Speaker
Description
Problem:
AI workloads are driving server memory capacity beyond the terabyte range, placing DRAM at the centre of both the hardware supply crisis and the environmental cost of modern computing [3]. Memory already accounts for 40–46% of total server energy consumption , rivalling or exceeding the processor [1, 2, 8, 9], especially for training and inference of large language models [5, 10]. In contrast, many sustainability studies across data-intensive research disciplines (e.g., genomics, astrophysics, climate modelling) still treat the processor as the dominant driver of carbon emissions, primarily because of the lack of a concrete method to account for DRAM’s energy [7]. Further, DRAM's environmental cost begins at the fab, and existing lifecycle assessments for research computing rarely capture DRAM's manufacturing-phase (embodied) emissions [4, 6] alongside its operational footprint [10].
Goal:
The goal of this interactive presentation is to give an in-depth look into how modern DRAM consumes energy and how we can start to reason about DRAM’s operational and embodied carbon emissions. I will first present a brief, discipline-agnostic introduction to how DRAM works and how it interacts with the processor, requiring no prior hardware knowledge. We will then look at how DRAM dictates modern computing system performance, and why memory has become a dominant power consumer in datacenters. We will next understand how to optimize applications to reduce the energy and carbon emissions associated with memory.
Expected Takeaway:
Audiences should expect to gain an understanding of how DRAM impacts the overall system’s performance, power, cost, and carbon emissions. Attendees will leave with a practical estimation methodology, applicable regardless of research domain, for accounting DRAM's operational and embodied carbon, directly supporting community efforts to build shared sustainability accounting standards. It is important to remember that efficiency gains alone are insufficient if workload growth continues to outpace them (Jevons paradox)[7].