A beginner's guide to deploying LLMs with AMD on Windows using PyTorch

I kind of feel the same way, but my take may not be quite as cynical. The way I see it, some companies just act really stupid, and there's a lot more room for forgiveness when you suspect incompetence than active disregard. There's a lot of talk about how amazing Lisa Su is as a leader, and while it's true that she has executed brilliantly on the goals she dictated for the company, the record will show that those goals were rather short-sighted in the long run.

It's been painfully obvious since at least 2018 that GPUs were going to explode in value as Nvidia developed an increasingly large monopoly on AI compute. In response to this, AMD put almost no-one on the driver stack. At the time, it was baffling. Now it's a near Blockbuster-level failure in corporate decision making. They reap what they sow.

But what will benefit the consumer in the long run is to break the monopoly. There are no real sides to take here other than "not Nvidia". If AMD provides a decent product, it makes sense to buy it to show that the support is there. It's not been sunshine and roses, but my recent experience with the 7900 XTX has been positive and substantially better than anything they had to offer in 2020. The fact that the AMD stack is still so weak is not a product of their lack of improvement, but rather how astonishingly far behind they were.

Having said that, the recent GPU prices from AMD have been baffling. If they don't improve, they'll continue to be in the dust.

What will likely break the monopoly is a Chinese competitor producing a worse but usable product within the next couple of years. AMD will be the next Compaq.

Indeed, but it will take more than a few years, the entire industry negging GPUs for a decade and half comes at a high price for them. And looking back historically, dominant companies like Intel had been sucking wind for over a decade before its ineptitude caught up with it. Microsoft was repeatedly punching itself in the face until Satya Nadella took over and pulled it back from the brink just by getting them to stop doing so. And a quarter century later, CSCO has nearly recovered to its dotcom peak stock value.

But Nvidia isn't inept yet, and despite some of its people being just as arrogant as Googlers at Google's prime, they still deliver a vastly better product and ecosystem than anyone else. The Chinese will crack that, but the first few generations are going to be a rough ride. Just check out the iterations on any knockoff of DeWalt tools they sell with cheaper parts until they nail the price/performance ratio to perfection to see how they roll. But also, Nvidia remains a moving target with near bottomless pockets now. Good luck with that.