Datacentres have become a backbone of the modern economy, they currently consume 2% of worldwide electricity, rising to 10% by 2030. While compute accounts for 40% of this energy, 80% of this compute energy is related to moving data between the main memory chips and the processor (CPU), a structural bottleneck within server design that causes slowdown and low CPU usage. Development of ground-breaking Processing-In-Memory (PIM) technology by UPMEM prevents this data movement by performing calculations in the memory chips, where the data resides. By eliminating the need to move data off chips, PIM bypasses structural bottlenecks. The power of UPMEM’s technology comes from leveraging existing industry protocols, chip technology and programing languages, and designing PIM modules that fit into standard memory slots. When compared to those using conventional memory, PIM equipped servers have proven to be up to 20x faster, 10x more efficient and cheaper for Big Data and AI applications.

Integrated digital diagnostics can support complex surgeries in many anatomies where brain tumour surgery is one of the most complex cases. Neurosurgeons face several challenges during brain tumour surgeries, such as critical tissue and brain tumour margins differentiation or the interpretation of large amount of data available provided by several independent devices. To overcome these challenges, STRATUM will develop a 3D Decision Support Tool for brain surgery guidance and diagnostics (reaching TRL7) based on multimodal data processing through Artificial Intelligence (AI) algorithms that will be integrated as an energy-efficient Point-of-Care computing tool. It will be developed following a co-creation methodology involving key stakeholders and end-users. STRATUM will pursue the following objectives: 1) To foster advances in personalized medicine based on multimodal data (including the emerging hyperspectral imaging modality) and AI. 2) To increase the intraoperative diagnostic accuracy of brain tumours, improving surgical outcomes and patients’ quality of life. 3) To reduce surgery time with respect to current neurosurgical operation durations. 4) To improve current cost- and energy-efficiency of neurosurgical workflows. 5) To demonstrate the prototype in a two-year clinical study in 3 clinical sites, including an early health technology assessment. 6) To prepare the preliminary business plan and the TRL9 roadmap after the project ending. An optimized integration and processing of available and new emerging data sources would aid surgeons in timely efficient and correct decision-making in tissue removal. This would maximize the degree of resection while simultaneously minimize the risk of neurological deficits. Moreover, time efficient surgical procedures not only benefit the patients directly by minimizing anaesthesia time and risks of e.g. postoperative infections, but also indirectly by optimizing resources of the health care system.

Low cost, high throughput DNA and RNA sequencing (HTS) data is now the main workforce for various genomics and transcriptomics applications. HTS technologies have already started to impact a broad range of research and clinical use for the life sciences. These include, but are not limited to 1) large-scale sequencing studies for population genomics and disease-causing mutation discovery including cancer, 2) metagenomics, 3) comparative genomics, 5) transcriptome profiling, and 6) outbreak detection and tracking including COVID-19, Ebola, and Zika. HTS also impacts the whole health care system in several directions. Although there is still much room for improvement, sequencing of personal genomes is now becoming a part of preventive and personalized medicine as HTS technologies make it possible to 1) identify genetic mutations that enable rare disease diagnosis, 2) determine cancer subtypes therefore guiding treatment options, and 3) characterize infections and antibiotic resistance. Currently all genomics data are processed in energy-hungry computer clusters and data centers, which also necessitate the transfer of data via the internet, which also consumes substantial amounts of energy and wastes valuable time. Therefore there is a need for fast, energy-efficient, and cost-efficient technologies that enable all forms of genomics research without requiring data centers and cloud platforms. In this project we aim to leverage the emerging processing-in-memory (PIM) technologies to enable such powerful edge computing. We will focus on co-designing algorithms and data structures commonly used in bioinformatics together with several types of PIM architectures to obtain the highest benefit in cost, energy, and time savings. BioPIM will also impact other fields that employ similar algorithms. Our designs and algorithms will not be limited to cheap hardware, and they will impact computation efficiency on all forms of computing environments including cloud platforms.

AI is increasingly becoming a significant factor in the CO2 footprint of the European economy. To avoid a conflict between sustainability and economic competitiveness and to allow the European economy to leverage AI for its leadership in a climate friendly way, new technologies to reduce the energy requirements of all parts of AI system are needed. A key problem is the fact that tools (e.g. PyTorch) and methods that currently drive the rapid spread and democratization of AI prioritize performance and functionality while paying little attention to the CO2 footprint. As a consequence, we see rapid growth in AI applications, but not much so in AI applications that are optimized for low power and sustainability. To change that we aim to develop an interactive design framework and associated models, methods and tools that will foster energy efficiency throughout the whole life-cycle of ML applications: from the design and exploration phase that includes exploratory iterations of training, testing and optimizing different system versions through the final training of the production systems (which often involves huge amounts of data, computation and epochs) and (where appropriate) continuous online re-training during deployment for the inference process. The framework will optimize the ML solutions based on the application tasks, across levels from hardware to model architecture. AI developers from all experience levels will be able to make use of the framework through its emphasis on human-centric interactive transparent design and functional knowledge cores, instead of the common blackbox and fully automated optimization approaches in AutoML. The framework will be made available on the AI4EU platform and disseminated through close collaboration with initiatives such as the ICT 48 networks. It will also be directly exploited by the industrial partners representing various parts of the relevant value chain: from software framework, through hardware to AI services.