- cross-posted to:
- [email protected]
- cross-posted to:
- [email protected]
This is again a big win on the red team at least for me. They developed a “fully open” 3B parameters model family trained from scratch on AMD Instinct™ MI300X GPUs.
AMD is excited to announce Instella, a family of fully open state-of-the-art 3-billion-parameter language models (LMs) […]. Instella models outperform existing fully open models of similar sizes and achieve competitive performance compared to state-of-the-art open-weight models such as Llama-3.2-3B, Gemma-2-2B, and Qwen-2.5-3B […].
As shown in this image (https://rocm.blogs.amd.com/_images/scaling_perf_instruct.png) this model outperforms current other “fully open” models, coming next to open weight only models.
A step further, thank you AMD.
PS : not doing AMD propaganda but thanks them to help and contribute to the Open Source World.
Stop snipping sentences in half as if that’s what I fucking said!
“So does AMD” means IT IS ARBITRARY, in direct contradiction to when you declared, “it’s not arbitrary.” I only mentioned “arbitrary control of the platform” referencing your defense of Nvidia’s “control of the API” which you still declare is absolute. Experience cognitive dissonance, god damn you!
AMD is not behind. AMD is excluded.
AMD’s compute capabilities are presumably on-par, dollar for dollar, just like their graphics capabilities. Yet I have to say “presumably” because Nvidia’s spent a king’s ransom ensuring that direct competition does not occur.
Nvidia has reserved the APIs for GPU compute to themselves. CUDA performance on AMD isn’t worse - it’s nothing. It is excluded. AMD does not have a chance to run CUDA programs. It could! - as evidenced by multiple projects which ran CUDA on AMD hardware - but the threat of legal action ensures one vendor controls the de-facto standard for GPU compute.
Having only one vendor that matters is what a monopoly is.
I’m snipping the relevant part so you or someone else can go look it up without it adding too much noise to my comment.
Arbitrary means there’s no reason or purpose for it. They want to iterate quickly, and being forced to stick to a standard gets in the way of that.
If you want a standard, OpenCL is right there. Performance sucks because neither major vendor has cared much about it, but that’s the cross platform API to use.
AMD is only excluded in the sense that Windows APIs don’t work on Linux directly. But just like how WINE exists, HIP exists (AMD created) to run the same code on both. But then you run into the very obvious issue that AMD has a motivation to optimize for their GPUs and not Nvidia’s.
But why deal with that abstraction if you only use Nvidia GPUs anyway?
No, a monopoly has exclusive control of a market. We also use the term to mean a company that abuses its massive market share and therefore acts like a monopoly, such as:
Each of those engaged/engages in anti-competitive behavior with lawsuits, technical barriers, and essentially bribes.
Not sharing your IP isn’t a “technical barrier,” no company is under any obligation to do that. AMD has demonstrated that it can build a bridge to CUDA and ROCm, so unless Nvidia is directly sabotaging users of that or a similar bridge (and not just changing their API for other reasons), they’re not acting like a monopoly.
Arbitrary means they don’t need a reason. Nvidia’s control is total. It was always describing a thing you said. This is the third time I’ve explained this, and you keep pretending I must mean something else. I’m tired of tabooing words to trick you into accepting your own stated opinions.
Like insisting monopoly is absolute, and in the next sentence, acknowledging the definition everyone actually uses. The next sentence.
Nobody’s talking about sharing. These are legal barriers to letting AMD do the work, themselves. They are not permitted to compete directly. Open options don’t matter because adoption is a feature you can’t design. CUDA has a monopoly.
The entire crux of this conversation is you suggesting Nvidia would sue AMD for announcing CUDA support. I shouldn’t have to convince you of this. What are you doing?
There are two common definitions of “monopoly”:
I’m using the first, and I don’t think Nvidia meets that mark. They’re not winning because they’re making unfair agreements with datacenters, making deals with software vendors to drop/not add ROCm and OneAPI support, or abusing the legal system to block competitors’ products. The closest they get is the threat of a legal battle if AMD violates their copyright or appears to violate their copyright.
Nvidia has copyrights on CUDA, so they get a say in how it’s used. Whether APIs can be copyrighted isn’t a totally solved legal question, and I’m guessing there are still some unsolved questions around fair use as well.
So I don’t see Nvidia threatening to sue AMD over copyright of CUDA as particularly anti-competitive, because it is their IP afterall.
Nvidia has simply out-innovated AMD, such as:
And what has AMD innovated? They’re largely playing catch-up because Nvidia takes more risks and invests more in GPU R&D.
This was also the case with Intel v AMD between the big lawsuit resolution and Ryzen, when AMD did essentially a Hail Mary with the chiplet design after years of lagging behind, which paid huge dividends in manufacturing costs, allowing them to focus on specific parts of the chip and enjoy better yields. AMD is sort of innovating by bringing chiplet design to GPUs, but it absolutely feels like too little too late, and that’s just going to improve profit margins, it won’t fix the feature gap.
