It became obvious when old e-waste Xeons were turned into viable, usable machines, years ago.
Something is obviously wrong with this entire industry, and I cannot wait for it to pop. THIS will be the excitement everyone is looking for.
It became obvious when old e-waste Xeons were turned into viable, usable machines, years ago.
Something is obviously wrong with this entire industry, and I cannot wait for it to pop. THIS will be the excitement everyone is looking for.
High-end GPUs are already useless for gaming (a low-end GPU is enough), their traditional source of demand. They're floating on artificial demand for a while now.
There are two markets that currently could use them: LLMs and Augmented Reality. Both of these are currently useless, and getting more useless by the day.
CPUs are just piggybacking on all of this.
So, lots of things hanging on unrealized promises. It will pop when there is no next use for super high-end GPUs.
War is a potential user of such devices, and I predict it could be the next thing after LLMs and AR. But then if war breaks out in such a scale to drive silicon prices up, lots of things are going to pop, and food and fuel will boom to such a magnitude that will make silicon look silly.
I think it will pop before it comes to the point of war driving it, and it will happen within our lifetimes (so, not a Nostradamus-style prediction that will only be realized long-after I'm dead).
This is the exact model in which WWII operated. Car and plane supply chains were practically nationalized to support the military industry.
If drones, surveillance, satellites become the main war tech, they'll all use silicon, and things will be fully nationalized.
We already have all sorts of hints of this. Doesn't need a genius to predict that it could be what happens to these industries.
The balance with food and fuel is more delicate though. A war with drones, satellites and surveillance is not like WWII, there's a commercial aspect to it. If you put it on paper, food and fuel project more power and thus, can move more money. Any public crisis can make people forget about GPUs and jeopardize the process of nationalization that is currently being implemented, which still depends on relatively peaceful international trade.
Dude, you're describing the 80s. We're in 2025.
GPUs will be used for automated surveillance, espionage, brainwashing and market manipulation. At least that's what the current batch of technologies implies.
The only thing stopping this from becoming a full dystopia is that delicate balance with food and fuel I mentioned earlier.
It has become pretty obvious that entire wealthy nations can starve if they make the wrong move. Turns out GPUs cannot produce calories, and there's a limit to how much of a market you can manipulate to produce calories for you.
From a market perspective, LLMs sell GPUs. Doesn't even matter if they work or not.
From the geopolitical tensions perspective, they're the perfect excuse to create infrastructure for a global analogue of the Great Firewall (something that the Chinese are pioneers of, and catching up to the plan).
From the software engineering perspective, LLMs are a nuissance, a distraction. They harm everyone.
Really? What about textures? Any ML that the new wave of games might use? For instance, while current LLMs powering NPC interactions would be pretty horrible, what about in 2 years time? You could have arbitrary dialogue trees AND dynamically voiced NPCs or PCs. This is categorically impossible without more VRAM.
> the perfect excuse to create infrastructure for a global analogue of the Great Firewall
Yes, let's have more censorship and kill the dream of the Internet even deader than it already is.
> From the software engineering perspective, LLMs are a nuissance, a distraction. They harm everyone.
You should be aware that reasonable minds can differ in this issue. I won't defend companies forcing the use of LLMs (it would be like forcing use of vim or any other tech you dislike), but I disagree about being a nuisance, distraction, or a universal harm. It's all down to choices and fit for use case.