i feel like "woke" has lost all its meaning to me
#programming
olive sable
azure lynx
red is just a saturated pink.
split steppe
flat uni theory
azure lynx
both colors are on the spectrum. don't worry.
split steppe
pink is not a color
olive sable
why not?
olive sable
both colors are on the spectrum. don't worry.
you did mean the autism spectrum right?
azure lynx
literally the electromagnetic spectrum but if the colors were people, they'd be pretty focused
split steppe
na pink is just diluted red
mighty bane
na pink is just diluted red
But at least it exists.
azure lynx
pink is red + white (and white is all the (light) colors mixed together)
mighty bane
Unlike purple.
jovial remnant
Leon is a very good example for potential, I've also seen one image comparison of a mage from Hogwarts legacy that's way better than what they used in their video examples
olive sable
purple is also just pink
split steppe
But at least it exists.
it doesnt exist because its not in the bible
azure lynx
yellow doesn't exist really for people. it's just green and red together.
olive sable
it doesnt exist because its not in the bible
you dont exist because you're not in the bible
checkmate liberal
split steppe
actually i do, my given name is from the bible, checkmate
split steppe
wouldnt you like to know weather boy
azure lynx
not going to dox anyone, but joshua is the most common name in the US.
or possibly michael. it depends on the year
(for boys anyway... )
split steppe
the name just picks whatever it feels like being, every year
azure lynx
but the name doesn't refer to you. checkmate.
split steppe
not from the us, check your mate 
azure lynx
don't have to be. the name i n the bible doesn't actually refer to you. so your existence isn't proven.
olive sable
i think most of us aren't from the usa
split steppe
it refers to me specifically actually, im the only surviving person from the bible
trim valve
i think most of us aren't from the usa
considering the time in the us right now id be suprised
olive sable
bro is noah
split steppe
considering the time in the us right now id be suprised
plenty of neets in the usa who dont sleep
azure lynx
oh. the legendary person who heard "there are those among you who will live to see me come again" speech.
they have no name,
split steppe
also its like morning in the usa
azure lynx
yeah. it's early.
split steppe
everyones awake already
trim valve
also its like morning in the usa
does your average neurocord chatter wake before 10am
azure lynx
depends if they've been the sleep yet
trim valve
split steppe
or me, every one of you is a robot!
azure lynx
i'm currently detached from my real time zone due to things i'm working on.
trim valve
dead internet theory but everyone is neur
split steppe
i'm currently detached from my real time zone due to things i'm working on.
man's in a pocket dimension rn saving time
the real hustler
trim valve
bouncy lines
azure lynx
davinci is my favorite enough free video editor.
sage crag
wrr
tender river
kdenlive 
trim valve
losing my mind
honest dagger
I wish davinci resolve had a light theme
I think your the first person to ever say that haha
nocturne olive
losing my mind
Doing some sort of wizardry?
trim valve
I think your the first person to ever say that haha
its useful when you have bright light (the sun) on your monitor
otherwise I get why resolve uses a dark theme
honest dagger
its useful when you have bright light (the sun) on your monitor
Ahhh that might actually be useful 😅
trim valve
I've had this gripe with other blackmagic tools before
couldn't see the "on" button in the atem audio mixer
sick owl
worldly plank
I'll keep it off, thank you very much.
silent island
How is this kid not traumatized? Lmaooo
trim valve
couldn't see the "on" button in the atem audio mixer
like for reference
the software looks like this
and I was on an old lcd screen in the middle of a field
honest dagger
like for reference
What are you working on? dangg
trim valve
that's just an example screenshot from blackmagic lol
rigid snow
that is audio silly
trim valve
some idiot (me) forgot that microphones need constant attention otherwise you have a lot of work to do in post
trim valve
NOT podcast
rough bloom
podcast 
maiden geyser
I think your the first person to ever say that haha
there can only be one dave
split steppe
just saw dlss5 and i'm astonished how detached these companies have become from reality
they're so proud of the destruction of artistic intent it's wild
it's so incredibly bad i don't even have words to express it
silent cloak
there can only be one dave
silent island
chat, should I upgrade my router?
ive been using the same router since 2011 so...
should I?
cant believe ts outlived my TV lmaooo
rough bloom
also depends on your internet connection I guess
maiden geyser
chat, should I upgrade my router?
wpa2 in 2k26
silent island
consider obtaining ethernet cable instead
even on wifi ts lowkey cursed
rough bloom
also depends on your internet connection I guess
if you have practically no local traffic then you don't need a fast local network 
so if you can saturate your internet connection it's good enough™
azure lynx
literally sitting less than the width of my hand from the computer.
