Meta ads work because Meta is a massively popular platform, and because it targets those ads relentlessly.

That’s… not possible with floss.

In other words, I’m struggling to understand what you are envisioning here.

…If you want Meta ads, why wouldn’t you just use Meta ads?

That part I don’t know. TBH last time I used CAD software, they used graphics APIs these chips wouldn’t support, but honestly I don’t know about the rendering part these days.

Not CAD. Pretty much anything that’s “real time graphics” needs ROPs.

But Blender rendering, yes. Works great on datacenter GPUs.

It goes well beyond that. There’s tons of stuff GPU compute is useful for; personally, I want a B200 for video/image processing and for more “old school” machine learning experimentation.

It’s not a software limitation. Datacenter GPUs physically do not have ROPs; they cannot render games.

Nvidia removed them on the H100. The A100 technically has some, but not enough to perform well.

AMD removed them even earlier, on the MI100 at the very least.

…But if you are interested in workstation compute stuff?

Datacenter GPUs are great at that. I look forward to picking up a firesale B100 someday.

Also, there is already a group hacking them on the Level 1 Tech forums. But they are trying to unlock the motherboards to use them for compute, not graphics.

An issue with LLM inference is that the software architecture changes significantly, very quickly, and the combination of Nvidia’s software support + some architectural flexibility makes GPUs quite advantageous for that.

Just as an example, everyone saw KV cache as this existential limiting factor for LLMs. Now, it’s tiny. Another is the old dense model paradigm vs. the almost comically sparse ones coming out. Even matrix-multiply compute could be rendered irrelevant if Bitnet gains traction.

And that’s not to speak of training.

Nvidia has done supply shenanigans like this in the past, too. Repeatedly. Oldschool hardware outlets covered it extensively.

The Datacenter GPUs can’t be used for graphics, unfortunately.

Now, if you want a compute firesale…

I thought this would be nonsense at first, but the methodology is interesting. It seems plausible. And they tested quite a good spread of open weights, like MiMo and LongCat (which usually fly under the radar), all the way back to Yi 34B.

They missed Russia’s LLM, but that’s okay; it’s very obscure.

It’s also “lucky” they found this correlation, and so many recent weights happened to be initialized with PyTorch, on the GPU: very specific conditions.

…Also, I think the FP4 precision is a dead giveaway too, as Hopper wouldn’t be able to train that effectively, right? I assumed they were using at least some smuggled B200s.

And the export restrictions were stupid anyway.

He’s always pushed side stuff, like robots and AR and artificial datasets, even in the LLM boom. But he literally designed their first GPU; I think he was there too long to ever lose graphics as a priority in his mind.

Graphics isn’t just gaming. It never was. It’s forever been a mix of workstation stuff too.

The fear is that a new CEO will just look at Nvidia as an AI ASIC maker, specifically.

It’s held true for the entire history of human civilization.

Large Chinese firms have a bunch (though not a ton) of Nvidia GPUs, so the “end state” is proof too.

To clarify some, I meant these older sins have fallen out of consumers’ attention because of the attention on generated assets. It shouldn’t. That context is important.

IMO it would be inconsistent to loudly boycott some indie dev over (say) some diffusion generated assets, then turn around and buy an Activision or EA game in the same breath. That kind of purist absolutism isn’t productive; the larger context matters.

You mean over the export restrictions? Those were always meant to protect Nvidia, but were both ineffective and ended up doing the opposite, by encouraging the development of Chinese chips (as Jensen tried to warn them of).

So… honestly, that isn’t a big deal.

What’s worse is the potential conflict of interest sleeping with an Nvidia executive.

I was just trolling a little by copying someone else’s thread, no bad blood here.

That makes sense. There are definitely some PS5 games that don’t make use of all its features/hardware.

…U mad bro?

This here is the correct answer.

That’s a good definition, in fact I am bookmarking that to remember later.