Hacker News Reader

Comment by rwmj | Hacker News Reader

The difference is the hardware will become obsolete in a few years, unlike the dark fibre (really, rights of way) that provided cheap connectivity for years after the crash.

View on HN

The hardware will become obsolete but we may end up with an oversupply of cheap power - overbuild on things like Nuclear and solar. That would be nice.

Do we have any good data on how many of the GPU burn out given hard usage?

Are we talkinga bout loosing 50% of the hardware as it fails? Or far less?

It's an inherently different thing. Fibre is infrastructure. It'll still function decades later. GPUs at this scale are consumables. I've heard 3-5 years lifespan before they fail out. This might be a low estimate, but even if you double it - they're 50% of the cost of datacenters. We're flushing entire countries worth of economic value down an Nvidia shaped toilet.

Anyone got any good testing data? Surely someone knows rough failure rate from constant usage?

It's not that they fail (which they also do), it's that they become obsolete. They consume more energy to do fewer operations relative to newer nodes. This is the basis behind Moore's law.

Ah gotcha, ya that will be interesting to see, but does seem like a more likely life of 7 to 10 years is possible.

I am curious on actual hardware failure rates due to heat etc.

I tried to do a calculation.

https://news.ycombinator.com/item?id=48762428

So 10% failure rate per gpu year. Some might be repairable and there might be a failure rate acceleration as time passes etc. So very approximate numbers.

It's obsolete for training but is it obsolete for inference?

The factories won't be obsolete, even if the chips are.

The fabs for current nodes will be obsoleted by newer nodes.

not if they completely crash manufacturing with their push for fascism; be more optimistic!

Reply on news.ycombinator.com