AMD’s Untether AI Deal - Bad Signs for GPU-Driven AI training

Posted by EconomyAgency8423@reddit | hardware | View on Reddit | 8 comments

“AMD’s acquisition of Untether’s engineering group is proof that the GPU vendors know model training is over, and that a decline in GPU revenue is around the corner,” Justin Kinsey, president of semiconductor recruitment firm SBT and an investor in Untether wrote on LinkedIn.

[-]

PM_ME_YOUR_HAGGIS_@reddit

The premise makes sense - the mega capex training era will slow down and specialised inference chips will be the new norm for prod deployment.

[-]

Jeep-Eep@reddit

This would be basically indistinguishable from the AI crash dieback of essentially useless LLMs and similar techs that's been in the wings for a while - this sounds like an IP set well optimized for actually useful forms of ML tech such as the stuff you find in modern client GPUs.

[-]

PorchettaM@reddit

I don't really get why training and inference would be mutually exclusive as the article seems to be assuming.

[-]

scytheavatar@reddit

As this article explains

Companies have devoted immense resources to building AI models over the last two years or so. Now training these massive models is more of a one-time affair that requires considerable computing power and Nvidia has been the biggest beneficiary of this, as its GPUs are regarded as the fastest and most efficient for these tasks. However, the AI landscape may be shifting. Incremental performance gains are expected to diminish as models grow larger in terms of several parameters. Separately, the availability of high-quality data for training models is likely to become a bottleneck as much of the Internet’s high-quality data is already run through large language models. Considering this, the significantly front-loaded AI training phase could wind down. The underlying economics of the end market for GPU chips and the broader AI ecosystem are weak, and most of Nvidia’s customers likely aren’t generating meaningful returns on their investments just yet.

[-]

PorchettaM@reddit

Long term, it makes sense for there to be diminishing returns on training. I just question it being "around the corner" as the OP article is claiming. I can see a prolonged period of GPUs and specialized hardware coexisting as inference demand ramps up before training demand can slow down.

[-]

Jeep-Eep@reddit

I don't terribly buy the idea of 'maturing models, lower training!!!' here for what we talk about with 'AI' these days... maybe for applications like say, the RedStone featureset, but for something supposed to public face? It would need regular patching.

[-]

chefchef97@reddit

Sounds familiar

[-]

Jeep-Eep@reddit

If it gets those bastards to stop dipping into the client models before this bubble BS - and it reads to me a coded acknowledgement the party is winding down without sending the shareholders into tantrums - is over, this is all the better.