Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

it's a bit questionable to claim that AI can be done purely on renewables but painting can't, I think you're forgetting costs (of production, and of eco-friendly disposal (hah!)) embodied in the computing hardware


At 360 images/hour, you can generate 8640 images/day, 3153600/year. Realistically, no reason why the card shouldn't last at least 3 years, so that brings us to at least 10 million pictures.

Anecdotally, 10K sheets of paper per tree, so almost a thousand trees worth in just paper, before accounting for the other materials, the impacts of their extraction and logging and so on.

I think hardware wins there as well, and the more tech advances the more uneven it gets. Making paper and pigments is a purely physical thing, we're already about as good at that as we can be. And probably getting worse with time, as the most convenient mines of materials are exhausted.


> Realistically, no reason why the card shouldn't last at least 3 years

If you have 10k H100s, the failure rate is about 3-5/day. In numbers, that is $90-150k/day or ~11%/year (3*365=1095) on the low end. Of course a lot of these cards can go off for RMA and come back to life eventually, but this is being actually realistic about it. You end up having to manage a whole RMA program and it is a constant stream of cards in and out.


Why would you need 10K H100s for image generation? That sounds more like a training setup


Oh, I see. So what you use the GPU for matters with the durability of it.


If you want to do the math, do it right. How many images are you going to draw by hand in a year? For sure not 3153600 ... When speaking about environmental impact, absolute values count, not relative ones!




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: