Computex: Nvidia: No more products for people

After the Computex keynote, you could actually think that Nvidia would have lost contact with reality. c’t 3003 listens more closely.

Transcription of the video

(Note: This transcript is intended for people who cannot watch or want the video above. The text does not return all information of the image track.)

Take a look here, as honest as Nvidia here at the Computex I have never heard a company talking about AI (translations from us): “Risk performance is now sales. Arithmetic performance is profit. The lack of revenue and profit is loss.” People? It's about agents. “Fourther we have built CPUs for people. This CPU is built for agents.” Content? It's about tokens and profit. “Tokens are now profitable units.” Profit? “Every token is profitable.” Profit. “Real performance is turnover. Performance per watt is your turnover.”

So, summarized: The tokens that get out of the back thanks to a lot of computes, which is not about making something meaningful with it, but it's about generating profit. “That’s exactly why you’re all so busy and your companies run so well. In fact, this looks like some of your share prices.”

So I don't know how to do this, but I don't have anything to do with my reality. What I see here on the Computex instead: The classic PC market is quite on the ground, where Nvidia is not quite innocent, because all of the complete “KI makes you all rich” psychosis are decayed and a AI data center is piled out of the ground after the other.

And because the component manufacturers can make more money, the memory goes mainly to the AI data centers. And that then leads to a four-year-old PC console not becoming cheaper, as usual in technology history, but more expensive. And much more expensive, equal to several hundred euros. And AMD even releases processors for 10 years old AM4 platform, so that you can continue to use the old DDR4 memory instead of buying too expensive new ones.

So, and now comes the most scorching. So you can say that Nvidia has really put the personal computer market into the tumbling. And then Nvidia comes around the corner and says, "Hey, the old PC, it's dead! Now we have the new PC.” “This will be the new PC. Over the last three years, it took so long to reinvent the PC completely.” Personal AI PC. You can even generate your tokens yourself, thanks to RTX Spark. And there is Microsoft in the boat and there are really all major PC manufacturers involved.

Okay, so I think there's a lot of journalistic about me now. Hold on.

Dear hackers, dear internet surfers, welcome here at ...

Nvidias destructive presentation at the Computex

Okay, what happened? At the Computex in Taiwan, the world's largest computer fair, where I am on my way for the first time, Nvidia held the opening presentation. Two hours. You may also have to say that Nvidia founder Jensen Huang is something like a folk hero here in Taiwan. At every snack where he ate something, there seem to be pictures hanging from him. So at least that seemed to me that way. Yes, and there were also, as part of the Computex, so Chairman Huang says very proudly, 70 keynote eye-catching parties where his presentation was just broadcast. At such a gucking party I also saw the keynote because invited to the lecture directly Nvidia did not have me. Funny.

Yes, and Nvidia has actually announced CPUs, so no GPUs, for notebooks, along with Microsoft. The details later. But first of all, what Jensen Huang said on the keynote. And I honestly found that dystopian. I found it jamming.

All right, that thing started with, how likely every AI company was: The world becomes better with AI and all diseases are cured and so on and so on. You know that. But as soon as it became concrete, then it was no longer about humanity. Jensen Huang even said: We used to build CPUs for people. Now it's about building CPUs for AI agencies and maximizing tokens.

Tokens, these are the basic units with which current AI systems, i.e. language models, work. And Nvidia and its direct clientele, clearly, who want as many people as possible to do with AI, so consume as many tokens as possible. In reality, however, this has already led employees, for example, at Amazon, to have simply made completely pointless, expensive AI requests, because they were afraid that they might be bad in the token rankings, because any, yes, I just say, maybe not quite so smart supervisors have just believed what Jensen Huang has said here the other day, namely that it is deeply alarming when developers are at least causing an alarm.

So that Jensen Huang and his clientele naturally want to consume a quarter of a million in token costs, that's clear. But that really is sold so that you can measure the value of employees in their token consumption, yes, that is, I must say, stupid.

Yes, and Jensen also showed at the end of the two-hour keynote what you can do with a lot of tokens: a very bad AI-generated song with a generic cartoon plus typical KI-Slop errors. Hey, are you looking for KAROOKE? Or robots glued together so funny? Yes, RESENANG ACCELABED COMFUTING.

Yes, and here in Taipeh, Jensen Huang has further driven this token extremeism. Very simple: “Tokens are sales.” Okay, so that was his story. computing power, i.e. in the case of computing power in AI servers, which calculates tokens, and tokens are sales, i.e. computing power is also direct sales. He didn't even say that, but he always repeated it.

“Tokens are now profitable units. Tokens are now profitable units of sales. Chip, rack, network, electricity, cooling and power grid must be assembled at the end because computing power is revenue. DSX Max LPS allows operators to install more GPUs with the same power consumption, generating billions of revenue. More current flows into converter generating computing power. computing power is now sales. computing power is profit. Power per watt is revenue. Because every token is profitable. Every token is turnover. The more you buy, the more you earn.”

