Connect With the Experts at GTC 2025
Get answers to your most pressing questions from NVIDIA technology experts.
Connect With the Experts sessions are a unique opportunity to meet in person with the people behind NVIDIA products and research at an interactive area of the San Jose Convention Center.
Each 50-minute session will give you a chance to get your questions answered and learn about the latest NVIDIA technologies—including generative AI, agentic AI, data center, robotics, data processing, accelerated computing, and more.
A Selection of Session Topics
We'll provide a brief overview of the NVIDIA Nsight™ family of tools and debuggers. You’ll learn about the latest features for debugging, profiling, and optimizing NVIDIA® CUDA® compute and graphics applications on the newest platforms. Several experts from the tools development and management teams will be available to talk about getting started, best practices, and advanced techniques for application debugging and optimization with NVIDIA developer tools. You can ask questions about anything from high-level concepts to specific tools details, and also discuss plans and suggestions for improving the tools in the future.
Key Takeaways:
- Join Connect With the Experts sessions that give you a unique opportunity to meet with the minds behind NVIDIA’s products and research, in either a group or 1:1 setting, to get your questions answered.
- Learn about CUDA and graphics developer tools, including what's new and helpful tips and tricks.
- Network with fellow developers in your technology area to share and learn from one another.
Join us and learn about quantum computing at NVIDIA! We'll show you how to use CUDA-Q and cuQuantum to accelerate your quantum computing research, from simulating algorithms and processors to running hybrid quantum-classical workloads. In this interactive session, we'll discuss how to take full advantage of NVIDIA’s most powerful GPUs to explore the benefits of quantum computing. Join NVIDIA quantum software experts to learn more about the CUDA-Q open-source platform for programming quantum processing units (QPUs), GPUs, and CPUs seamlessly, and how it integrates with cuQuantum to accelerate applications simulation. Discuss your ideas with the team and give us feedback on your experience using NVIDIA’s quantum computing software.
Key Takeaways:
- Learn to develop hybrid quantum-classical applications using CUDA-Q.
- Learn how to accelerate your quantum computing simulations using cuQuantum.
- Discover how to effortlessly switch between accelerated simulations and physical QPU runs in CUDA-Q.
- Learn how to incorporate noise into your quantum circuit simulations using CUDA-Q.
- Discuss ideas and give feedback to NVIDIA Quantum team members.
Join NVIDIA experts to explore the field of digital health. Discuss how NVIDIA NIM™ microservices and blueprints empower developers, researchers, and enterprises to quickly create generative AI solutions in healthcare incorporating clinical documentation and agentic health workflows. Dive into the latest advancements in generative AI foundational models, learning about their applications in understanding, summarization, and how they intersect in real-world clinical applications. And engage directly with our product managers, developer relations, and solution architects in discussions that address your biggest challenges.
Key Takeaways:
- Learn how to accelerate digital health solution development with technologies like NVIDIA NIM and Blueprints.
- Understand the role of generative AI and its numerous applications in understanding and summarizing clinical data.
- Connect the technologies to real-world clinical applications and scenarios.
- Engage directly with NVIDIA product managers, developer relations, and solution architects to discuss challenges and find solutions.
Real-time edge AI and robotic systems are transforming medical workflows and creating revolutionary devices that are advancing the standard of care for disease detection, diagnosis, and interventions. NVIDIA's platforms—MONAI, Holoscan, Isaac™, Omniverse™, and Metropolis—span model training, digital twins/simulations, and runtime deployment. Together, they give developers an end-to-end, accelerated, full-stack infrastructure that enables rapid prototyping and scalable development of real-time AI applications. Learn from NVIDIA experts on how our platforms can save you time and money in bringing AI-enabled systems to market—from smart diagnostic devices and innovative imaging systems to surgical robots trained using digital twin technologies.
Key Takeaways:
- Learn about the range of NVIDIA platforms capabilities for developing and deploying a diversity of accelerated, edge-AI medical applications.