but it's also plugged in at Gbit.
yay comcast equipment
rough bloom
even on wifi ts lowkey cursed
try Cloudflare speed test mabe https://speed.cloudflare.com/
it'll also show packet loss and latency
silent island
<:Hmm:889032126622728192> try Cloudflare speed test mabe https://speed.cloudflar...
uhhh is dis gud?
rough bloom
yeah that seems pretty good
latency is high but I think that's just because of the speed test works
but no packet loss, even on WiFi
fast pagoda
36 might be the highest ive ever recorded
tender river
rough bloom
packet loss 
fast pagoda
lol it really just says >1Gbit if you go above that
midnight sigil
sage crag
mobile data in this room is faster
moved phone by 30cm
rough bloom
or http2 maybe
seems to be pretty much the same result
tender river
fast pagoda
the cli speedtests suck
tender river
sage crag
how can it say my internet is unsuitable when this is the fastest my internet has ever been
fast pagoda
theyre daring you to be beter
willow path
tender river
fast pagoda
speedtest++ is fairly accurate
speedtest-cli uploads liek 2 mb so it's super random values
rough bloom
mobile data in this room is faster
mobile data NOT faster
sage crag
<:NeuroHuh:1071414397752332429>
they have fibre to the box and fibre to the property, i have vdsl
froge
tender river
rice very neutral
fast pagoda
im about to rip someone's head off at verizon
tender river
i tried putting mitarashi dango glaze (made of soy sauce and sugar) on bread
tender river
i tried putting mitarashi dango glaze (made of soy sauce and sugar) on bread
it tasted like
soy sauce
but on dango it tasted like
sweet soy sauce
sage crag
trim valve
<:neuroPogHD:1458850559976144959> mobile data NOT faster
I can see my mobile data tower
rough bloom
I can see my mobile data tower <:neuroPogHD:1458850559976144959>
I figured out why the result was bad 
mobile ISP contract renewed
trim valve
I can see my mobile data tower <:neuroPogHD:1458850559976144959>
would post picture but even google ai can doxx me from the building
sage crag
armed and ready dt pass b rank
trim valve
would post picture but even google ai can doxx me from the building
sometimes I google stuff and get hyper specific results even in incognito so i question its accuracy
but do not want to find out
tender river
armed and ready dt pass b rank
fast pagoda
verizon decided there's no way anyone could ever want to so you can't disable 5g band to just use 4g
rough bloom
fast pagoda
and the 5g tower they have near me has some sort of known issue
for months
so i just have no data half the time
trim valve
do you not get that option in your phone settings?
trim valve
i forget the us does that
tender river
m
we have some preinstalled apps in russia
in our case its mostly just the government wanting to prop up local businesses
trim valve
candy crush?
fast pagoda
oh yeah king software very niche local dev
when i called verizon about the simply non functional nature of their 5g tower here
they were like
yeah known issue
we'll give you a discount
but
they did give me a discount
$10/month
in order to do so
they removeed the $20/month off discount i already had
so they increased my bill by $10/month as a treat
tender river
fast pagoda
and the service still doesn't work a month later 
wish i could just tell people "workin on it" for months at work
azure lynx
wish i could just tell people "workin on it" for months at work
you can. for about 3 months usually.
flint dome
why are you using so much active time
i have the same question with my ram
8gb idle
like what is windows doing
trim valve
how much total?
flint dome
16
flint dome
its eating half my ram
my fedora takes up only 1
and its bloated
are they just genuinely throwing updates at the wall with an llm and seeing what passes production tests
trim valve
i mean
do a task that uses a bunch of memory and then kill that task
see how much memory you're actually using
nocturne olive
I've actually done that test on a laptop with 8GB RAM
The result was that it was still using like 1.8GB after
On Win10
After I had debloated it a bunch
(the test was running 2000 instances of minesweeper and then killing them)
flint dome
its showing ghost values for some reason
when I go through theres not a lot of apps using memory but the overall usage is 8
did I get a virus lmfao
maiden geyser
<:what:586938028065423377>
pagefile on ssd
trim valve
what
are we just making shit up these days
oh wait no I do in fact have pagefile: yes
idk where my pagefile even is
apparently I have a 3GB pagefile
somewhere
maiden geyser
the image says there is a pagefile on C:
nocturne olive
when I go through theres not a lot of apps using memory but the overall usage is...
