Once the overall AI bubble bursts/deflates, would they be able to re-purpose it for other workloads e.g. cloud gaming?
Though, maybe Midjourney or other niche AI companies could buy it for the right price I guess.
back
4 comments
The server/ML GPUs are not great for gaming, they strip out all the specialized shader/pixel units to cram in more general compute cores.
In theory you could write games with just those but aside from the amount of work needed I'm not sure if the performance would be good enough, the specialized texture samplers etc can be faster than general purpose compute shaders.
Also, for cloud gaming you want very low latency, so few GPUs all over the world in local POPs, not a lot of GPUs in few large data centers.
This feels like saying when the dot-com bubble burst, could the servers be re-purposed as mainframes. Whatever about valuations and individual stocks, AI/LLM workloads are only going up.
I am more on the AI skeptic side (LLMs are not a path to general intelligence, most positions cannot be 100% replaced with any version of product I have seem to date), but I fully agree. Some form of AI text generation is likely to be here forever. Maybe it gets vastly more efficient per clock cycle, but compared to N years ago, all of the tech vendors need more compute to offer this service.
It doesn’t matter if its general intelligence. What matters is if they can get workers and consumers using it every day. Which increasingly seems to be the case.
Putting aside possible specialization of hardware and differences for applications, cloud gaming was tried and largely flopped. Related for the power intense use-cases where it could be useful to outsource had too much latency involved. Remember Strada?
"Trainium accelerator" doesn't sound like it would be terribly useful outside it's current niche. It's not clear how much of the spend is that kind of thing versus general purpose compute and storage.
For the previous generation, AWS had a split where Trainium == model training and Inferentia == inference. With the current generation, they are also recommending smaller Trainium configurations for inference.
Even if nobody trains another model, the Trainium chips will be useful for inference of existing models that have been recompiled for it.