- Understand the architecture of Holoscan hardware to enable low-latency, edge-AI workflows for multi-sensor applications.
- Explore how a three-computer combination for training, simulation, and robotic control provides a framework for the future of digital surgery.
Do you want to learn how to get started accelerating large language models and other state-of-the-art deep learning models? Join this interactive support session with the cuDNN engineering staff to get direct help and surface your questions, issues, and ideas specific to your use case.
Key Takeaways:
- Learn how to use cuDNN's state-of-the-art Flash Attention variants.
- See how to use cuDNN's Frontend for improved ease of use and development time.
- Understand cuDNN's Graph API for fusions.
Visit us to discuss ways to get started or deepen your understanding of NVIDIA tools that support vision AI applications. Our topic experts are offering one-on-one dialogue on topics related to training, fine-tuning, and deploying a range of vision AI models. Ask about NVIDIA Metropolis, VLMs, foundation models, NIM microservices, reference workflows, DeepStream, TAO, TensorRT™, Triton™ Inference Server, Jetson™, and more.
Key Takeaways:
- Get started with the latest in VLMs and foundation models.
- Talk through barriers you might be facing with visual AI projects.
- Learn how to quickly incorporate generative AI and NIM into Metropolis deployments.
Join NVIDIA experts to explore the field of digital biology. We’ll discuss the latest advancements in accelerated computing and foundation models for drug discovery and genomics—from bioinformatics tooling to protein structure prediction, molecular dynamics to molecular generation, and much more. Learn how NVIDIA BioNeMo™ and Parabricks, together with NVIDIA NIM microservices and blueprints, empower developers, researchers, and enterprises to quickly create generative AI solutions across chemistry, biology, and genomics. Engage directly with our product managers, developer relations, and solution architects in discussions that address your challenges and seek answers to your questions.
Key Takeaways:
- Learn how the BioNeMo framework enables the development and training of large-scale AI models for drug discovery.
- Learn how to accelerate your drug discovery process through the BioNeMo NIM and NIM Agent Blueprints.
- Engage directly with NVIDIA product managers, developer relations, and solution architects to discuss challenges and find solutions.
- Discover how to reduce time to insights for genomics data using Parabricks, RAPIDS™ single-cell, and VISTA-2D.
The IDC has predicted that by 2025, 80% of all data will be unstructured. Join us for an interactive session and learn how you can use GPUs to perform semantic search and data mining at scale on embeddings created from unstructured data. We'll dive into the cuVS library and provide an overview of the various databases and libraries that have integrated it to achieve state-of-the-art performance for building and searching vector indexes.
Key Takeaways:
- See how cuVS provides state-of-the-art performance for vector search applications on the GPU.
- Learn how cuVS powers several vector search libraries and databases.
- See how vector search indexes can be built up to 20X faster on the GPU and searched up to 10X faster than with CPUs.
Did you know GPUs can reduce the cost of your Apache Spark workloads? Come join this session to find out how. Whether you’re new to Apache Spark or a long-time user, we're here to have a conversation about large-scale data processing, your production workloads, optimal configurations, and how to introduce GPUs to your environment. We can also answer questions about job qualification and tuning with GPUs. Get questions answered about potential performance gains, cost savings, or power usage reduction possible with GPU-accelerated Spark. Learn about the migration journey others have already undertaken.
Key Takeaways:
- Learn how batch ETL workloads can be accelerated with GPUs for cost savings, performance, and data center footprint power reduction.
- See the tooling available to help you identify workloads that are best suited for GPUs.
- Learn about the differences between GPU types and how RAPIDS Spark workloads benefit from each one.
- Discover how organizations with petabytes of data and thousands of workloads have undertaken this migration. Hear about the roadmap for RAPIDS Spark.
