyy go and buy something flashy ad put it under a desk and hope that nobody realises that the building heating, aircon and physical security aren't equipped to manage these.
Researchers basically have to courier hard drives between research institutions and HPC resources because they can't move mult-TB to PB data sets over the wires fast enough.
"Australia has no national strategy to acquire and sustain state-of-the-art high-performance computing and data (HPCD) – putting the country's future prosperity and security at risk."
I very much doubt that either of Australia's prosperity or security are going to be compromised in any way by a lack of contenders on the Exascale Interglobal Willy-wavers list.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
