3 ms·
GCP has a really strong ecosystem around AI. If you end up dealing with some type of ML problem with absurd memory requirements, TPUs can be useful there. A lot
by jeffshek 7y ago
GCP has a really strong ecosystem around AI. If you end up dealing with some type of ML problem with absurd memory requirements, TPUs can be useful there. A lot of recent Transformer architectures have these requirements.
Also, a huge amount of cloud costs are based on egress, so having it on one cloud saves you resources.
- DaiPlusPlus 7y ago> having it on one cloud saves you resources. All the major cloud storage providers tout geo-replication so that if an earthquake or meteor takes out a datacenter there’s still other copies. But nothing can protect against an administrative SNAFU where you - or your cloud provider - and accidentally or intentionally deletes your account and everything’s gone in an instant (yes, there’s usually recourse and backups to recover from - but that’s hours or days of downtime). Sometimes it isn’t even the cloud provider’s fault: see Adobe’s blameshifting onto the US executive for their dropping all their Venezuelan customers. While Azure has the “Black Forest” unit which is legally - and technologically - firewalled off the main Azure cloud - and you can also license Azure Stack for running many (most?) of Azure’s platform on-prem or self-hosted, I haven’t seen similar offerings from AWS or Google.
- jeffshek 7y agoI’m biased toward my use case - but if I have to move a bunch of terabytes for a ML data pipeline - I arguably am having stuff on GCP. To your point though, yes there is business risk of having it on one cloud.
- mamon 7y agoGCP also considers it copyright infrigement if you use their AI services for something actually useful, like, for example, self-driving car software. So it is better to stay away from it, unless you only use it for fun.