Join this engaging session led by NVIDIA experts who collaborate closely with the global academic research community, supporting leading researchers in using NVIDIA software technologies to drive scientific advancements across diverse fields. We'll provide insights into how various NVIDIA teams, representing different functional areas, can assist in accelerating your scientific discoveries at multiple levels. The team will outline their approach to working with individual researchers and groups, share success stories, and introduce a range of developer support programs, including but not limited to the Academic Grant Program, Deep Learning Institute courses and Teaching Kits, and Researcher Advisory Councils. We'll also focus on aligning discussions with specific research domains such as robotics, biology, generative AI, data processing, and climate science—while showcasing tools and resources available to help researchers enhance research outcomes.
Key Takeaways:
- See how NVIDIA helps the academic research community adopt NVIDIA software technologies.
- Learn about how different teams at NVIDIA can assist in accelerating your scientific discoveries on multiple levels.
- Learn about various supporting programs NVIDIA offers today for research communities.
- Discover various frameworks and tools that can help researchers accelerate their research outcomes across diverse fields of science.
- Learn how to empower your lab quickly in AI knowledge and beyond.
Just as critical to the performance of AI model training, tail latency can negatively affect real-time AI inference, financial trading, and large-scale data analytics, where even microsecond delays can affect outcomes. The approach to zero-tail latency and accelerated networking with enhanced programmable I/O represent a paradigm shift in the performance and scalability of modern computing infrastructures. This session will open the discussion to dive into the significance of accelerated networking technologies in delivering zero-tail latency, ensuring that every data transfer meets demanding performance requirements even beyond 100k GPU scale.
Key Takeaways:
- Achieve zero-tail latency to ensure that all requests, even those in the "tail" of the distribution, experience minimal delays, leading to more consistent performance, reducing jitter, and enabling faster decision-making processes.
- Learn best practices for building large-scale networks for AI.
- Understand how networking impacts the performance of various AI workloads.
- See how accelerated networking, when coupled with programmable I/O, enables dynamic optimization and offloading of tasks directly to network hardware, dramatically reducing overhead and enhancing throughput.
- Learn about advanced adaptive routing to load-balance network traffic at the packet level and telemetry-based congestion control, enabling the highest effective bandwidth for the AI fabric.
Join this interactive session to deep dive into the world of large language model (LLM) inference. Learn about techniques to reduce computational overhead and discover strategies for scaling your applications to handle high-concurrency demands. Whether you’re exploring best practices for application deployment with a NIM or low-level model optimizations with Tensor-RT-LLM to accelerate inference and minimize latency, this session will cover it all.
Key Takeaways:
- Application Deployment: Learn how to seamlessly integrate LLMs into your infrastructure using a NIM and discover strategies for scaling your applications to handle high-concurrency demands.
- Performance Optimization: Explore low-level model optimizations using state-of-the-art advanced techniques with TRT-LLM.
- Inference Workflows: Learn about NVIDIA's reference workflows using NIM Agent Blueprints.
Next-generation robots will need to sense, plan, and act autonomously. A "sim-first" approach is required to ensure that the entire robot stack performs. Connect with NVIDIA experts to learn how you use NVIDIA Isaac Sim™ for your robotics workflows, including generating synthetic data for perception model training and software-in-loop testing to validate your robot stack.
Key Takeaways:
- Learn about synthetic data generation for training AI models.
- Explore software-in-the-loop for validating robot stack.
- See how using generative AI can speed up scene creation, data generation, and scenario testing.
- Learn how to connect up ROS nodes with Isaac Sim for robot validation.
We'll introduce the newest TensorCore programming guide, explaining how to build up NVIDIA Blackwell architecture TensorCore GEMM, use new low-precision TensorCore, map the above concepts to CUTLASS, and better use CUTLASS. We'll also share techniques to optimize memory bandwidth and maximize latency hiding. A straightforward rule of thumb can predict the bandwidth use of a kernel, simplifying the allocation of staging buffers in shared memory.
Key Takeaways:
- Learn how to program with the new Blackwell Tensor Core.
- See how the new TMA instructions and persistent mode work with the Blackwell Tensor Core.
