3 ms·
> no serious ML research will use it for training anything else than toy models They will use it to edit code, deploy and remotely run jobs. Serious deep learn
by visarga 6y ago
> no serious ML research will use it for training anything else than toy models
They will use it to edit code, deploy and remotely run jobs. Serious deep learning on laptops is a non-starter, on any laptop, on account of heat dissipation.
- ChuckNorris89 6y agoDepends what do you mean by 'serious'? Training Tesla's FSD neural net on a laptop? A student training some models for courses or self study?
- WanderPanda 6y agoExactly! I would guess that maybe 80% of monthly active deep learning users would benefit from a small accelerator like that. In contrast to that, probably less than 1% of the deep learning flops will benefit
- 908B64B197 6y ago> A student training some models for courses or self study? Serious schools have computing clusters. ML researchers might be interested to know which undergrad is actually training models complex enough to benefit from better hardware.
- EveYoung 6y agoWith WFH nowadays, I wouldn't even want a GPU workstation running at home considering the noise and power consumption. And with the high costs of a capable DS team, runing these workloads in the cloud doesn't make much of a difference either.
- 3np 6y agoSounds like you haven't seen the cloud costs for processing units handling these workloads, then... They're severely overpriced compared to normal CPU loads.
- chungus_khan 6y agoAnd for that purpose it's nice, but as good as any other semi-modern laptop. Over quarantine I amused some coworkers by doing an entire day's work (not ML, but not stuff that's run locally) on an old VT220, with very little change in workflow. Being able to do light stuff on one's own consumer hardware without having to buy something new is still incredibly helpful to students and other people trying to learn, as well as hobbyists though.
- ravi-delia 6y agoI mean it's 'better' in that it runs cooler and faster for longer than anything price equivalent, but that advantage will eventually be normal through the industry.