Talks

High-Performance Computing (HPC) drives discovery across science and industry and underpins the rapid advances in AI. At the heart of modern HPC platforms is the Graphics Processing Unit (GPU), which delivers the bulk of compute power but also dominates energy consumption. As GPU architectures increasingly prioritize low-precision arithmetic for AI workloads, HPC applications that depend on higher precision face new programmability challenges alongside new opportunities in mixed-precision computing. Crucially, the energy efficiency of GPU applications depends not only on compute utilization but also on memory traffic patterns, and the fastest implementation is not always the most energy efficient. Reliable exploration of these trade-offs is further complicated by the limited accuracy and temporal resolution of current power measurement tools. Combined with the vast, discontinuous design spaces inherent to GPU programming, manual optimization is infeasible. Automatic performance tuning, or auto-tuning, offers a proven approach to this problem, automatically searching for optimal configurations across algorithm, application, and hardware parameters. To address the emerging demands of mixed-precision computing and energy-aware execution, the field is now moving toward constrained and multi-objective optimization to enable systematic exploration of the trade-offs between performance, energy consumption, and numerical accuracy. In this talk, I will highlight key challenges, recent developments, and future directions in GPU auto-tuning.

00:00:00 Slide 1
00:02:23 Slide 2
00:02:58 Slide 3
00:03:44 Slide 4
00:04:27 Slide 5
00:06:59 Slide 6
00:07:42 Slide 7
00:10:02 Slide 8
00:11:03 Slide 9
00:12:09 Slide 10
00:13:00 Slide 11
00:14:16 Slide 12
00:15:10 Slide 13
00:15:34 Slide 14
00:16:18 Slide 15
00:17:21 Slide 16
00:18:45 Slide 17
00:20:38 Slide 18
00:21:00 Slide 19
00:22:27 Slide 20
00:23:58 Slide 21
00:24:43 Slide 22
00:25:28 Slide 23
00:26:26 Slide 24
00:27:57 Slide 25
00:28:46 Slide 26
00:29:37 Slide 27
00:31:48 Slide 28
00:34:09 Slide 29
00:36:38 Slide 30
00:38:46 Slide 31
00:39:02 Slide 32
00:39:36 Slide 33

Mitigation starts by treating technical debt and digital waste as sustainability problems. Idle compute, chatty microservices, oversized payloads, unbounded concurrency, always-on resilience, and forgotten data, environments, or duplicate workflows all become permanent excess: more kilowatt-hours, more emissions, more cloud spend. We uncover where energy is wasted in everyday systems, from CPU boost events with no user value and I/O waits misread as CPU issues to zombie instances and hidden digital waste across storage, pipelines, and tooling. These findings become a green debt backlog tied to measurable reductions in energy, cost, and operational heat. The talk reframes performance as value per watt. It explores event-driven designs, lean data contracts, carbon-aware compute placement, and CI/CD checks for efficiency. Attendees leave with practical habits to mitigate waste, innovate greener systems, and sustain gains long after the conference.

00:00:00 Slide 1
00:00:06 Slide 2
00:02:59 Slide 3
00:04:17 Slide 4
00:04:46 Slide 5
00:05:12 Slide 6
00:05:37 Slide 7
00:06:24 Slide 8
00:08:43 Slide 9
00:08:58 Slide 10
00:10:54 Slide 11
00:11:36 Slide 12
00:11:53 Slide 13
00:12:14 Slide 14
00:13:11 Slide 15
00:13:38 Slide 16
00:14:49 Slide 17
00:14:59 Slide 18
00:15:42 Slide 19
00:15:55 Slide 20
00:17:11 Slide 21
00:17:31 Slide 22
00:18:03 Slide 23
00:18:24 Slide 24
00:19:08 Slide 25
00:20:02 Slide 26
00:21:54 Slide 27
00:23:06 Slide 28
00:23:16 Slide 29
00:24:42 Slide 30
00:25:21 Slide 31
00:30:17 Slide 32

The proposed ISIS-II Neutron and Muon Source underwent a Life Cycle Assessment (LCA) during the early feasibility and design stage. The potential environmental impacts were evaluated across construction, operation, and decommissioning to identify and integrate environmental sustainability practises from its inception. As with most modern accelerators, computing is - and will be - essential for the design and operation of ISIS-II, yet predicting the computing impact for a facility that is proposed to run from 2040 to 2100 comes with its challenges. This work shares those challenges and explores the assumptions made to attempt to predict the required computing resources of a future accelerator.

