Descripción del proyecto
Comprender los límites de la inexactitud podría conducir a algoritmos más eficaces para la exaescala
Los cálculos matriciales son habituales en muchos campos. Nos acercamos a la era de la informática a exaescala —y de las matrices a exaescala—, que requerirá unos volúmenes de trabajo casi insondables. Una forma habitual de reducir este volumen de trabajo es introducir cálculos matriciales inexactos o aproximados. Sin embargo, para garantizar que la inexactitud procedente de múltiples fuentes que se propaga a través de los cálculos matriciales a exaescala no genere errores en las conclusiones que extraemos de los resultados, necesitamos un enfoque sistemático para analizar conjuntamente múltiples fuentes de inexactitud y estudiar su interacción. El equipo del proyecto inEXASCALE, financiado por el Consejo Europeo de Investigación, analizará rigurosamente la combinación de errores procedentes de múltiples fuentes y evaluará su efecto sobre el comportamiento numérico y la calidad de las soluciones en cálculos matriciales. Esto podría dar lugar a nuevos algoritmos para problemas a exaescala que aprovechen la inexactitud para equilibrar el rendimiento y la precisión.
Objetivo
Scientific computing inherently involves multiple sources of inexactness, from discretization or simplification of the problem, to noisy data, to finite precision rounding errors, to approximations made to increase parallelism, to stopping computations intentionally to improve efficiency. The standard state-of-the-art approach is to analyze different sources of error separately. There is currently no solid foundation or systematic approach for combining multiple sources of inexactness together and studying their interaction. Developing reliable approaches for exascale requires filling this gap, which must start with establishing a new rigorous foundation for analyzing multiple sources of error in matrix computations. Without this basis, the quest for efficiency in areas vitally depending on matrix computations, including, for example, data science and machine learning, will remain reliant on a trial-and-error approach.
This project aims to break the current modular approach to the analysis and design of algorithms for matrix computations by understanding how different sources of inexactness interact while being propagated through a computation and their effect on numerical behavior and solution quality. Our holistic approach, rooted in rigorous theoretical analysis, will reveal opportunities for developing new algorithms for exascale problems that exploit inexactness to balance performance and accuracy.
The project is structured around four fundamental objectives:
WP1: Analysis of exascale matrix computations subject to multiple sources of inexactness
WP2: Development of new algorithms that exploit inexactness that are both fast and provably accurate
WP3: Making error analysis of exascale computations meaningful in practice
WP4: Exploring emerging sources of inexactness beyond the exascale era
Our approach will lead to new methodologies that can change current paradigms.
Ámbito científico
Palabras clave
Programa(s)
- HORIZON.1.1 - European Research Council (ERC) Main Programme
Régimen de financiación
HORIZON-ERC - HORIZON ERC GrantsInstitución de acogida
116 36 Praha 1
Chequia