Expos at the NVIDIA AI Summit 2024 – India : US Pioneer Global VC DIFCHQ SFO India Singapore – Riyadh Swiss Our Mind

The NVIDIA AI Summit 2024, held from October 23 to 25 at the Jio World Convention Centre in Mumbai, marked a significant milestone in India’s journey toward becoming a global leader in artificial intelligence. This inaugural event brought together industry leaders, innovators, and technology enthusiasts to explore the transformative potential of AI across various sectors. Notable speakers included NVIDIA CEO Jensen Huang and Reliance Industries Chairman Mukesh Ambani, who engaged in discussions about India’s role in the AI landscape and the importance of building robust digital infrastructure.

NVIDIA AI Summit 2024 - India
Figure 1. NVIDIA AI Summit 2024 – India

The NVIDIA AI Summit – 2024 featured over 60 sessions, live demonstrations, and hands-on workshops focused on critical topics such as AI infrastructure development, foundational models for Indian languages, and fostering innovation within India’s vibrant startup ecosystem. As India positions itself as a hub for deep technology, this summit highlighted the collaborative efforts needed to harness AI’s capabilities to address pressing regional challenges while empowering the workforce through upskilling initiatives.

This article covers the presentation overview by Jensen Huang, the fireside chat with Mukesh Ambani, and some of the notable sessions that our team from OpenCV University attended at the event.

Key Highlights from Jensen Huang’s Presentation

During his presentation, Jensen Huang covered a range of topics that showcased NVIDIA’s advancements and strategic vision for AI, particularly in the context of India. Below are the key points discussed:

NVIDIA Accelerating Industry

NVIDIA CUDA accelerating different industries.
Figure 2. NVIDIA CUDA accelerating different industries.

NVIDIA’s GPU Acceleration: Huang emphasized NVIDIA’s software suite, which boosts AI capabilities across a range of applications. These tools are essential for optimizing performance and efficiency in machine learning and data processing. Highlights included:

  • cuDNN for deep learning
  • Modulus for AI Physics
  • cuOPT for decision optimization
  • cuDF for data processing
  • cuVS for vector search
  • Parabricks for gene sequencing
  • cuQuantum for quantum simulation
  • AI Aerial for radio signal processing
  • cuDSS for CAE (Computer-Aided Engineering)
  • cuLitho for computational lithography

Each tool is designed to support specific, high-demand workflows in AI and data science.

Evolution of Software

Software 1.0 and Software 2.0 powered by ML: The transition from traditional software development to machine learning-centric approaches was emphasized, illustrating how AI is reshaping software engineering paradigms.

Blackwell Architecture

Blackwell Powering Innovations: Huang introduced Blackwell, a revolutionary architecture combining seven chips designed to power NVIDIA’s foundational models and applications like Omniverse and CUDA libraries. This architecture represents a significant leap in computational capabilities.

Scaling Laws in AI

Shift in Scaling Laws: He discussed the implications of scaling laws on inference compute, particularly in relation to models like ChatGPT-o1, indicating how advancements in AI are driving increased computational demands.

NVIDIA’s Impact in India

NVIDIA Powering India: Huang outlined various partnerships with Indian entities such as CDAC, UPI, Flipkart, Reliance Jio Brain, and others. He stressed the importance of local collaborations to enhance AI infrastructure and capabilities.

Enterprise AI Solutions

NVIDIA Enterprise AI: The focus on creating intelligent agents through NVIDIA’s enterprise solutions was discussed, highlighting the role of NVIDIA NeMo in developing sophisticated AI models.

India’s Software Prowess

AI Software Development: Huang praised India’s talent and innovation in creating advanced AI software solutions, positioning the country as a leader in the global AI landscape.

Industrial Applications of AI

Physical Industrial AI: He introduced concepts of integrating Omniverse with DGX and AGX systems to enhance industrial processes across sectors such as manufacturing and automotive.

AI Factory Concept

Creating an AI Factory: Huang described a vision for an “AI Factory” that combines various industries – science, enterprise, healthcare, and manufacturing – projecting it as a $100 trillion industry opportunity.

Digital Twins and Simulation

Digital Twin Technology: The presentation concluded with a focus on digital twin technology and physics simulation for real-time monitoring and improved operational efficiency. This innovation aims to address challenges like development time and cost while enhancing risk mitigation strategies.

Huang’s insights not only showcased NVIDIA’s technological advancements but also reinforced India’s potential as a global hub for AI innovation.

Fireside Chat with Mukesh Ambani

Jensen Huang and Mukesh Ambani at the NVIDIA AI Summit - India: Fireside Chat.
Figure 3. Jensen Huang and Mukesh Ambani at the NVIDIA AI Summit – India: Fireside Chat.

During his fireside chat with Mukesh Ambani, Chairman of Reliance Industries, Jensen Huang emphasized that “NVIDIA is AI, and India is the epicenter of the IT industry,” highlighting the symbiotic relationship between NVIDIA’s technological advancements and India’s tech ecosystem.