00:00:00 Slide 1
00:00:35 Slide 2
00:01:12 Slide 3
00:02:37 Slide 4
00:03:45 Slide 5
00:04:00 Slide 6
00:05:50 Slide 7
00:07:46 Slide 8
00:08:18 Slide 9
00:09:52 Slide 10
00:11:11 Slide 11
00:12:15 Slide 12
00:13:20 Slide 13
00:14:26 Slide 14
00:14:45 Slide 15
00:15:46 Slide 16
00:17:30 Slide 17
00:18:24 Slide 18
00:18:57 Slide 19
00:19:34 Slide 20
00:20:47 Slide 21
00:22:23 Slide 22
00:27:30 Slide 23

In many science and engineering disciplines, limited representative training data, poor reproducibility and low interpretability hinder the complete integration of AI. The environmental impact of this growing technology is also of particular concern. Physics-informed machine learning (PIML) seeks to answer the aforementioned challenges, with many techniques leveraging physical knowledge to reduce training data requirements. We aim to explore how and when these approaches reduce model emissions. In this initial work, we embed physical insight into our ML models, assessing performance and emissions jointly on two simple benchmarks, one synthetic and one from an engineering lab experiment. The work demonstrates how a PIML approach can reduce emissions based on reduced training data, yet also highlights the increased model complexity from additional hyperparameters to be optimised, and the tradeoff between these factors to decrease overall emissions.

00:00:00 Slide 1
00:00:23 Slide 2
00:00:46 Slide 3
00:02:01 Slide 4
00:03:13 Slide 5
00:04:28 Slide 6
00:05:10 Slide 7
00:06:42 Slide 8
00:07:11 Slide 9
00:08:04 Slide 10
00:09:55 Slide 11
00:10:37 Slide 12
00:10:55 Slide 13
00:11:22 Slide 14
00:11:37 Slide 15
00:12:27 Slide 16
00:12:58 Slide 17
00:14:08 Slide 18
00:16:02 Slide 19
00:16:49 Slide 20
00:18:14 Slide 21
00:19:42 Slide 22

The Department of Aeronautics at Imperial College London is addressing the environmental and accessibility challenges of modern research by deploying a scalable repository architecture. This system integrates a custom InvenioRDM interface with Ceph object storage to manage massive computational datasets in alignment with FAIR principles. By leveraging software-defined storage on commodity hardware, the department avoids carbon-intensive "forklift upgrades," allowing for sustainable, incremental capacity growth. The infrastructure features a self-healing, S3-compatible backend designed to eliminate "dark data" through domain-specific metadata curation. To reduce energy consumption associated with unnecessary network egress, the platform supports flexible retrieval, enabling researchers to inspect granular data subsets rather than downloading entire multi-terabyte files. Ultimately, this ecosystem prevents redundant, energy-heavy re-computations by transforming primary data into a permanent, reusable asset. The proposed talk details the lifecycle of this transition from ad-hoc management to an integrated, environmentally conscious research framework.

00:00:00 Slide 1
00:00:01 Slide 2
00:00:32 Slide 3
00:01:41 Slide 4
00:02:48 Slide 5
00:04:28 Slide 6
00:06:12 Slide 7
00:08:01 Slide 8
00:08:50 Slide 9
00:09:38 Slide 10
00:10:23 Slide 11
00:10:43 Slide 12
00:11:13 Slide 13
00:13:31 Slide 14
00:14:44 Slide 15
00:16:34 Slide 16

Astronomers have a unique perspective on the Earth, its fragility, and the absence of a 'planet B'. Yet the carbon footprint of their activities and instruments remains substantial. The amount of data generated by these increasingly precise technical instruments, and the subsequent computing infrastructure needed to process and store it, poses a key challenge to sustainability. The Wide-field Spectroscopic Telescope (WST) is a 12-meter spectroscopic facility currently under development and potentially operational in Chile in 2040. One of its key science cases is the so-called 'time domain astrophysics', which requires rapid follow-up observations of transient sources in the night sky. The WST will be equipped with a provisional number of ~600 detectors, which will gather data continuously every night over an expected lifetime of 50 years. The expected data volume ranges from 1 to 3 PB per year. In this talk, we will present the first estimates of the carbon footprint associated with the data reduction pipeline and storage infrastructures. We will highlight the WST's ongoing effort to incorporate sustainability considerations into hardware selection for data processing and storage, and explore how location affects the carbon footprint of the different solutions. We will also discuss how integrating sustainability early in the design process of research infrastructures can effectively mitigate the environmental impact.

