That’s a lot of vagueness though.

I’m interested in something physical. Something I could buy, that could realistically upgrade an Arc B580 in (say) indie games, or KCDII, or llama.cpp.

This is nowhere close yet.

And I dont doubt they’re getting investment, but I’m just skeptical because I’ve seen this story a hundred times before, even in China. And even then, what’s promised to be general hardware often evolves into something for a very nich need. As a recent example, see Tenstorrent.

What’s stopping you?

My feelings:

• Neat.

• Future is interesting.

• But proof is not in the puddin’ yet.

It’s easy to say “they’ll scale it up. They’ll get to optimizing the software,” but I’ve seen waaay too many hardware makers mess up that step, and fade into the vast graveyard of their peers.

Like, does anyone remember Centaur’s exciting, 8 core x86 CPUs, with a huge block for ML inference, from about a decade ago?

https://chipsandcheese.com/p/via-part-4-a-deep-dive-into-centaurs-last-cpu-core-cns

It was competitive with Zen 1, not even counting the accelerator.

The answer is “probably not.” Even the Wikichip Fuse writeups on the architecture are gone.

Doesn’t matter(for this, specifically) if it’s not performant on LLM inference engines.

And I’m not just talking about CUDA. Even GGUF Vulkan (for example) has all sorts of vendor quirks that can absolutely trash performance. VLLM is often a joke on AMD, with certain models, on certain cards, even with dev support.

Of what, though? Huawei NPUs are datacenter hardware.

As much as we hate it, Nvidia gaming GPUs are ultimately cheap consumer devices, and they’re very good at hybrid CPU+GPU inference.

I think Intel has the best chance of pulling a rabbit out of a hat with Arc. They have a usable platform already, hardware “close enough” to Nvidia that LLM compatibility isn’t a nightmare. And they have nothing to lose, no illusion of “protecting datacenter cards” like AMD has.

They’ve released two open weights LLMs, trained on AMD hardware.

…And yes. They are archaic jokes. I could have trained a better model if I was in charge of it, which is sad.

And don’t even get me started on hardware and library footgunning.

Yeah. I mean, I have a Ryzen desktop and a 2020 GPU, and Mimo 2.5 is a bit faster and mind bogglingly better than frontier models from like… two years ago? And frontier models are plateauing, I think.

Still, my worry is that we consumer won’t HAVE any hardware. Many don’t even own a laptop these days, and it feels like they’ll just drop desktops (and work will just use thin clients) if they’re too cost prohibitive for people to buy.

Oh don’t mistake me, they are not consumer friendly.

They are just trying to sell enterprise GPUs directly to “consumer” businesses and the cloud providers they use, instead of through literally fraudulent middlemen like OpenAI.

This is what pretty much everyone with hardware is doing, including Huawei, Tenstorrent, Cerebras, even AMD. Maybe I misinterpreted you, but hardly anyone cares about individual self-hosters.

Apple does, though. MLX is actually getting pretty cool. But they’ll always be quite insular, anti-consumer in other ways, and they still seem detached from what the community is largely doing.

Nvidia sees the writing in the wall too, hence the big Nemotron effort now. They’ve been pushing open models, but no one can hear them over Altman’s lies.

AMD… is… trying.

Some other companies have made pretty interesting efforts too, like LG and IBM. Huawei already publish a big model to promote their ASICs, and is planning another in weeks. Even some Russian company trained a big open LLM from scratch, though it wasn’t very good TBH.

And this is not even looking outside the LLM space, where all sorts of interesting models are published.

I’m on a RTX 3090.

At the moment, in Windows 11, Chromium hardware acceleration is busted for me. YouTube is stuttering like crazy. IDK why; my Nvidia drivers are up-to-date.

Also, my computers stutters and freezes when trying to render .HEIC photos from my camera. The CPU churns in the background but doesn’t do anything. RAW previews are busted and show up as static, too.

Meanwhile, on CachyOS linux:

• Hardware acceleration is just… working.

• The KDE photo viewer previews everything, perfectly. And quickly.

• Its been this way all year. I’ve had this linux partition for over two years now.

Now, I’m not a linux fanboy. I still do most of my gaming on (stripped/neutered) Windows 11, as its usually faster. I don’t like that its faster, but when I A/B test games that need the performance, that’s what my benchmarks say.

But the days of Nvidia linux being wonky are over for me. The days where Windows is the “stable OS” are gone. In fact, the situation is flipped: I’m babying Windows 11 to try and keep it together, while Linux Just Works… as long as I don’t push it into areas it doesn’t like.

Yes.

• LACT for GPU: https://github.com/ilya-zlobintsev/LACT

• For CPU: https://wiki.archlinux.org/title/CPU_frequency_scaling#Configuring_frequency_boosting

The default KDE power saver profile also disables turbo, and is configurable exactly like you asked.

But, like others said, we can’t really help without any hardware info.

The twist:

It’s actually a flat piece of paper.

Like what? RT is not blocked for me; I read an article from there once in awhile. Neither is the Russian government website or anything else I can think to test.

In response to:

Because lemmy.ml federates with the broader Lemmyverse. So if instances like .world is blocked, that same content is visible on .ml so it’s blocked too. Defederate from us and we might have a chance.

lemmy.ml was not blocked because of lemmy.world. Or other federated instances, as far as I can tell through testing a few.

Check yourself. Go to any IP tester or http resolver, or try your own VPN.

As of this post, Lemmy.world is not blocked in China. Neither is piefed.social.

But lemmy.ml is blocked in China.

Don’t really see how that matters.

In other words, it’s… unfortunate, but necessary?

Every time I see such memes, I remind myself that lemmy.ml is blocked in:

• China (including Hong Kong)

• Iran

• North Korea

It is currently unblocked in:

• The United States

• Israel (and Palestine)

• Ukraine

• EU

• Pretty much the rest of the world

• Russia

Though I worry about Russia, as they’ve already blocked Bluesky and at least one part of the fediverse: https://lemmy.world/post/14156170

(And FYI I checked these regions myself, with a VPN, and then when another external tool, just now).

If you have to ask us, not your SO in your main relationship, then yes.

If you can talk to your SO about it candidly, then no.

It doesn’t even matter if there’s sexting or not: if you feel the need to keep it secret, that’s the problem.

Yeah.

Not gonna lie, it was a terrible idea from him.

The first thing the AI Bros want, and the absolute last thing I want, is for my country and social security to get entangled in a pyramid scheme.

I’m all for collaborative AI effort (like China) or some national funding to train open models as a utility (like Europe is trying but kind of struggling to get together). I’d be alright funding “infrastructure” companies like Huggingface or Cerebras, or maybe AMD/Intel with very very specific conditions. But Bernie’s proposal is basically the worst of everything.