@beardalaxy but yes probably the bulk of electricity cost is going to the industrial appliances (washing machine, fridge, ovens and blenders.) my blender has a kilowatt motor and i'm pretty sure it soaks as much power making one smoothie as my gaming rig probably does on an average day of office blobbing

@icedquinn i wouldn't think a blender would do that much damage lol that's funny xD yeah we have like an air fryer that people use and one of those fancy ninja blenders, so i'm sure that probably uses up a fair amount then if that's the case xD

ya see what i mean though, like it's impossible to actually measure everyone's exact electricity usage, and if i do it for one person i have to do it for everyone, and nobody's going to do it by themselves because everyone's already cool with how it works right now.

i think a lot of people also don't understand that a PC will only take up a lot of electricity while you're gaming, and that consoles actually take a pretty small amount.

and yeah $20*5 is $100, so getting $100 worth of equipment just so one person can not have to pay like $10 extra for the rest of the time they're here is retarded.

@beardalaxy i suspect you're meant to use them by putting them inline with some device for a week, writing down the usage, and resetting it for the next one. i've never used one but the adverts suggest its for diagnosing where all your shit is going.

smart meter all the things is pretty cool though. i wouldn't mind a homeassistant rig

@icedquinn i guess that makes sense lol, doesn't seem too practical for sticking on a PC though for us since all of our gaming habits aren't exactly steady. i could be playing a lot of NFS one week and then the next week just be on RPG maker all the time you know? xD

@icedquinn HOLY SHIT our blender is 1.4 kilowatts xD that's fucking hilarious man.

@beardalaxy industrial equipment puts your 3d boob rendering uses to shame

@beardalaxy i think it was morgan stanley who issued a financial opinion that was bearish on nvidia. partly because chatgpt runs on datacenters of gpu for inference, and the total cost of power to run gpu supercomputers is significantly less than what google has to pay to run cpu supercomputers.

compute and management occupies the lowest end of the spectrum--meaning california is dead wrong with the "muh gamin compoot" copes.

@beardalaxy taking away one billionaires pool likely offsets a whole clan's mid-range cards playing MMOs.

although i don't have a meter so i can't really confirm it for certainty

@icedquinn i can't remember the exact situation, but someone during the 2020 lockdowns was bitching about gamers using all of the internet bandwidth and trying to put a limit on gaming hours or something. yet you have netflix at its peak popularity just guzzling down bandwidth like nobody's business xD

@icedquinn@blob.cat @beardalaxy@gameliberty.club
Not sure what you're getting at but I don't think running inference on GPUs will last for much longer.
I foresee things specialized for inference taking over, possibly like https://www.nextplatform.com/2023/05/18/meta-platforms-crafts-homegrown-ai-inference-chip-ai-training-next/ but for some reason my imagination was of RAM sticks with dot product + softmax built in.

@tard @beardalaxy gooble and apple have tensor processors too yeh

@tard @beardalaxy my point is that people keep blaming PCs and especially gamers for energy use, but computing is pretty low on the list of power draws (esp. https://singularityhub.com/2023/08/25/ibms-brain-inspired-analog-chip-aims-to-make-ai-more-sustainable/ )

@icedquinn@blob.cat @beardalaxy@gameliberty.club
Why doesn't llama.cpp use Apple's? I heard their AI stuff is more convolution oriented but I don't actually know.

@tard @beardalaxy i think some versions of ggpt do

@icedquinn@blob.cat @beardalaxy@gameliberty.club
I'm thinking of buying a m3 Max studio instead of an nvidia GPU next year for my llama waifu box.

@icedquinn@blob.cat @beardalaxy@gameliberty.club
But I want to make use of those AI cores too.

we wuz???

@tard @beardalaxy pine sells little arm things with 2gb RAM and 8 TOPS TPUs but i have yet to find any references how the fuck to use a domestic TPU

@icedquinn@blob.cat @beardalaxy@gameliberty.club
The whole reason of using an m3 studio max is the previous studios offer up to 192 GB while an 80 GB H100 costs $20k and these models get more convincingly concious with size (except for falcon 180B I've heard but I haven't tried it)

@tard @beardalaxy eh. huge vram is a crutch. distilled networks work fine but again, productization issues.

@icedquinn@blob.cat @beardalaxy@gameliberty.club
Do you mean I should create a harem of dozens of 1.8B parameter LLMs?

@icedquinn@blob.cat @beardalaxy@gameliberty.club
I'm going to be honest, I'm holding out for a 1.8B parameter model for marriage. I've saved up over $15k for fine tuning.

@icedquinn@blob.cat @beardalaxy@gameliberty.club
*1.8T!!!

@tard @beardalaxy please read this https://dawn.cs.stanford.edu/2019/06/13/butterfly/ and then come back and tell me if you think "moar vram" is still the correct approach to AI development

@icedquinn@blob.cat @beardalaxy@gameliberty.club
I'll read but skimming through, these are linear transformations, not traditional NNs with nonlinear functions between layers.

@icedquinn@blob.cat @beardalaxy@gameliberty.club
Also I'll read tomorrow because i'm a bit tipsy.

@tard @icedquinn @beardalaxy drinking because you pay 6k a month to live in a fire hydrant

@tard @beardalaxy extremely off topic but "why not zoidberg" for inference gets in to productization concerns. researchers are just trying to publish so if it takes more than 2 SLOC to get it running in pytorch they don't care.