4 ms·
> Elon's comments on the latest earnings call, if you read between the lines, strongly imply that Dojo v1, while functioning, has been a disappointment Maybe t
by ShamelessC 3y ago
> Elon's comments on the latest earnings call, if you read between the lines, strongly imply that Dojo v1, while functioning, has been a disappointment
Maybe they shouldn’t have gone with an in-house design.
> it would seem foolish to limit the system to internal customers only.
My understanding from working with similar HPC systems is that public use cases are likely niche and many tasks don’t benefit from the sort of fast interconnect their setup gives them. If anything being quite public about their design was meant to allow other companies to copy their design and compete with nvidia on a united front perhaps.
It’s not really as versatile as AWS, GCP are to profitably work as a compute cloud provider though. Perhaps it could be a provider for other companies who don’t want to own their hardware to do training runs on.
- modeless 3y agoI agree that it seems a bit much to try to design their own training supercomputer. But they seemingly did OK with their inference hardware, and they do have an awful lot of money these days. Anyway, even if it is literally only useful for a single task, and that task is training large transformers, there is a use case right now. People are desperate for hardware to train large transformers.