Windows doesn't tell you about all of its memory using components
trim valve
not that it got used
trim valve
Windows doesn't tell you about all of its memory using components
crazy how the process list doesn't show memory not being used by processes
nocturne olive
Windows do be silly like that
rough bloom
Linux does the exact same thing
idle dune
What a cute cat
flint dome
Windows doesn't tell you about all of its memory using components
im going to build my own zilog z80 cpu running cp/m and ooga booga like the old times
free of any bloat as I will rewrite the os in rust
tender river
running rust compiler on z80 cpu
young oyster
<:enub:1163912230585237655>
olive sable
tall people have bigger eyes, so more light enters their eyes, so they can see better in the dark
but the image is less sharp
so short people see in 4K
azure lynx
the reason tall people can see better in the dark is because their eyes can pick up longer wavelengths.
like infrared.
olive sable
that shouldnt matter since the cones in the eyes pick up very litle to no infrared
azure lynx
if the person was larger in every way (including at the atomic level) it would work
azure lynx
larger antennae can pick up longer wavelengths.
the optin in the cones in your eye would respond to lower frequency light coz they're bigger now
makes as much sense as saying tall people can see better in the dark.
olive sable
humans only see 380nm to 750nm
thats cuz IR doesn't have enough energy to trigger the chemical reation in our cones.
azure lynx
i was being silly. short people don't have higher resolution eyes either.
sage crag
the reason tall people can see better in the dark is because their eyes can pick...
height doesnt upscale the size of your cones
cell size remain uniform across height
olive sable
i think they're doing a hypothetical
azure lynx
and larger eyes would increase sharpness too
olive sable
if the person was larger in every way (including at the atomic level) it would w...
i guess if the electrons were further from the nucleus, but without increasing the size of either so the force keeping them together doesn't scale. then it would be able to see IR
tender river
height doesnt upscale the size of your cones
it does
if you're over 3 meters tall you can see microwaves
sage crag
and larger eyes would increase sharpness too
depends on the density of photoreceptors rather than the surface area
lens size does have an impact on visual clarity though
olive sable
if you're over 3 meters tall you can see microwaves
only 1 in N people can see this object
sage crag
and having more space for the focus to be adjusted allows better range of distances, though you'd probably be long sighted compared to a regular sized human
azure lynx
the larger the aperture, the higher the resolving power.
sage crag
and having more space for the focus to be adjusted allows better range of distan...
due to the minimum adjustment being larger
olive sable
lens size does have an impact on visual clarity though
the reason im going on this dumb tangent is cuz i was thinking of lenses ye
azure lynx
i remember reading that Legolas in lord of the rings would need eyes like a meter across to be able to see as far as he did in the story
azure lynx
magic lenses floating in front of them that focus it all down to their eyes
that'd work
coz magic
like binoculars
but magic
sage crag
elfs (elves) may have completely different internal anatomy to humans
maybe they have four stomachs like cows
olive sable
elfs (elves) 
fickle rain
Or they’re more magic than physical and thus don’t give a shit about the limitations
sage crag
`elfs (elves) `<:neuroCatUuh:1131629092475768884>
made up word linguistics debate
elfs elves
say one and someone will cry
azure lynx
https://www.youtube.com/watch?v=Rk2izv-c_ts <= was this one
olive sable
elves might have a mirror at the back of their head, and the eyes extend into the body, allowing them to be 1m long
olive sable
say one and someone will cry
olive sable
what about tolkien?
azure lynx
so the solution is "Elves don't exist, so they can do anything they want."
olive sable
Tolkien strongly preferred the plural elves (and dwarves/dwarvish) over "elfs," correcting editors to maintain this usage.
huh
azure lynx
the elves themselves spoke Elvish and not English, so didn't give their opinion.
olive sable
the only thing i can really blame about the lord of the rings trilogy is.
they used ballistas to attack a city, but that city use trebuchets to defend
it should have been the other way around
azure lynx
my main fault was the trilogy takes longer to watch than the flight from Los Angeles to Auckland. definitely too long.
sage crag
the elves themselves spoke Elvish and not English, so didn't give their opinion.
https://en.wikipedia.org/wiki/Elvish_languages_of_Middle-earth
there are a few
olive sable
my main fault was the trilogy takes longer to watch than the flight from Los Ang...
i bingewatched the whole thing a month or so ago
and the pacing is actually pretty good
just a long story
azure lynx
i was actually pretty sick of all the media about it (and am originally from NZ so it was x 10 over there) and have only watched the 1st one.
azure lynx
i've read the hobbit. i think it took less time to read than the movie.
like, i read it in a weekend while i was visiting a friend
without being rude
olive sable
i didnt read any of the books.
i was supposed to for a book report tho. but i ended up winging it
sage crag
i've read the hobbit. i think it took less time to read than the movie.
