This is a guest post by Zhang Dixuan, Atlas datacentre domain general manager at Huawei’s intelligent computing business department
The quest for knowledge has driven us to discover the highest mountains and has taken us to the moon some 60 years ago. But how do we find out things we do not yet know that we need to know?
Huawei has been working on this conundrum and has what it believes is the answer: artificial intelligence (AI).
The convergence of 5G, virtual and augmented reality, machine learning and other emerging technologies will help people see beyond distance, distortion, surface and history. Coupled with the use of AI, we will be able to see things we couldn’t imagine and better understand what we already see.
One application is outer space navigation, an area that is fraught with challenges. This is where AI can help to achieve the seemingly impossible and help us to unlock the secrets of the universe.
Consider the Square Kilometre Array (SKA), a global science project supported by 13 countries to build the world’s largest and most powerful radio-telescope to overcome the limitation of the human eye.
The SKA is called an array because it has dozens of large antennas and other types of radio wave receivers, which are arranged and linked together via fibre optic cables. These receivers and antennas span one square kilometre.
According to Philip Diamond, director-general of SKA Organisation, the SKA antennas pick up huge volumes of data which are sent to supercomputers, where the data is processed to allow astronomers to undertake scientific investigations.
But that is just part of the story. The data needs to be handled in new ways to ensure vast amounts of data that grow to as much as 600 petabytes a year can be processed.
Equipped with thousands of Ascend AI processors, Huawei’s Atlas 900 AI cluster delivers 256 to 1,024 PFLOPS at FP16 – the equivalent of 500,000 PCs. This makes it the world’s fastest AI training cluster capable of quickly sifting through the data generated by the radio telescopes.
The cluster is applied to a sky map of the Southern Hemisphere which has some 200,000 stars. The SKA radio telescopes at two facilities in Australia and a sister site in South Africa will collect data from these celestial bodies and send the raw data – equivalent to six times the amount of data on the internet – into the Atlas cluster for analysis.
Diamond hopes such exceptional computing power will help “uncover the secrets of the cosmic dawn and the birth of stars and galaxies.”
The process will initially require around 50 petaflops of dedicated digital signal processing power before rising to 250 petaflops as demand increases. This will speed up analysis – it is expected to take just 10 seconds to locate and identify a specific star, compared to some 169 days usually taken by an astronomer. Scientists at the Shanghai Observatory have already completed a prototype SKA datacentre, demonstrating ways to store and access the data remotely.
But the power of the Atlas 900 is not confined to unlocking the mysteries of the universe. Through AI training, it can be used to recognise objects using image, video and voice data – some of the most difficult aspects of AI learning. These datasets can then be applied to other complex tasks such as weather forecasting, performing medical diagnoses and so on.
This may be one small step in technology, but a giant leap for mankind.