Nvidia Computex 2023 keynote highlights

What does Nvidia have to show off when it comes to the big league

Table of Contents Table of Contents

Computex has finally started once again, we see what Nvidia brings to its 2023 keynote. Returning to public speeches, since the last time one was given four years ago. And with the boom of recent AI, Nvida’s shares and market capitalization have been booming, as it reached nearly the one trillion dollar market value.

But we don’t expect much for gamers, but rather the more lucrative area of data centers is the highlight of this talk. Especially as they already jumped the gun the week before with the release of the RTX 4060 Ti. With a two-hour long talk packed already, we see why it might have been cut to the previous week.

AI focus

As we might expect, AI is a focus of this event. One of the biggest contributors to Nvidia’s ever-growing range of products, and the next focus after it sold plenty of graphics cards to crypto miners. Now it needs a replacement and that comes with powering AI.

We see how the introduction to the keynote is produced by AI, helping give some of the script, it is also then voiced by AI. Along with some of the imaging provided. Giving the big push for what the next best graphics cards will be used for.

Nvidia ACE

Jensen has announced AI animation for games. Nvidia Ace gives a quicker creation of characters and stories without a script and design. You can generate the whole of these interactions more easily without writing the whole thing yourself.

With a cutscene showing off the use of scripting with just background information on what the character intends. This would be a much quicker process of creation as the avatar can also be generated with AI.

Accelerated computing and deep learning

Jensen points out how the constantly increasing power and possibilities of computing is slowing down over the years. So the focus should be on power efficiency and the two factors that came at the right time accelerated computing and deep learning.

The computing model comes to data centers, as that’s where the computers are. Jensen shows off that for USD 10M you get 960 CPU servers, running at 11GWh, for 1xLLM (large language model). But the GPU servers, are much more, with the same amount of USD 10M, you only get 48 GPU servers, 3.2GWh, and 44x LLM.

Since most data centers are power-limited, Jensen shows how much more efficient GPU servers are. “The more buy, the more you save”, as you only need two GPU servers to have one large language model, not the 960 servers required.

Nvidia also flexes being able to set up the use of data centers a lot faster. This means companies start earning a lot faster from it, along with the lower depreciation of the hardware. With the utilization of the graphics cards a lot higher and it gets to much better use out of it.

Hopper HGX H100

Showing off their data center hardware, the H100 with 8 GPUs on a board weighs in at 60 lbs and will set you back $300,000. And the heavy setup is what is running those large language models. Showcasing Nvidia’s reinvention of what computers are, with a design and invention of the whole model available, and what is the true product it cares most about.

Generative AI

Another transformation of what is capable of AI is the ability to create a wide range of fields. Another look at the generative AI Jensen shows off the power of what AI can do. Providing text samples and creating an avatar and synth speech from it, furthering its capabilities. Along with Google’s text-to-music to create s, Voicemod creates a song.

Nvidia is currently working with 1,600 Gen AI companies to work in different sectors to create new uses for AI to help with what it can do. With biology

Grace Hopper in production

200 billion transistors with a lot of computing power but at a low power draw. But also both CPU and GPU with access to a lot of memory at a high bandwidth. As others are limited by the memory available, the GH200 is a much more efficient usage of data without having to split it like the HGX H100 computer it had previously shown off.

This does show why Nvidia would want to take over ARM, as with just 8 of these with an NVLINK Switch, you get a whole pod of them with 960GB/s bandwidth. So with 256 Superchips, creates one ExaFLOPS of power and 144TB of GPU memory that is shared. So with 150 miles of optical fiber, 2,112 60mm fans, 70K CFM, and 40K lbs you get the biggest GPU. 230 TB/sec, 1 EXFLOP of computing is a much more efficient computing installation for big companies.

Maxine 3D is another utilization of the Grace Hopper supercomputer. It combines the previous announcements of eye contact mode and auto-translation which will enhance conferences. So communication is upgraded including 3D mapping just from a 2D camera in any hardware, so it’s more responsive and inclusive.

MGX

Nvidia announces the open modular server design for Accelerated computing. The chassis is multi-generation standardized for TTM and improve return on investment. Configurable to plenty of solutions, it’s another part of the market it can attack to fill with its hardware.

Spectrum-X Switch

48 PCBs, connect up the Infiniband capabilities in a singular switch it aims to lower the jittering and packet loss with error-correcting. It has 100 billion transistors with TSMC’s 4N process. Providing 51.2T of bandwidth, with 64 x 800G ports or 128 x 400G ports. It is end-to-end optimized for AI workloads. Along with managing the data more effectively between parts.

Omniverse

Omniverse is also expanding, the tool allowing collaboration and creation, it is also working on AI integration. Giving prompts can help change the background of the models you have shown or similar things. Working towards just giving it ideas rather than doing it yourself.

VOD