the hobbit is a children's book
olive sable
sage crag
its still a good book, but its very short
azure lynx
i know ;] it's quite good though. i only read it coz i saw it was so small
olive sable
i like books short
rigid snow
azure lynx
like i said, i was visiting someone
sage crag
i like books short
the hobbit is 95000 words apparently
azure lynx
was it really?
sage crag
novels are usually anywhere in the range of 70000 to 120000 words
azure lynx
that seems long for what i remember
olive sable
i read the first chapter, and then made the assignment with help form chatgpt
to get an idea on how the story goes, i wrote most of it myself
azure lynx
how many partial degrees do you think chatgpt has earned since they switch it on?
olive sable
a couple million
olive sable
thats more than 95K
damn
oh
no i was reading the wrong one
specifically the hobit is 95k ye
azure lynx
wow. i must've been ploughing though it. then again, i have no idea how many pages or words books have these days as i use a kindle
olive sable
i lose focus after about 4000 words
sage crag
according to this chart
average reading fluency
but 350 seems a bit low for an avid reader
sage crag
unsure, my reading speed is around the 400wpm range though, which is why i brought up the 4 hours
azure lynx
if i'm focused i can read as fast as my eyes move.
sage crag
if i'm focused i can read as fast as my eyes move.
eye movement is naturally the bottleneck for reading speed
azure lynx
there's a trick where you move your finger down the page and you can read faster, but that only works for about 2 pages for me before i start zoning out
sage crag
moving eyes between chunks reduces eye strain and improves speed
azure lynx
and i like to read dialog "out loud" in my head
sage crag
moving eyes between chunks reduces eye strain and improves speed
rather than continuously moving the eyes
sage crag
and i like to read dialog "out loud" in my head
i read at whatever speed is comfortable to absorb all the information
azure lynx
places with lots of dialog are slower usually... depending on how they speak. if they're being direct and to the point it's probably faster than if they're being all clever.
sage crag
i read at whatever speed is comfortable to absorb all the information
unless im bored and then i start skimming between major events
but that's unusual
azure lynx
when i'm reading quickly the danger is i get three pages on and realize i have just been thinking about lunch for the last 3 pages
azure lynx
blinking is the next bottleneck
tender river
this random reading test says i have 391wpm
i unfocus very easily though
azure lynx
when i'm reading quickly the danger is i get three pages on and realize i have j...
really the danger is when I read a particularly clever line and I'm just repeating it to myself and marvelling while moving my eyes over the words.
rigid snow
me
azure lynx
me: squirrel
sage crag
non-skimmed
but easy article
you could guess these questions easily even if you skimmed the text
tender river
i wouldnt be surprised if your peak reading speed was higher
though i suspect most of the difference comes from better focus
azure lynx
i think the trick would be to write the article using completely ficticious groups with counterintuitive properties.
azure lynx
like the logic problem "If all things that are smoked are good for you, and cigarettes are smoked, are cigarettes good for you?" ... the answer is "Yes." (even though the premises is false, if it was true, then cigarettes would be good for you. i just asked about that set of facts in that specific (not true) state.
cigarettes are not good for you, but "IF all things that were smoked were good for you" THEN it would be true. even though that is not true.
sage crag
but language doesnt exist in a vacuum
azure lynx
the logic was what is being questioned, not the facts. the reading test also was a test of background knowledge
sage crag
i exist in a vaccum
bred get moldy if not vacuum seal ye
azure lynx
in space nobody can hear you wrrr
language doesn't exist in a vacuum because sound needs a medium to travel through.
i guess sign language could though. so language does exist in a vacuum. (deliberately misinterpreting the meaning of "exist in a vacuum" for humorous purposes)
the roomba uses a protocol to define what mode it should be in and how it should behave in various situations. this is a lanaguage and the protocol specifies what is a valid "sentence". so language does exist in a vacuum.
split steppe
i exist in a vaccum
i exist in a false vacuum
olive sable
in space nobody can hear you wrrr
iirc if you touch them you can hear them
split steppe
transferring vibrations from one to another
olive sable
thats one way to put it
azure lynx
today is a sad day. a Computerfile video was titled "Writing a guitar sound in C". They lied. It was writing a guitar processing effect in C. And (at least the initial one) was just an overdrive effect.
writing a gutiar sound would've been much more interesting... simulating the strings.
olive sable
what is the difference?
split steppe
i kinda gave up on computerphile because 90% of their videos became about ai in recent years
it's just boring seeing the same topic over and over again
azure lynx
for an effect you still need a guitar.
