Hacker News Reader

Comment by Nelson69 | Hacker News Reader

Not just is it intact, but is there really an alternative ML/AI platform out there? They're still the best (only?) game in town with the full stack. And you can go buy or lease access to their parts right now and get radical performance improvements for your model training.

View on HN

kseems like a lot of startups are trying to build things equivalent to TPUs. Also one can use FPGAs for inference(?) maybe also training?

I think a lot of people see an opportunity to try to take a slice of the big compute market for ML/AI. who knows if these alternatives really are viable....

but google has a lot of smart people and building your own silicon seems like a great idea these days..so I think other people will attack their stack.

AI is a pretty tractable workload for GPUs/TPUs/FPGAs. Inference is moving to many more places, the days of Nvidia being the solution for AI are over. It will still be the goto for research, but as things mature, the nets will literally run anywhere there is a MAC (multiply accumulate).

It may run, but will it run as cost-effectively for the same amount of performance as it would on a GPU?

yeah, you have GPU like chips being developed by a lot of companies invested in AI. From server chips for training, down to cell phone chips for cheap inference.

I am really hoping to see AMD entering the battlefield. AMD GPUs are more than capable of handling the workload it is just a software problem.

Reply on news.ycombinator.com