CSIRO's Bracewell AI system comprises 114 PowerEdge C4130 servers connected by a 13 x 100Gbps 36p EDR InfiniBand switch fabric from Melanox.
The system has 3,192 Xeon cores, 1,634,304 CUDA cores and 29TB RAM.
Although the price was not disclosed, the project had a $4 million budget. It went live earlier this month.
Data61 computer vision group leader Nick Barnes explained that the limited resolution and brightness levels mean the bionic eye has to be able to identify important objects in the scene picked up by the camera so they can be made visible to the user.
The project is applying deep learning to the task of interpreting scenes, as this seems more fruitful than the traditional approach of explicitly-coded rules.
Three people will be fitted with the next-generation bionic eye. It will involve a head-mounted camera, an external processor to simplify the images convert them to a series of impulses transmitted to the implanted device.
“When we conducted our first human trial, participants had to be fully supervised and were mostly limited to the laboratory, but for our next trial we’re aiming to get participants out of the lab and into the real world, controlling the whole system themselves,” said Barnes.
The participants will be allowed to use the system for as long as they wish.
Bracewell will be put to use in other areas, including virtual screening for therapeutic treatments, traffic and logistics optimisation, modelling of new material structures and compositions, and pattern analysis. It mostly runs Linux, but nodes can be dynamically provisioned to run Windows when required by researchers.
“This is a critical enabler for CSIRO science, engineering and innovation," said CSIRO deputy chief information officer and head of scientific computing, Angus Macoustra.
"As a leading global research organisation, it’s important to sustain our global competitiveness by maintaining the currency and performance of our computing and data infrastructures.
“The power of this new system is that it allows our researchers to tackle challenging workloads and ultimately enable CSIRO research to solve real-world issues. The system will nearly double the aggregate computational power available to CSIRO researchers, and will help transform the way we do scientific research and development."
Macoustra said Bracewell is still being tuned, but performance in the petaflop range is expected. Barnes suggested it will be at least one double-precision petaflop, but pointed out that the Tesla GPU's support for half-precision calculations is important for deep learning tasks. He expects to see at least six half-precision petaflops.
Dell EMC ANZ high performance computing lead Andrew Underwood said “Our system removes the complexity from the installation, management and use of artificial intelligence frameworks, and has enabled CSIRO to speed up its time to results for scientific outcomes, which will in turn boost Australia’s competitiveness in the global economy.”
CSIRO's new system is named Bracewell after Ronald N. Bracewell, an Australian astronomer and engineer whose work led to fundamental advances in medical imaging.