Today Lenovo announced that the company is collaborating with the STFC Hartree Centre in the U.K. on improving the energy-efficiency of HPC systems.
In this RCE Podcast, Brock Palen and Jeff Squyres discuss the Open Compute Project with Thomas Sohmers from Rex Computing. “Thomas Sohmers is the founder and CEO of REX Computing. His experience includes working at the MIT Institute for Soldier Nanotechnologies for 3 years as both an end user of HPC systems, and later transitioning into designing and building them at the lab. This experience led to starting REX Computing in 2013 as a recipient of the Peter Thiel ’20 under 20′ Fellowship, where he leads the architectural design and business operations.”
“The end of Dennard scaling has made all computing power limited, so that performance is determined by energy efficiency. With improvements in process technology offering little increase in efficiency innovations in architecture and circuits are required to maintain the expected performance scaling. The large scale parallelism and deep storage hierarchy of future machines poses programming challenges. Future programming systems must allow the programmer to express their code in a high-level, target-independent manner and optimize the target-dependent decisions of mapping available parallelism in time and space. This talk will discuss these challenges in more detail and introduce some of the technologies being developed to address them.”
Penguin Computing just announced the Altus Altus 2a30, a building block for the first application optimized accelerated processing unit (APU) clusters, making seamless GPU and CPU memory sharing on clusters a reality based on heterogeneous system architecture (HSA) from AMD. The shared memory capability involves very lightweight context switches to switch instantaneously between the GPU and CPU, whichever code runs best at a given moment.
“A crazy idea was born at ISC14 while answering questions about the new energy metrics in Allinea Performance Reports and Allinea MAP – could the information in these reports show us an easy way to increase energy efficiency without having to change the program? The idea was to use CPU frequency scaling to run memory-bound jobs at a lower clock frequency. In Lazy Energy Efficiency Challenge One, I found that on a synthetic benchmark I could increase the iterations per watt by 19% on a memory-bound code.”