October 2024

When Harry met Sally or When AI Met HPC

When Sally met Harry

At the beginning of the XXI century, in a world where technology and humanity are deeply intertwined, Dr Sally Albright, a brilliant computer scientist and biologist, was working on a groundbreaking AI project to solve complex global issues. In a twist of fate, a catastrophic lab accident involving an experimental quantum computer infused Sally with Harry the AI-based system she created, granting her extraordinary superpowers and transforming her into Quantum Mind.

All I want is you AI

HPAI in numbers

• 6 years old (est. 2017)
• 2 people (2017) 25 people (2024) + 15 new before 16/12/2024
• 7 researchers (2 visiting) 12 engineers 8 students
• ~3M€ gathered in competitive funding
• 5 private collaborations
• +50 papers +2,000 citations 2 AI PhDs 10+ AI Master

​

 Convergence of Computational Power and Intelligence�

• AI leverages HPC's massive processing capabilities.
• HPC clusters can distribute DL training tasks across hundreds or thousands of GPUs and CPUs working in tandem, dramatically reducing training time
• The vast computational resources of HPC allow for extensive hyperparameter optimization, improving DL models performance
• HPC enhances AI-based systems' ability to analyse complex datasets
• HPC systems offer high-speed computation and data throughput, enabling AI-based applications to analyse massive volumes of data quickly
• Synergy enables solving previously intractable problems
• In fields like astrophysics, climate science, and genomics, HPC allows AI-based systems to process and analyse intricate scientific simulations
• AI + HPC drive innovation across multiple scientific disciplines.
• This integration is not only accelerating existing research but also opening up entirely new areas of inquiry that were previously inaccessible due to computational limitations.

​

Computación (en FLOPs totales) requerido para el Entrenamiento de LLMs

Source: “Computing Power and the Governance of Artificial Intelligence” (2024) arXiv:2402.08797v1

• If you add up the LLMs with more than 0.5 trillion parameters, the total is 16.3 trillion parameters.

​

• Training a model with one trillion parameters costs approximately $937,411,765, and this does not include water expenses.

​

• An LLM with one trillion parameters, in English, requires 20.5 trillion tokens. There is insufficient unique information available to effectively train all these models, resulting in a high level of redundancy.

​

• There is no information on the cost of curating the data to train these models. It is known that they have done it in China

Transdisciplinary Collaboration�

• Brings together experts from diverse fields:
• Computer Science
• Mathematics
• Domain-specific sciences (e.g. biology, physics, medicine)

​

Transdisciplinary Collaboration�

• Brings together experts from diverse fields:
• Computer Science
• Mathematics
• Domain-specific sciences (e.g. biology, physics, medicine)
• Fosters new collaborative research paradigms
• Enables cross-pollination of ideas and methodologies

​

ALOE

13

13

Technological Integration

• Development of specialized hardware (e.g., GPUs for AI workloads)
• Creation of software frameworks bridging AI and HPC
• Advancement of hybrid computing systems (CPU + GPU)
• Containerization for scalable and portable AI/HPC applications

Data-Driven Scientific Discovery

• AI-powered analysis of massive scientific datasets
• Enhancement of traditional HPC simulations with AI-based techniques
• Acceleration of research in fields like:
• Genomics
• Climate science
• Astrophysics
• Enables new approaches to complex problem-solving

Synthetic Content Detection (images/videos)

• SuSy model
• SuperResolution and GenAI (Dalle-3, StableDiffusion, Midjourney)
• AI Act obliges
• Media, broadcasters, insurance, fraud in finance, …

​

[28]

Data quality

17

17

Why Data is like Water, Not Oil

• Data is an essential resource that ought to be high-quality and travel securely from the source system to the user and viceversa.
• Data governance is the springboard to a secure quality AI-based applications.
• Data Management to ingest only the cleanest data
• Data Security to protect the economy

​

Ethical and Societal Considerations

19

19

Ethical and Societal Considerations

• Raises important questions across multiple disciplines:
• Philosophy
• Law
• Social sciences
• Environmental issues
• Addresses challenges in:
• Data privacy, security and cybersecurity
• Ethical use of AI-based in decision-making
• Societal impacts of advanced computing technologies
• Necessitates development of new governance frameworks

Trustworthy AI a la americane

21

21

Some creative jobs maybe will go away, but�maybe they shouldn’t be there in the first place

To warm or not to warm (the Earth)?

Green Computing

• With 100 megatonnes of CO₂ emissions per year, similar to American commercial aviation, the contribution of data centers and high-performance computing facilities to climate change is substantial.
• So far, rapidly increasing demand has been paralleled by increasingly energy-efficient facilities, with overall electricity consumption of data centers somewhat stable.

​

AI critical important challenges

• The concentration of power in the hands of an oligopoly of those controlling today’s AI systems (Alphabet, Alibaba, Amazon, Meta, Twitter, etc).

��Ignorantia juris non excusat��

Some like it hot

28

28

Horizon 2050 (future was here)

• Last five years have been a turning point for AI and robotics. Everybody (Alibaba, Alphabet (aka Google), Amazon, IBM, Meta (aka Facebook), Microsoft, OpenAI etc.) buys AI.

​

• Machine Learning and Deep Learning are now routinely used for speech recognition, machine translation, robotic control, risk management, cybersecurity, etc.

​

• By 2025, McKinsey predicts an economic value of fifty Billion€ for AI-related technologies (already surpassed).

​

​

​

• Are these are good news for the world?

​

​

Sapere Audet��Thanks!

www.bsc.es

31

31