Today, AMD shared their goal to increase the energy efficiency for their EPYC CPUs and Instinct accelerators for AI training and HPC by up to 30 times. HPC is an acronym for High-Performance Computing programs. They are processed through accelerated computational nodes. It is expected that the goal can not be completed before 2025, which includes AMD’s high-processing CPUs and GPU accelerators which they utilize for AI training, and HPC accelerated CPU configurations.
If AMD wants to achieve this goal, it will have to increase the power efficiency of its computational nodes at a higher rate. It has to be more than 2.5 times faster than the standard set by the industry over the last five years.
If you are wondering what exactly are Accelerated compute nodes. They are extremely powerful and most advanced systems in the world that are used for research and supercomputer tests that the majority of standard systems are not capable of processing. Scientists also use accelerated computational nodes to create discoveries and breakthroughs in several fields, like climate estimations and alternative energy solutions. They also play a part in AI where Accelerated compute nodes enable studies of neural networks studying like speech recognition, language translation, and expert recommendation systems. AMD’s plan will save several billions of kilowatt-hours of electricity by 2025.
AMD is also planning to use their “segment-specific data center power utilization effectiveness (PUE) with equipment utilization taken into account.” The company power consumptions for both CPUs and GPUs are set on specific segment utilization percentages, which include both idle and non-idle. The power consumption is multiplied by PUE to get energy use for the calculation of the performance per watt. AMD’s baseline for power consumption uses the industry energy rates that were calculated during 2015 to 2020 and are then “extrapolated to 2025.” The measure of Energy per operation improvement for the next 5 years comes through estimated global volumes which is then multiplied by the TEC, or Typical Energy Consumption to get each segment’s actual energy utilization worldwide.