- Learn how the memory subsystem has changed in the latest NVIDIA GPUs, including Blackwell.
- Discover a simple rule of thumb to estimate the memory throughput of a kernel.
- Explore several techniques to make the most out of the increased bandwidth in the latest GPUs.
Deploying AI models for real-world applications demands efficient inference strategies to handle their intense memory and compute requirements. Learn the key phases of LLM inference and provide practical techniques for developers and data scientists to optimize model performance, reduce costs, and achieve scalable deployment. We'll also cover state-of-the-art attention mechanism improvements, such as multi-query and flash attention, to optimize memory use during inference. Additionally, we'll explore model-serving strategies like in-flight batching and speculative inference, which can significantly enhance GPU utilization while minimizing latency in production environments.
Key Takeaways:
- Understanding AI Inference: Deep dive into the prefill and decode phases of LLM inference and their impact on GPU utilization and performance.
- Advanced Model Serving Strategies: Discover methods such as in-flight batching and speculative inference to boost GPU utilization and minimize latency during deployment.
- Efficient Inference Techniques: Understand how quantization and sparsity can reduce memory footprints and accelerate LLM inference for production environments.
Compute for humanoid robotics is crucial to deliver optimized performance and advanced AI capabilities. NVIDIA Jetson™ Thor for humanoid robotics features the next-gen NVIDIA Blackwell architecture with integrated safety features and high-performance computing to ensure safe and efficient interactions in real-world environments. Connect with our experts to understand how to unlock power and performance for humanoid robotics applications from software, AI models, and libraries to compute. Join us to accelerate your humanoid robotics projects and drive innovation in embodied AI.
Key Takeaways:
- Unlock unprecedented performance for humanoid robotics applications using Jetson Thor.
- Learn about NVIDIA's extensive robotics platform, including advanced software tools, AI models, and libraries specifically designed for humanoid robotics.
- Accelerate development cycles and drive innovation in embodied AI applications using Jetson Thor.
NVIDIA Grace™ Hopper and Grace Blackwell are a new breed of memory-converged CPU-GPU superchips engineered to meet the challenges of AI, high-performance computing, and data processing. Learn about the unique architecture of these superchips and their versatile form factors that simplify data center deployment. We'll address such questions as how the 900 GB/s NVLink™ chip-to-chip interconnect enables the CPU and GPU to access all system-allocated memory, how you can seamlessly migrate your x86-based workloads to the Arm-based Grace CPU, and the performance gains you can expect to unlock for your applications.
Key Takeaways:
- Understand how the NVIDIA GH200 Grace Hopper Superchip combines an Arm-based Grace CPU with a Hopper GPU over a high-bandwidth and memory-coherent NVLink chip-to-chip interconnect.
- Explore programming models, performance libraries, compilers, debuggers, and profiling tools to accelerate scientific computing workflow performance while realizing power and energy efficiency.
- Get expert advice on using open-source libraries—including RAPIDS for data science and machine learning and the PyTorch ecosystem for deep learning on Grace Hopper—across a broad range of academic, research, and industry use cases.
- See how to use Grace Hopper platforms as a path for deploying applications on Grace Blackwell systems.
Get unique insights from NVIDIA's Kaggle Grandmasters, who have competed and earned top honors in hundreds of international data science competitions. The team brings deep expertise in generative AI, computer vision, natural language processing, forecasting, recommender systems, tabular data, signals, and graphs. Discover the latest tools and techniques winning competitions today, hear best practices for GPU acceleration across diverse use cases, and explore how these experts approach complex challenges. Bring your questions and learn directly from data science champions on optimizing model performance for real-world applications.
Key Takeaways:
- Learn directly from data science champions on optimizing model performance for real-world applications.
- Discover the latest tools and techniques winning data science competitions today.
- Hear best practices for GPU acceleration across diverse use cases.
- Explore how Kaggle Grandmasters from NVIDIA approach complex challenges.
