Hacker News Reader

Comment by mschuster91 | Hacker News Reader

Latency.

Latency for interaction with AI? For something the user use browsers and electron apps. A ping goes around the whole world before your pressed key is visible on the screen.

> A ping goes around the whole world before your pressed key is visible on the screen.

I don't disagree with your sentiment, but as someone who manages infrastructure which is literally on the opposite side of the planet (from me) RTT is long enough to be objectively horrible for anything interactive like typing over SSH. At least 250ms capital to capital for Europe to Oceania + last mile latency makes 300ms a more realistic minimum + actual processing time + client side rendering and it really starts to feel like a half second from the full human perspective... Under ideal conditions.

> Latency for interaction with AI?

It's not just AI. Video streaming or game streaming both depend on close proximity to the end user, alone because the data volume would completely overwhelm backhaul links... that's why Youtube, Netflix and a lot of CDNs all have colo facilities close to major population centers and their regionally dominant ISPs.

It can be significantly cheaper to pay colo fees at an ISP' s data center than to pay interconnection fees for bulk transfers.

> Video streaming or game streaming both depend on close proximity to the end user

That's a straw man for this discussion. Yes, there will be a need to place some DCs close to users, but most of them could be located closer to energy production, attached to a backbone network, and the vast majority of customers would not suffer any degradation of service from it.

Increasing latency can also reduce TCP throughput ceiling, increase in-flight concurrency, and increase server memory consumption.

Dafuck? No way lol

Playing from eu on us servers is not playable

>A ping goes around the whole world before your pressed key is visible on the screen.

I don't think that gaming is the focus of the DC discussed in this article. I don't think gaming will ever require a massive buildup as AI nowadays.

Are they more playable if the server is in the nearest coast compared to the middle of the US? If you're in the US, for something like AI inference, then RTT within the continental US is probably reasonable. Similar is likely true through most of Europe.

Doesn’t matter for training or most AI use cases. We are talking tens of ms latency here.

I’m gonna guess availability of employees may matter more.

Once built, the DC has like, ten employees. Nine security guards and a part-time on-site technician.

Does it matter for training?

I'd assume that site selection managers understand how recommending "only useful for training" locations might impact their careers.

Whats the relevance of fibre optic latency to electric lines?

Either the data center is in the wind belt and you run fiber to the people, or the data center is near the people and you run power to the source. Either way you have to do some kind of cable run, and that means land.

There was a massive dark fiber buildout in 2000. The hyperscalers already have long haul fiber optic infrastructure everywhere in US.

fiber is a lot easier to run under houses than hvdc.

Not all loads require low latency to end users.

Reply on news.ycombinator.com