Forwarded from Nic
The meme "be patient dwarves" is a local meme among Vanomas's viewers that has spread beyond the depths of the internet. At this point, it no longer matters who the quote originally belongs to. The essence of the trend is to attribute it to various and unexpected characters.
Forwarded from Nic
This has even more print screens with a citation being attributed to Gimly from Two Towers and Transformers movies, Hitler, etc
Under the comments this post where I found it:
V.:
Like you can run a good few models at home, i know for a fact there is no shot that even the enterprise ones burn that much power.
Edoardo | 00uno00:
For context, the most carbon-intensive image generation
model (stable-diffusion-xl-base-1.0) generates 1,594 grams
of 𝐶𝑂2𝑒𝑞 for 1,000 inferences, which is roughly the equivalent to
4.1 miles driven by an average gasoline-powered passenger vehi-
cle [51], whereas the least carbon-intensive text generation model
(distilbert-base-uncased) generates as much carbon as 0.0006
miles driven by a similar vehicle, i.e. 6,833 times less
V.:
An inference in this context refers to running an input through the model, in this case, putting in a prompt and generating an image.
So, 1000 Inferences, 1000 prompts generating an image, for stable diffusions biggest model at the time of writing the study, produced (according to the methodology of the study) 1.594 grams of C02 emissions.
Generating a single image according to this study is therefore roughly equivalent to 0.001594grams of C02/image generated, or (according to the calculations of this study), moving your car 0,0041 miles/about 6,6 meters).
Converting this to the power needed to charge a phone is a bit tricky, but for the sake of this i will provide some different fossil fuel sources to illustrate my point.
A very bad fuel source, Hard coal, generates about 338g CO2/kwh produced ( my souce for this is the 1st google result, feel free to do more research but i feel like this is probably adequate https://www.volker-quaschning.de/datserv/CO2-spez/index_e.php). For the cost of 0.001594g of CO2 we can generate ~4.71598e-6kWh or ~ 4.71598mWh. Average phone battery in my head is like 3000mWh, which may be a little small these days but thats where i would put that, so this would be roughly equal to charging your phone by 0.157199%, not quite enough to charge it very much at all. And thats being generous and assuming 100% charger efficiency and no loss in the power lines, switching stations, etc, post generation.
But lets see a more efficient power source than hard coal. Lets compare one of the better fuels you can burn, Natural Gas. Natural gas generates about 201g/C02 per generated kWh, which means we can generate roughly 7.93035e-6 kWh or ~7.93035 mWh, about a 0.264345% charge for your phone.
This is still a pretty big amount for a singular query, but like. its not close to 50%. Unless i am missing a major piece of info here, that claim is wrong, and very much so.
I do not like generative AI. I think it is inhuman, it is theft, it is evil, it only serves to enrich people that would grind me into mulch at the slightest indication of profit.
But i also dont think we need to fucking lie about it to make that point.
And the claim that one prompt consumes that much power makes no thermodynamic sense either. I am kinda done making my point so i dont feel like doing more math, but if you look at the power consumption of one nvidia A100 gpu(300/400W), the datacenter gpu these things usually run on, burning that much power on a singular prompt makes no sense. The amount of time these things would have to spend on a singular prompt wouldnt make the math of charging your phone by 50% work out at all.
V.:
Like you can run a good few models at home, i know for a fact there is no shot that even the enterprise ones burn that much power.
Edoardo | 00uno00:
For context, the most carbon-intensive image generation
model (stable-diffusion-xl-base-1.0) generates 1,594 grams
of 𝐶𝑂2𝑒𝑞 for 1,000 inferences, which is roughly the equivalent to
4.1 miles driven by an average gasoline-powered passenger vehi-
cle [51], whereas the least carbon-intensive text generation model
(distilbert-base-uncased) generates as much carbon as 0.0006
miles driven by a similar vehicle, i.e. 6,833 times less
V.:
An inference in this context refers to running an input through the model, in this case, putting in a prompt and generating an image.
So, 1000 Inferences, 1000 prompts generating an image, for stable diffusions biggest model at the time of writing the study, produced (according to the methodology of the study) 1.594 grams of C02 emissions.
Generating a single image according to this study is therefore roughly equivalent to 0.001594grams of C02/image generated, or (according to the calculations of this study), moving your car 0,0041 miles/about 6,6 meters).
Converting this to the power needed to charge a phone is a bit tricky, but for the sake of this i will provide some different fossil fuel sources to illustrate my point.
A very bad fuel source, Hard coal, generates about 338g CO2/kwh produced ( my souce for this is the 1st google result, feel free to do more research but i feel like this is probably adequate https://www.volker-quaschning.de/datserv/CO2-spez/index_e.php). For the cost of 0.001594g of CO2 we can generate ~4.71598e-6kWh or ~ 4.71598mWh. Average phone battery in my head is like 3000mWh, which may be a little small these days but thats where i would put that, so this would be roughly equal to charging your phone by 0.157199%, not quite enough to charge it very much at all. And thats being generous and assuming 100% charger efficiency and no loss in the power lines, switching stations, etc, post generation.
But lets see a more efficient power source than hard coal. Lets compare one of the better fuels you can burn, Natural Gas. Natural gas generates about 201g/C02 per generated kWh, which means we can generate roughly 7.93035e-6 kWh or ~7.93035 mWh, about a 0.264345% charge for your phone.
This is still a pretty big amount for a singular query, but like. its not close to 50%. Unless i am missing a major piece of info here, that claim is wrong, and very much so.
I do not like generative AI. I think it is inhuman, it is theft, it is evil, it only serves to enrich people that would grind me into mulch at the slightest indication of profit.
But i also dont think we need to fucking lie about it to make that point.
And the claim that one prompt consumes that much power makes no thermodynamic sense either. I am kinda done making my point so i dont feel like doing more math, but if you look at the power consumption of one nvidia A100 gpu(300/400W), the datacenter gpu these things usually run on, burning that much power on a singular prompt makes no sense. The amount of time these things would have to spend on a singular prompt wouldnt make the math of charging your phone by 50% work out at all.
Volker Quaschning - Erneuerbare Energien und Klimaschutz
Specific carbon dioxide emissions of various fuels
In conclusion: The second panel of the infographic above is inaccurate. We don't need to misrepresent data to have enough reasons to hate genAI. While generating a single image doesn't charge your phone to 50%, training the actual model still consumes way too much energy, and, as a result, produces way too much carbon