Jiwei Network reported (text / Sun Qinzhou), the biannual GTC is one of the world's largest AI technology conferences, at the 2023 Spring GTC Conference held last week, NVIDIA released a series of hot topics involving large-scale language AI, generative AI, industrial metaverse and other hot topics, but behind these hot topics, NVIDIA's ambitions seem to be more than that.
Where is the consumer metaverse?
From the "first year of VR" to the "meta-universe", compared with the hot consumer market of the meta-universe concept in previous years, this year's meta-universe market is obviously much cooler.
According to Counterpoint, global XR headset shipments fell 21% year-over-year in the fourth quarter of 2022, along with declines across device types, a phenomenon that Counterpoint attributed to a lack of high-profile products and compelling use cases in recent quarters.
Meta, which dominates XR devices, is naturally one of the biggest victims of the weakening XR market environment, according to its official financial report, the Meta Reality Labs division generated revenue of $2.159 billion in 2022, down at least $100 million from 2021.
Compared with the consumer-grade metaverse, the industrial metaverse seems more "tempting" for NVIDIA. Ahmed Banafa, a lecturer and academic advisor at San Jose State University, once spoke to Jiwei Interview about the difference between the two: "In the Industrial Internet of Things (in this case, the metaverse), you have a limited number of customers, but they tend to have a lot of money. And consumer IoT will have many headaches, such as laws and regulations and privacy. ”
At the meeting, NVIDIA released the NVIDIAOmniverse™Cloud cloud service platform and announced a partnership with Microsoft to host the service in Microsoft Azure. Customers can build a digital twin of a factory in an industrial metaverse or test the performance of a product's digital twin in the metaverse. This accelerates product development, optimizes manufacturing processes, and more.
In addition, customer companies can link real-time data from sensors in the physical world to digital twins on the cloud, allowing customers to build more accurate digital twins in the metaverse.
(Image source: NVIDIA)
However, compared with the industrial metaverse, NVIDIA's AI layout for many years seems to be more valued.
From ChatGPT to computational lithography, the "iPhone moment" of the AI industry
At least, judging from Lao Huang's keynote speech, NVIDIA is well prepared for this. In the keynote speech, the company launched four inference platforms for generative AI: NVIDIA L4 for AI video, NVIDIA L40 for image generation, NVIDIA GraceHopper for graphics recommendation, vector databases, and graph neural networks, and NVIDIA H100NVL for the current popular language model AI ChatGPT. It also said that the DGXAI supercomputer launched for AI computing power suppliers has been fully put into production.
Among them, the NVIDIA H100NVL, designed for large-scale language model deployment, has 12 times the inference performance on GPT-3 models compared to the previous generation A100 - and the HGXA100 is the only GPU with real-time ChatGPT capabilities.
(Image source: NVIDIA)
At the same time, in addition to selling hardware directly, NVIDIA also launched NVIDIA DGXCloud, which allows companies to rent the computing infrastructure and software needed for AI or other training models on a monthly subscription basis.
Of course, the biggest "king bomb" of the entire conference is that NVIDIA pulled ASML, TSMC and Synopsys to introduce accelerated GPU computing into the field of lithography. By running the NVIDIA cuLitho software library for computational lithography on the GPU, it offers up to 40x the performance leap over current lithography technology, which means that only 500 NVIDIA DGXH100 systems can do the work of a system that previously required 40,000 CPUs.
(Image source: NVIDIA)
In order to prove NVIDIA's determination in AI, Huang Jenxun even took the iPhone in the smartphone industry as an example and repeatedly emphasized in his speech that we are now in the "iPhone moment" of AI.
Perhaps computing power infrastructure is NVIDIA's goal
But we believe that for NVIDIA, it doesn't matter whether it is AI or the metaverse that is in the "iPhone moment", their real goal is actually the data center that provides computing power as infrastructure behind them.
Because whether it's moving to the "metaverse" of industrial productivity, generative and conversational AI on fire recently, or even computing lithography, which is called a "technological bomb" in the speech, they all have one thing in common: they all require more or different computing power than ever before, so these service providers or potential customers for these use cases will have to upgrade their data centers if they don't want to be left behind in the coming waves.
For example, according to semianalysis, if Google deployed ChatGPT directly into each search, it would need about 512820.51 A100HGX servers, for a total of 4102568 A100 GPUs. The cost of these servers and networks alone can exceed $100 billion.
You know, the total cost of Google completing a traditional search request is only about 1.06 cents, and the cost of introducing ChatGPT per conversation needs to increase by 1.42 cents, of which hardware expenses account for 0.36 cents.
(Image source: SemiAnalysis)
Although the actual situation is not as exaggerated as estimated, after all, Google can save computing power needs by reducing the number of times AI is used, but there is no doubt that for NVIDIA, which provides graphics cards, just as the progress of 3D graphics games promoted the development of GPUs, the construction of data center infrastructure driven by AI and industrial metaverse is a huge opportunity.
What are the opportunities? According to NVIDIA's official revenue forecast, from the second quarter of fiscal 2023, the revenue of the data center market will replace the gaming market (that is, the market where gaming PCs and the like are located) as NVIDIA's main source of revenue, and in the third quarter, data center revenue will jump to twice the revenue of the game market. In fact, it's not just them, its old rival AMD's recent financial report also showed that consumer electronics revenue was about $1.6 billion, while data center revenue reached $1.7 billion. Lucas Kah, an analyst at Golin Management Consulting, told me in an interview with iJiwei: "We are seeing an inflection point in data center revenue for the first time, eclipsing gaming and PC revenue."
(Image source: NVIDIA)
Although we cannot determine whether AI is reaching the iPhone moment or another "gimmick", what we can be sure of is that neither AI nor the metaverse is the end of NVIDIA, but a springboard for it to achieve data center supremacy.