CABLING
one of the factors driving the need for increased HPC infrastructure . AI models continue to proliferate in both size and number , with applications from scientific research to retail . Chinese researchers from Tsinghua University and Alibaba are training a 10bn parameter multi-modal model using NVIDIA ' s Megatron-LM training code , and Google recently trained a 1.6trnparameter AI language model . Cantrell warns that , “ Even though significant strides have been made towards more energysparing algorithms and models , all signs point to increasingly complex models and progressively sophisticated AI . This inevitably means a continued reliance on high intensity compute and , in turn , energy consumption .”
While data centres are getting more efficient all the time , the sheer amount of power required to support HPC workloads like these means that efficiency alone isn ’ t enough ; ultra-efficient designs need to be paired with site selections that give access to renewable power and the kinds of low ambient temperatures that further reduce electricity demand for cooling . “ No matter how efficient AI systems – and the data centres that house them – are , HPC and machine learning applications located in facilities powered by fossil fuels will always carry a carbon cost ,” says Cantrell . “ More and more , enterprises are looking for ways to mitigate the Scope 3 emissions generated by their externally hosted high intensity compute , and consequently , seek enhanced connectivity and more choice when it comes to location . New subsea cables like the IRIS system will be in greater demand , as they facilitate access to data centres powered by renewable energy – enabling organisations all around the world to process and store their data sustainably .” datacentremagazine . com 71