Better late than never. Back in March NVIDIA held the biggest GTC ever! Unfortunately I had a death in the family which cut my visit short. I left before my scheduled talk on Thursday. I apologize for my absence , but I’m sure you understand. For people I did get to meet at the show, thank you for introducing yourself!
Big Time
Since a lot of this is “old news” by now, I’ll point out some of my observations from the show. First, the conference had a much different feel than the previous (2019) one. Just the sheer number of people (estimated to be ~ 16.5K) put a strain on the resources of the San Jose Convention Center. It felt like at least twice the number of people as the last one. The show has outgrown the venue. This is for several reasons, the most obvious being that NVIDIA is now valued at over a trillion dollars. That means it has reached critical mass where all financial and technology folks have to pay attention.
Another big point. It’s the “show” part. While Jensen Huang has given more keynotes than you can shake a stick at, NVIDIA cranked it up a notch. NVIDIA rented out a sports arena for the keynote, with rock concert worthy stagecraft. Looky here:
Quite the show.
AI
There was several mentions of AI at the show. In each sentence. A little scary what happens when the tech marketing groups get rolling. There’s no such thing as subtle during these types of event.
Exhibit Floor
In the time that I was there, the exhibit floor was always jam packed. There were many amazing sites to see. There were two ‘theaters’ were groups would give lightning talks. Some were vendors, some were researchers. The talks would be ~ 10 minutes, just enough to get a feel for a broad range of subjects.
The exhibit floor was divided into different domain specific areas. The edge computing area is where most of the Jetson crowd was located. There was also a robotics area, where a lot of interesting work was being presented. I had a nerdgasm when I saw the Disney Research BD-X droids there:
I had done extensive research and guessing how they work. Disney Imagineering held a session at GTC which told us exactly how they do! One interesting fact is that the gait of the droid uses a reinforcement learning model developed on ISAAC. It takes a RTX4090 ~48 hours to create the gait model. There are two Jetson Orin NX in the droid. One in the head and one in the body. This gives it enough compute to do some very interesting things! After all, they are research projects. Since GTC, these droids have been seen at Disneyland, working for a living. People, especially kids, love these things.
Unitree had some doggies there!
Sessions
There were hundred of sessions, so I can’t even begin to tell you what was covered. The great thing about sessions is that you can find a topic that you are tangentially interest in, jump in a session about it, and have a great overview of where the state of the art is in that subject.
I’ll also mention that almost anything that involves computation (virtually all science, technology and business fields now) are represented. As an example, I went to a session about the Sphere in Las Vegas, Nevada. They were talking numbers that it’s very hard to get your head around. For example, there are ~ 166,000 speakers in the venue. The speakers beam form audio to any given seat. Which means every seat in the house has an optimal audio experience. Just getting the data to the correct place on the network is a herculean task. The 50 minute show now playing is 14 petabytes (PETABYTES!) of uncompressed video and audio. They use NVIDIA switches to help move the data around. Definitely big league stuff.
Very many of the sessions were recorded, and available on demand: https://www.nvidia.com/gtc/ Also, some of the more popular sessions are available on YouTube ( 25 of them as of this writing):
Announcements
There were many announcements at the show. Here we’ll focus on the some of the Jetson related ones.
Blackwell
Blackwell is the name of the new architecture for NVIDIA GPUs. Like previous generations, NVIDIA will be offering these in many configurations. We’re still a little way off from all the technical details, but the data center GPUs will be two dies connected via a high speed transport fabric in the same package, similar to the architecture of Apple M3 Pro chips.
As you recall, NVIDIA cancelled the Atlan generation of Jetson. This year they announced the Thor generation. This will be Blackwell architecture, with reportedly 6-10X the amount of performance as the Orin line. This will open up a lot of new application spaces.
Robotics
The first Jetson Thor chips will be for automotive. They also announced Project GROOT, a general-purpose foundation model for humanoid robots. GROOT will run on Thor, with what appears to be all of the humanoid vendors (with the exception of TESLA) being on board. All this is tied together with ISAAC, of course.
There is a big emphasis being put on digital twins. The idea behind digital twins is that you build environments and robots in simulation which exactly replicate their real world counterparts. You train the robots to work in that environment. Then you deploy the trained models onto the real world robots, and everything goes to plan.
There were many real-world cases of working examples. In the end, this should save a lot of time and money.
Conclusion
GTC 2024 was quite the experience. If you want inspiration, consider attending one of these events. Everyone is there. There are lab sessions where you can have any questions you have answered by the engineers who work on it day in and day out. Well worth the investment. If there is one tip I have, you should go over the sessions carefully and find ones that you find interesting. Planning is key here. There’s way to many bright and shinies to distract. But those can be fun too!
The JetsonHacks schedule should be coming back to more normal in a few weeks. Lots of projects to catch up on!
The post NVIDIA GTC 2024 appeared first on JetsonHacks.