00:00:00 Slide 1
00:00:29 Slide 2
00:01:38 Slide 3
00:02:08 Slide 4
00:02:26 Slide 5
00:03:03 Slide 6
00:03:46 Slide 7
00:05:04 Slide 8
00:06:00 Slide 9
00:07:03 Slide 10
00:07:26 Slide 11
00:08:59 Slide 12
00:09:31 Slide 13
00:10:33 Slide 14
00:11:14 Slide 15
00:11:56 Slide 16
00:12:56 Slide 17
00:13:31 Slide 18
00:14:56 Slide 19
00:15:32 Slide 20
00:17:02 Slide 21
00:17:52 Slide 22
00:18:44 Slide 23

Time- and location-shifting of computational workloads is widely proposed to reduce data-centre emissions by exploiting variation in electricity carbon intensity. However, CO$_2$-only optimization can shift burdens to places where impacts are experienced locally, such as water withdrawals in stressed basins, worsened air-pollution exposure, and increased stress on constrained grids. We present Orca, a sustainability-aware workload shifting framework that jointly considers global climate impacts and heterogeneous local criteria. Orca integrates region- and time-dependent signals for carbon, water-stress--weighted water use, air-pollution exposure proxies, and grid-stress indicators, and formulates scheduling as a multi-objective optimization problem. Using Pareto analysis, preference weighting, and optional impact caps, Orca exposes and mitigates trade-offs between emissions reduction and local burdens. A three-region case study shows that CO$_2$-optimal shifting can worsen local outcomes, while Orca produces context-sensitive schedules that better balance global and local sustainability objectives.

00:00:00 Slide 1
00:00:32 Slide 2
00:01:02 Slide 3
00:01:19 Slide 4
00:01:43 Slide 5
00:01:58 Slide 6
00:02:26 Slide 7
00:02:44 Slide 8
00:03:06 Slide 9
00:03:19 Slide 10
00:04:28 Slide 11
00:05:11 Slide 12
00:06:30 Slide 13
00:08:04 Slide 14
00:09:20 Slide 15
00:09:31 Slide 16
00:10:07 Slide 17
00:10:22 Slide 18
00:11:06 Slide 19
00:12:53 Slide 20
00:13:49 Slide 21
00:14:34 Slide 22
00:15:49 Slide 23
00:17:52 Slide 24
00:18:45 Slide 25

No description

00:00:00 Slide 1
00:00:37 Slide 2
00:02:20 Slide 3
00:03:48 Slide 4
00:05:56 Slide 5
00:06:44 Slide 6
00:07:45 Slide 7
00:09:18 Slide 8
00:10:52 Slide 9
00:12:52 Slide 10
00:14:51 Slide 11
00:15:53 Slide 12
00:17:26 Slide 13
00:19:10 Slide 14
00:20:44 Slide 15
00:21:01 Slide 16
00:22:10 Slide 17
00:22:34 Slide 18
00:23:04 Slide 19
00:23:27 Slide 20
00:23:41 Slide 21
00:23:53 Slide 22
00:24:06 Slide 23
00:24:24 Slide 24
00:24:37 Slide 25
00:24:48 Slide 26
00:25:17 Slide 27
00:26:37 Slide 28
00:28:02 Slide 29
00:28:42 Slide 30
00:30:13 Slide 31
00:31:32 Slide 32
00:32:26 Slide 33
00:32:53 Slide 34
00:33:05 Slide 35
00:33:15 Slide 36
00:33:46 Slide 37
00:34:01 Slide 38

Sustainability and software engineering are often treated as separate concerns. They do intersect. One often focuses on energy, carbon, and environmental limits. The other focuses on code, systems, longevity, reproducibility, and delivery. This talk asks if the link runs deeper. Both fields face similar pressures. Demand rises. Resources are finite. Complexity grows. Hidden costs accumulate. I draw on work in network systems, media, broadcasting, open source, mentoring, platform-scale systems, and research computing. If so, better software engineering is not merely adjacent to sustainability. The two can be symbiotic. Through projects and examples, the session explores how to build systems that scale better and waste less. It also asks how systems can create more capable participants. Finally, I consider what opportunities lie beyond net zero, and how we build for that world.

00:00:00 Slide 1
00:00:14 Slide 2
00:01:29 Slide 3
00:02:13 Slide 4
00:02:31 Slide 5
00:02:54 Slide 6
00:03:13 Slide 7
00:03:32 Slide 8
00:07:38 Slide 9
00:09:10 Slide 10
00:14:17 Slide 11
00:15:28 Slide 12
00:17:21 Slide 13
00:19:32 Slide 14
00:21:11 Slide 15
00:23:15 Slide 16
00:26:16 Slide 17
00:27:20 Slide 18
00:32:31 Slide 19
00:33:06 Slide 20
00:35:15 Slide 21