AMD says future PCs will run 30B parameter models at 100T/s
AMD says future PCs will run 30B parameter models at 100T/s
They're gonna need a heck of a lot of memory bandwidth – not to mention capacity – to do it
You're viewing a single thread.
That's so unnecessary. The way we are going about "AI" is like brute forcing a password. Sure you'll get the job done, but it's the least efficient way to do it. I'm not saying I have the answer, but why not try and find a more efficient way to get the same results instead of building unnecessarily powerful PC's? I can only imagine how much redundancy is present in those LLM black boxes. It's just a word calculator.
I'm gonna go off on a limb and say that they're likely doing both.
Also going to go off on a limb and say both is better, plus AMD being a compute company has a.. Ahem.. Vested interest?... In the further utilization of ever increasing amounts of compute.
Well they are definitely prioritizing one over the other.
You definitely sound educated and certainly not like a novice just bitching. In terms of AI research, which methods being explored (or not) do you feel need more direct investment?
Look man, he said he doesn't understand anything. Why don't you just accept that everyone working in AI is stupid and there's a completely better way to do everything.
/s
Man, I sure do wonder which one is going to shown visible returns sooner? Fundamentally reworking how the models work, or simply duct taping more processing power to it?
Obviously the brute force method is going to show the most returns immediately, you're just throwing more resources at it. Efficiency gains take time. While it's absolutely a much bigger deal with AI, that's pretty much the path all all these things have. Crypto mining, ray tracing, 3d graphics, hell even all the way back to 2d graphics.
There's no magic "make run more efficient" button.
While I agree with you to a certain extent, these technologies always take different paths depending on those priorities. The thing with 3d and 2d graphics was that they were working with limiting technology. In fact I would even use that as an argument against just "building a better machine." Back then they had to make software work with the limitations of the hardware. You couldn't just duct tape two SNES's together and get better performance. They had to be efficient or have no product to even release. Nowadays you can just buy more computing power. Even when it comes to graphics there are so many companies that release unoptimized software onto the market because the consumer can just "build a better machine." Crypto has so much unnecessary redundancy that all of the computations just get thrown out the window while only 1 computer gets to add to the Blockchain and gets that reward for the actual mining.
Those older industries had more limitations than we did so they had to make it as efficient as possible. Now we have so much computing power there is no incentive to make things more efficient save for long term viability. Which none of these companies give a shit about as long as they are making money. I'm not saying they need to hit the magic "efficiency" button. I'm just saying they're lazy and making everyone else pay the price.
I'm not saying I have the answer, but why not try and find a more efficient way
??? You don't understand the problem yet you claim there's a better way that everyone has missed until now?
Well sure. That's everything.
"Planes are so unnecessary, why haven't they found a better way."
"CPU's have billions of transistors, why haven't they found a better way."