Energy and computing : Identifying and addressing bottlenecks
Place : IRIG/SPINTEC, auditorium 445 CEA Building 10.05 (access to CEA requires an entry authorization. Request it before 08th at admin.spintec@cea.fr)
video conference : https://univ-grenoble-alpes-fr.zoom.us/j/98769867024?pwd=dXNnT3RMeThjYStybGVQSUN0TVdJdz09
Meeting ID: 987 6986 7024
Passcode: 025918
Abstract : From low-power IoT devices to high-performance computing nodes, energy is often a key consideration. We face multiple significant energy-related challenges: availability, cost, and sustainability, to name only the most frequent. Understanding how the energy is spent is crucial for addressing bottlenecks in both software and hardware aspects. This presentation is twofold. First, we present a low-power audio recording device, a first high-level energy breakdown shows that storage is responsible for half of the total energy usage. Under these conditions, we implement and compare multiple data reduction strategies. Secondly, we take a look into high-performance systems. We propose a methodology to evaluate the energy consumption of GPUs at the instruction level. We base our methodology on measurements provided by a built-in sensor and performance counters reported by profiling tools. Our results show that data movement across the cache hierarchy dominates the energy usage.
Biography : I defended my PhD in December 2023 in Montpellier on the subject of embedded computing for wildlife monitoring under severe energy constraints. I followed with a one-year ATER contract to study inference acceleration around both high-performance computing systems (with GPUs) and low-power devices (with ADAM, a RISC-V based platform). I now occupy an assistant professor position at Phelma and Spintec. I mainly teach HDL and programming, and my research will focus on emerging computing paradigms, where I hope to contribute on the subject of system-level integration of smart memories.