And that's wild because it's really very misleading. It's about what to do with the tokens. Well, the ones that sell the tokens and create the servers, for which it doesn't matter. But Jensen Huang pretends to benefit. So also those who only buy tokens. And jobs would not be lost either. This is total nonsens, because the employees triple their productivity – respect. Yeah, he really said that.

“The sum of 3 trillion dollars now generates almost three times as much output. This effectively corresponds to a productivity of 9 trillion dollars from 3 trillion dollars."

And as a result, the profit and the employers will naturally also introduce new people when the profit increases.

“The number of software engineers is increasing. People talk about AI reducing the number of jobs. Complete nonsens. It ensures that more software engineers are set. And the reason is quite simple: if you can set a software engineer, and thus generate 9 trillion dollars in productive work: Why would you not just stop?”

But this bill, it's something you can take apart. Just how Nvidia measures productivity. They just looked at the commit volume on GitHub. So how many code changes are introduced into software projects. And yes, the Commits have been extremely high in recent times, by very likely AI agents. That's right. But I will let you discuss in the comments whether the number of commits on GitHub really is one to one with productivity.

And anyway, that only software developers find themselves in the bill, so people who can undoubtedly start most with AI systems today. However, there are several professions where there is certainly no productivity gain by AI.

Anyway, the Nvidia keynote seemed to be like billionaires for billionaires. I'm obviously not the target group. Although I find AI technically quite interesting and do a lot of work on it, I definitely do not triple my productivity. And that tokens should be profitable value units for me? Yeah, I don't know now. So that seemed to me to be very decoupled from reality, or at least from my reality.

I don't want to be around all the time. I actually found an example from the keynote really very interesting. Here's how I lost the flap from my remote control, which looks like that, let's do a CAD file. And then, with the help of the CUDA-X library collection, the AI agency builds a file that eats a 3D printer, and then you can simply print it out. So there's the idea behind it that you provide special libraries and frameworks that can be explicitly served by AI agencies. This could actually be useful, even if I don't really know if this must be absolutely CUDA libraries that only run on Nvidia hardware.

The new PC platform RTX Spark

So, but now to the most interesting, at least half concrete announcement of Nvidia. From autumn, Windows notebooks will come out, which run completely with Nvidia architecture. So not only the graphics unit, but also the CPU. The whole thing hasn't happened so far, so at least not for Windows.

The whole is called RTX Spark, not DGX Spark, as these small AI boxes we have already tested, but RTX Spark. The whole thing is ARM-based, so what Qualcomm already does with Windows and Apple stops with macOS. Just the competition architecture of x64, so what Intel and AMD do.

Yes, and with 20 ARM CPU cores, i.e. with Cortex X925 and A725, and 6144 Shader cores, i.e. the Blackwell architecture, the RTX Park processor corresponds to an end-user version of the system on chip GB10, i.e. the chip from the DGX Spark. Only the DGX Spark pulls up to 216 watts. And that is, of course, difficult with such a notebook. And that's why Nvidia will certainly have much optimized the RTX Spark. But there has not been any concrete. Only that it is the most efficient chip ever from Nvidia and that the battery lasts a working day. At least, Microsoft says on a website about its RTX Spark variant. This thing is called Surface Laptop Ultra.

Yes, and Microsoft are not the only ones to build RTX parking devices. In fact, Nvidia has all, really all relevant players in the boat, so Acer, Asus, Dell, Gigabyte, HP, Lenovo and MSI.

Yes, and so, as I said, one should be able to generate his tokens and execute the AI agencies locally. So what I did in my OpenCore video with an AMD Strix-Halo system. Yeah, and you should be able to play. But unfortunately there are no benchmarks and specs yet.

So the RTX-Spark devices are to give it with up to 128 GB of unified LPDDR5X memory, which creates a throughput of 300 GB per second. CPU and GPU are connected with 600 GB per second via NVLink. And yes, it was.

Nvidia says the first devices are coming in autumn. Here at the Computex you can hear a bit behind held-up hand that this could be athletic because it is still crumbling with Microsoft in Windows-11 support. Yeah, look.

In any case, I find it really interesting that the AI mania triggered by Nvidia in the data centers of the classic PC market is getting really smoky, so because of the storage prices, and then Nvidia says: You also need really doll AI performance on your laptops. And now I don't know if the target group does that. Because Microsoft's copilot Plus PCs are there, I say, the euphoria is not the same.

Yeah, that's what it's like to know if you can do anything with these Nvidia calculators, what you can't do with an equally expensive x64 calculator or a MacBook.

Yes, keyword is equally expensive: So how expensive an RTX parking part with 128 gigabytes should be, that is unclear to date. So I see here on the Computex mainly that the manufacturers introduce all 8-GB notebooks, which is actually quite useless because they just have to keep up with the MacBook Neo, because the memory is so expensive because of Nvidia.

So everything's difficult. I'm curious how this goes on. What do you think? Bye from Taiwan!

c't 3003 is the YouTube channel of c't. The videos on c’t 3003 are independent content and independent of the articles in c’t magazine. The editors Jan-Keno Janssen, Lukas Rumpler, Sahin Erengil and Pascal Schewe publish a video every week.

(jkj)