Huang discussed the transformative impact of AI across various sectors, asserting that India is well-positioned to lead this revolution due to its rich talent pool and robust digital infrastructure. He noted that NVIDIA’s commitment to upskilling Indian IT professionals has already trained over 200,000 individuals in AI technologies.

Furthermore, he announced plans to significantly enhance India’s AI computing capabilities and compute power within a year.

The discussion also touched on the necessity of building large language models tailored to India’s diverse linguistic landscape, which Huang described as one of the most challenging yet promising opportunities for AI development. He concluded with a vision of collaboration between NVIDIA and Indian partners to foster innovation and solidify India’s position as a global leader in AI.

Notable Sessions from The NVIDIA AI Summit

In the following subsections, we will go through the overview of some of the sessions that the OpenCV University team attended during the two days of the summit.

Lip-Sync Model by Invideo Inc

Anshul Khandelwal, co-founder and CTO of Invideo Inc showcased their latest lip-sync model for generative AI and content creators.

Their frontier model uses NVIDIA’s StyleGAN (Generative Adversarial Network) to generate lip-sync videos. Trained on more than 2500 hours of data, the pipeline can automate the entire process of content generation including, video, audio, and syncing the lips with audio of the generated faces.

Lip Sync model of Invideo in action.
Figure 4. Lip Sync model of Invideo in action.

The speaker further discussed how they are scaling inference using Elixir and ONNX.

One of the highlights of the talk was the learnings from training such a model at scale which are summarized below:

  • At scale, most generative AI models eventually converge, given enough training data and iterations.
  • Dataset quality plays an important role.
  • At scale inference costs start to rise.

Overall, the talk went in-depth into the technical capabilities of training a generative and summarized the learnings well for someone who is willing to take up such a project.

Climate Science with Earth-2

Triggered by climate change, Climate Science has been one of the most debated topics in the last few years.

In this session, Jeff Larkin and Manish Modani from NVIDIA present Earth-2, a platform for next-generation weather and climate predictions.

Earth-2 combines four different models for global weather forecasting – FourCaseNet, GraphCast, Cordiff, and Pangu-Weather.

Different deep learning models available in the Earth-2 platform.
Figure 5. Different deep learning models available in the Earth-2 platform.

These tools help climate scientists predict and prevent damages and loss of lives from floods, hurricanes, and other natural disasters faster than ever before.

It is a boon for countries like India where each state can face adverse weather conditions several times a year.

Using ResDiff to visualize KM-scale climate data in high resolution.
Figure 6. Using ResDiff to visualize KM-scale climate data in high resolution.

High resolution kilometer scale visualizations help get a clearer picture of the approaching hurricanes and create climate simulations faster.

Powered by NVIDIA NIM for deployment and NVIDIA Modulus for fine-tuning, Earth-2 has become a game changer for weather forecasting, climate scientists, and architectures for building better cities.

Video World Foundation Models

In this session, Niket Agarwal from NVIDIA presented the entire AI lifecycle and stack involved in creating a video foundation model.

In the past few years, after LLMs, video generation models have gained much popularity. However, they are more complex to build because:

  • Models need to understand and comprehend videos.
  • Curating large labeled video datasets is time-consuming and resource intensive.
Video Foundation Models (VFM) are starting to emerge as one of primary applications of generative AI.
Figure 7. Video Foundation Models (VFM) are starting to emerge as one of the primary applications of generative AI.

The first step, video dataset curation, includes the following steps:

  • Extracting interesting and relevant clips from a video. This step uses an AI model that can recognize semantically relevant frames.
  • Using a second AI model for high quality data filtering for aesthetics and watermarks.
  • Finally, a good VLM is used for captioning the video clips.

The speaker pointed out how the entire data curation process involved 10s of AI models exchanging data between them in Terabytes scale. This leads to creating a Petabyte scale dataset that is finally used to train the video model.

Such intensity of data transfer and training demands 1000s of GPUs, specifically H100 and L40S GPUs.

The video curator architecture developed by NVIDIA to manage the Petabytes scale video data.
Figure 8. The video curator architecture developed by NVIDIA to manage the Petabytes scale video data.

The author ends the session by addressing some key challenges in the data pipeline which include:

  • The need for domain specific videos to teach the model varied concepts.
  • Removing duplicated content from the dataset.
  • Video understanding and creating an internal video search engine to enhance manual understanding of the dataset.

Indian LLMs for Global AI Leadership

Presented by Aakrit Vaish, this session included the current and future landscape of Indic LLMs.

Indic LLMs will power different sectors of the Indian economy.
Figure 9. Indic LLMs will power different sectors of the Indian economy.

The primary focus of this session was threefold:

  • LLMs are not predominantly trained on Indic languages as more than 50% of the internet data is in English.
  • High quality and filtered data is hugely lacking and important for training Indic LLMs from the ground up.
  • For the Indic language landscape, small language models for specific use cases will be more important than LLMs.

To cater to the first two issues, IndiaAI Mission (of which Aakrit is a member) is creating the IndiaAI platform for datasets. We can term this as the Hugging Face of Indian datasets. This led to a discussion of the large number of Indian startups that are working in the dataset curation domain.

https://learnopencv.com/nvidia-ai-summit-2024-india-overview/