The much anticipated Developer Conference for the Era of AI is off to a brilliant start. Every year NVIDIA hosts a hugely popular conference called GPU Technology Conference (GTC) to share the latest AI technology with developers, engineers, and researchers. GTC is a great opportunity to discover and explore transformations in the industry driven by the power of AI, collaborative virtual worlds, and beyond. Like previous years, we expect some exciting announcements this year. If you are passionate about AI, Computer Vision, and Deep Learning, you will find many many interesting sessions at GTC. If you haven’t already done so, there is still time to subscribe and register for FREE for a chance to win the Ultimate Play NVIDIA RTX 3080 Ti GPU.
If you missed the Day 1 sessions, we will recap some of the highlights and share information on the sessions to look forward to on Day 2. We will also share links of some of the recorded talks, so you can watch them later.
Developer Training Workshops
Day 1 was hugely focused on developers and developer education. The day kicked off with several full-day workshops on various topics such as
- Applications of AI for anomaly detection [DLIW1573]: This workshop was about identifying intrusions in telecommunications networks using machine learning-based anomaly detection. Although the specific example was from telecommunications, the same ideas can be applied to detecting fraudulent transactions, security etc. The participants were taught to implement three anomaly detection techniques:
- CUDA accelerated gradient boosting with XG Boost (no deep learning)
- Auto-encoders (based on deep learning), and
- Generative Adversarial Networks (GANs)
- Fundamentals of Accelerated Data Science [DLIW1574]: If you are interested in ‘traditional’ data science (not involving deep learning), then this workshop is for you. This workshop included hands-on practice on accelerating data science pipelines from loading data all the way to analysis. The instructors introduced CUDA-based libraries like cuGRAPH, cuML, and cuDF, and best of all, you do not need deep expertise in CUDA to take advantage of any of these libraries.
- Fundamentals of Deep Learning [DLIW2325]: This workshop provided a high-level overview of the fundamentals of deep learning and how they apply to computer vision and natural language processing.
- CUDA Python [DLIW2458]: Do you ever write code in NumPy, only to realize it is slow? Do you wish your NumPy code could be accelerated by GPUs? If yes, then this workshop is for you. It introduced Numba, a ‘compiler’ for python code. Numba takes existing Python or NumPy code and compiles it to run on GPUs.
- AI-based Cybersecurity [DLIW2515]: NVIDIA has created a dedicated framework for cybersecurity called ‘Morpheus’, introduced in the video below. This workshop was an introduction to Morpheus showing how cybersecurity experts can use GPU acceleration to guard massive systems against zero-day attacks.
- Learning Deep Learning [DLIT2675]: We know that NVIDIA is a leader in AI software, hardware, and products. So, wouldn’t it be great if someone from NVIDIA wrote a book about Deep Learning, explaining DL from the point of view of someone with a front-row seat to the AI revolution? That’s exactly what ‘Learning Deep Learning’ is all about. NVIDIA’s Magnus Ekman has written this massive 752-page explainer on various aspects of deep learning, from computer vision to natural language processing to transformers. The best thing about this book is that all code samples are provided in both TensorFlow and PyTorch. This session was an introduction to the various topics covered in this book. Luckily, the code samples from the book are available publicly on GitHub, so you should definitely take a look at them.
Day 1 also contained many workshops related to natural language processing and time series forecasting, which we mention here briefly:
- Building Conversational AI Applications [DLIW2460]: This workshop was an introduction to NVIDIA’s RIVA framework for building speech AI applications for use cases such as automatic speech recognition, named entity recognition etc. The speakers highlighted the fact that RIVA integrates seamlessly with NVIDIA TAO transfer learning framework which makes training and deploying models easier (we will share more about TAO tomorrow).
- Transformer-based Natural Language Processing [DLIW2330]: As the name suggests, this workshop walked through advances in NLP over the last few years culminating in transformer models like BERT and Megatron.
- AI for predictive maintenance [DLIW2331]: Using LSTMs for predicting machine failure and reducing downtime.
Although registration to GTC is free, these are paid workshops for a fee of $149 each. If you are interested in workshops, please have a look at the agenda for the next 2 days, since multiple copies of every workshop are available for many time zones.