AMD has poor GPU market share because they consistently released mediocre products, and were late to each of the big shifts in GPU tech. They have no right to Nvidia’s IP, and the longer they keep playing catchup, the harder it’ll be to unseat Nvidia’s dominant position.
But freely sharing it, and half-assing the AMD version - that would be monopoly behavior.
Uh huh.
And a bridge is proof that deliberate incompatibility is not anti-competitive… but you’d endorse Nvidia suing if AMD really used it.
Uh huh.
Thank goodness you’ve backed this up with court cases saying APIs are bad to withhold and fine to re-implement, or I’d start to think it’s all a bit fucking silly.
Nvidia performs fire and motion. “It’s just cover fire so that they can move forward and you can’t.” They tried proprietary physics, proprietary sound, proprietary hair, and a bunch of other nonsense besides the two and a half things that turned out to matter.
Every single one of them was pulled out of their ass specifically so they could go “AMD can’t do [blank]!” and avert direct competition. They used to bribe developers into implementing their latest gimmick. I don’t know offhand if they told devs not to bother making it work on AMD hardware. But we can guess. Then - the moment AMD proved their hardware could obviously do the same kind of computations - Nvidia dropped that gimmick and pulled out a fresh one.
CUDA is the only one of those ATI-era stupid GPU tricks that AMD did not recreate. That’s why it’s still around, and HairWorks is not, even though games still have hair. Expect raytracing to stop getting any attention once it’s not an “incomparable advantage.” (DLSS might stick around simply because modern game performance blows.)
AMD has poor GPU market share because Nvidia had a higher high end and used every ounce of their market advantage to lock in customers and prevent fair competition.
And you’d endorse Nvidia suing if AMD made it apples-to-apples.
What? I never claimed that.
When did I endorse Nvidia doing anything?
That’s called R&D. You try a bunch of stuff and keep what works. And yeah, marketing depts will spin that the best they can to get people on board. That’s how competition works.
Source? That sounds like monopolistic behavior if done with clear market dominance.
This sounds like AMD not innovating on their own and instead playing catchup.
I doubt that. Raytracing is a bit gimmicky, but it also dramatically improves visuals, so it’s unlikely to go away anytime soon. Maybe they’ll generalize it to a CUDA routine, but I doubt the approach is disappearing.
The first part matters for gaming cards, where people like to be on the “winning team” or whatever. That’s why AMD has been slurping up CPU marketshare in the DIY market, they took the performance crown and that matters for average consumers.
But that doesn’t matter as much for datacenters, which care a lot more about efficiency and cost. Workstations care about top end performance, but they’re not a big part of the market.
AMD lost the datacenter for two reasons:
Nvidia established a lead and maintained it. That’s all.
“It’s their IP, after all.”
Competition being that thing with only one vendor for [blank].
Where that vendor expends treasure to ensure games run well only on their hardware.
… the entire back half of the previous comment is about how Nvidia actively prevented AMD from innovating, by forcing them to play catch-up. What you consider a counterargument is endlessly fascinating.
CUDA’s monopoly and CUDA’s monopoly. One option that matters, one vendor who sells it. All to run your own god damned code.
Nvidia established a lead and exploited it. We have a word for that.
That’s a statement of fact, not endorsement.
So, supporting their customers is somehow a bad thing?
That’s not how that works. AMD could have made something useful instead of just chasing whatever Nvidia was doing. When you’re in an arms race, you don’t just keep up, you leap frog. AMD did neither.
Again, CUDA doesn’t have a monopoly. There are multiple compute frameworks that work on Nvidia GPUs, like OpenCL, OpenMD, and OpenACC, so you’re not even locked in to CUDA if you want to target Nvidia. AMD actually seems to do more locking in than Nvidia by only providing ROCm for workstation cards.
CUDA has dominant market share, but it doesn’t have a monopoly.
And yes, Nvidia absolutely exploited their lead. I have yet to see evidence that they did so in an anti-competitive way, they just threw a bunch of money at R&D until they found things that worked. That’s their competitive advantage. AMD has been exploiting their performance lead in CPUs and have gobbled up market share from Intel. That’s how business works, when you have the better product, you make the most of it.
Having a better software ecosystem isn’t anti-competitive. That’s called having a good software ecosystem.
‘I’d approve this lawsuit’ is not fact.
Devs are not Nvidia’s customers.
Fire and motion is an advantage that only works if you have the lead… like most monopoly power.
CUDA’s competitors do not matter, because their market share is fuck-all. That’s how actual monopolies work. Not even Standard Oil ever had all of the oil. They just had enough that they decided how things works.
ROCm is open-source, so fuck right off with that bullshit. Non-workstation cards support ROCm because of this.
Dominant marketshare is what monopoly meeeeans. The only cogent nuh-uhs are bickering about exploiting a dominant marketshare, which hey look, is what you keep admitting Nvidia does. They just created proprietary way to run your own god damned code, in such a way that you, personally, would totally understand if Nvidia sued the pants off of AMD, if AMD ever dared to use the bridge that you, personally, insist is proof that Nvidia’s not excluding AMD.
Maybe you’re just full of shit.