JUPITER Advanced Research Vehicle for Inference Services

Descripción del proyecto

Acelerar la inteligencia artificial y prepararla para el mundo real

La inteligencia artificial está pasando del reconocimiento al razonamiento. Entrenar modelos de inteligencia artificial potentes para razonar, responder y adaptarse a escala es el siguiente paso en la evolución histórica de la inteligencia artificial. La cuestión principal ya no es lo rápido que puede aprender un modelo, sino lo eficiente que puede pensar. El proyecto JARVIS, financiado con fondos europeos, tiene como objetivo salvar esta brecha. En concreto, trabajará con JUPITER, que es el superordenador más potente de Europa y el cuarto más rápido del mundo. El proyecto proporcionará una plataforma de alto rendimiento específica para la fase en la que un modelo de inteligencia artificial realiza realmente las tareas. El objetivo es construir una inteligencia artificial de nueva generación que no solo sea más rápida y eficiente energéticamente, sino que también esté preparada para las exigencias del mundo real.

Objetivo

JAIF will serve customers on the exascale supercomputer complex around JUPITER, scheduled to become operational in mid-2025. JUPITER is poised to be one of the strongest AI training machines in the world that allows, due to its superfast connectivity, full parallelism of training models on thousands of nodes. JAIF will help exploit JUPITER's outstanding capabilities and capacities, and leverage its transformative power for AI training.

In addition, the call offers the exciting opportunity to implement an experimental supercomputing platform for AI calculations. JAIF proposes to implement JARVIS, the JUPITER Advanced Research Vehicle for Inference Services. In JARVIS, a variety of experimental technological components of European origin will be integrated. JARVIS operates as a high-performance and highly scalable inference system that ideally complements JUPITER's AI training capabilities by achieving maximum throughput in terms of AI inference accessed through industry-standard cloud technologies.

JARVIS’ features are becoming very important for types of foundation models like largest language models (LLM) or reinforced LMMs like OpenAI’s o1 or DeepSeek-R1. The recent emergence of inference techniques, referred to as Inference Time-Scaling or as Test-Time Compute will require much larger scalable inference server capacities as the inference cycles become much longer due to the simulated reasoning iterations. In contrast to JUPITERs fully scalable connectivity, JARVIS will feature many islands of tightly connected nodes comprising between 4 and 16 large-memory nodes given current model sizes and those anticipated for the coming years. It is expected that this trend in AI algorithms will make the direct combination and tightest integration of training and inference systems on a modular hardware basis virtually essential.