Developer Breakout: Advanced Physics Simulation [SE2312]
Like breakout rooms in Zoom, Developer Breakouts were somewhat smaller and more intimate sessions, giving developers many more opportunities to interact with the speakers and ask questions. Luckily all developer breakouts were recorded and can be viewed now. This session was conducted by Michelle Lu, Andrew Reidmeyer, Kier Storey, Michal Hapala and Ales Borovicka. All of them are working on various aspects of physics simulation such as rigid body dynamics, soft body simulation, fluid flow, and ray tracing enabled rendering.
Michelle kicked off this session with a very cool demo video showing multiple rigid body simulation. The video showed multiple boxes in a warehouse colliding with each other and moving under the influence of a ‘force field’, a new feature in the latest version of PhysX, NVIDIA’s physics simulation engine. This was followed by another cool demo showing how even complicated physical interactions like screwing a bolt into a nut can be simulated. In this demo, a robot in Omniverse, picked up a bolt and rotated it to screw it into place on a nut.
For people in automotives, an amazing demo of gear interactions was up next, showing a simulation of a dune buggy driving around. The video clearly showed how interactions between gears in the buggy were driving the tyres into motion and causing the buggy to move. Later part of the video showed that even the engine of the buggy (a 4 cylinder one) was accurately simulated with the lateral forces from the pistons causing visible vibrations in the chassis.
This amazing demo prompted someone in the audience to ask whether this could some day replace professional finite element modelling software like ANSYS. While NVIDIA remained mostly tight-lipped about this for now, they did state that they are doing some validation against ANSYS for certain use-cases.
The second half of the session included many demos of soft-body simulations such as jelly cubes, clothes, and even a simulation of cloth in a windy environment. Michelle stressed that all these demos will be available for free with the latest release of Omniverse. Most importantly, these features are available on even consumer-grade GPUs like RTX 3080 and 3090 and do not require Quadro or A6000 (though the extra CUDA cores of the latter will definitely help). This was a very visual session, the full contents of which cannot be communicated in a text form, so we strongly recommend you to watch the recording here.
This presentation makes it clear that NVIDIA is continuing to invest heavily in all aspects of physics simulation and rendering. In a few short years, we might be able to simulate unstructured environments like agriculture, construction, and traffic. With faster simulation, we might see reinforcement learning finally take off in these environments.
Developer Breakout: 3D Deep Learning with Kaolin [SE42115]
If you do not know Kaolin, do not worry you are not alone. Several attendees of this session too had no idea what it is. Kaolin is an open-source library by NVIDIA to simplify 3D Deep Learning. For example, 3D meshes, point clouds, and volumetric representations can be visualized with Kaolin. The lead developers of Kaolin, Jean-Francois Lafleche (JF) and Clement Fuji Tsang (CF) led this session and introduced how Kaolin fits into the 3D deep learning workflow.
The feature that caught our attention is Structured Point Cloud (SPCs). We all know how precious GPU memory is. Since point clouds are dense, they can take up a lot of GPU memory. SPCs are a way to exploit the structure in the data to reduce memory footprint and thus render huge point clouds, which would otherwise be too big to visualize.
The developers avoided a direct comparison of Kaolin with PyTorch 3D, stating that Kaolin is mostly community driven and even the choice of supporting PyTorch (over TensorFlow) is largely community driven.
They shared an excellent list of resources for getting started with Kaolin:
- Kaolin documentation
- Kaolin v0.10.0 changelogs
- DMTet paper
- DMtet tutorial
- Understanding SPCs tutorial
- Omniverse Kaolin App documentation
JF and CF also mentioned that Kaolin examples are a good reference for getting started with 3D deep learning. Let us know in the comments if you would like us to create blog posts about 3D deep learning.
Why you should attend tomorrow
This brings us to the end of Day 1. The first day was just about getting our feet wet, since the real fun starts tomorrow. There are nearly 100 sessions planned for tomorrow with topics such as:
- Keynote with the announcement of Jetson Orin
- Developing Edge AI applications with DeepStream
- Demystifying Unified memory on Jetson
- 3D deep learning with Omniverse
- AI models made simple with NVIDIA TAO
- Getting started with edge AI on Jetson
- Developing, training and testing AI robots in simulation
- Introduction to autonomous vehicles
As we mentioned before, if you register with our link, you could win a RTX 3080 Ti GPU, so please register here. GTC registration link