azure lynx
it's saying "for this array, perform this function on all the samples"
which isn't very interesting until you get to resonant filters
split steppe
today is a sad day. a Computerfile video was titled "Writing a guitar sound in C...
if you were simulating the strings maybe, but in practice even if they did the guitar sound actually, it's just sine waves
aint nobody simulating the actual strings
could tie back to their FFT videos maybe
azure lynx
yeah. i've played with an open sourced/public domain VST guiitar synthesizer before to make it so it would do electric guitar sounds instead of acoustic. it used simulated strings.
i wrote some fun sounding (but extremely cheesy) guitar solos that even did some effects like slides and some harmonics with it, but it took a hell of a lot of processor. now you could probably push most of that to gpu
split steppe
i wonder, daws are notoriously cpu heavy because of the dependent calculations you need for the whole simulation
not sure if a gpu would do a good job, it's almost never been used afaik
azure lynx
if you're only doing the generation of one instrument with it it would probably be fine, especially if you know the latency... assuming you're not using it live
for live, get a real guitarist ;]
split steppe
duh :V
but like in general, i dunno if a gpu would help much, audio sim isn't really a parallel task
split steppe
it being a vector op doesn't mean much, what about the dependencies?
azure lynx
basically they chop the string into sections and simulate each section. the only dependencies other than the position of the guitar pick are the two sections on either side
patent shard
split steppe
hmm, maybe i'm imagining what the string sim does wrong, got a visual example for it?
azure lynx
split steppe
https://www.mathworks.com/help/audio/ug/simulation-of-a-plucked-string.html
digi...
page doesn't seem to say much, example needs a whole ass account
azure lynx
there is a version.
so it basically makes an array of delay lines and uses those to hold the string position, bouncing the signal off the end of the string and/or where it is pressed. because it allows you do choose where you strum the strings and how much and what angle you put the pick on, you can make realistic sounds with it
most of what I did was removing parts coz i wanted a rougher sound. ;/
split steppe
when it comes to simulating real sound this comes to mind as the perfect example https://www.engine-sim.parts/
azure lynx
yeah same kind of "physically based rendering" but for sound idea
the pipes like exhausts in this simulation are delay lines like the strings except obviously they have different rules for how they work
trim valve
noway screeps update
olive sable
opaque sigil
es2026 screeps? 
minor sail
<:neuroPogHD:1458850559976144959> noway screeps update
from 10 to 24 is a big ass jump
minor sail
Awww yeahhh! Ya boy just got an employment offer!
I've been partially / fully unemployed for just under 2 years now
azure lynx
jobs are overrated.
i mean, except for the money part.
and if you're lucky, the access to cool toys part.
ok ngl some jobs kick ass
3/5 better than the alternative
leaden crest
how do yall survive without employment
I dont know much about adulthood
but isnt money kinda required
warped narwhal
if you live anywhere other than America - Government benefits
If you live in America - idk bro good luck I guess?
leaden crest
ah american here that makes sense
we are born into the world with more debt than we will ever have wealth
amber fractal
Staying with my parents here, thank goodness they are so kind
olive sable
how do yall survive without employment
i dont have to pay rent, food, electricity, nor any other bills
also gov benefits i guess
altho most of it goes into cool tech
timber dagger
working on my presentation while listening to neuro21 radio station and money machine comes up
my focus was never higher for those 3 minutes
minor sail
how do yall survive without employment
I sold my home and moved in with my parents lol
chrome rampart
olive sable
proven merlin
Okay guys im very confused. Why is FP16 so much slower on this (or probably other pascal cards aswell) P40?
180Gflops fp16 vs 11tflops fp32? Why not just run the models in fp32 then? Vram useage id guess, but cant we at runtime trade some computation per layer to cast the values between fp16 and 32? Perhaps cast the next layer while the current one runs?
https://www.techpowerup.com/gpu-specs/tesla-p40.c2878
amber fractal
Okay guys im very confused. Why is FP16 so much slower on this (or probably othe...
That is tech called AMP, and is normally used to go from f32 to f16
proven merlin
Well f16 is supposedly slower here.. can amp do the opposite? And how much does it matter for Q4/Q5 gguf models?
nocturne olive
Okay guys im very confused. Why is FP16 so much slower on this (or probably othe...
No native FP16 presumably
proven merlin
No native FP16 presumably
I guess so too but why not just use fp32 then?
nocturne olive
In the case of my 3090 it has 1:1 FP32 and FP16 because it emulates FP16 or something
nocturne olive
I guess so too but why not just use fp32 then?
Memory
proven merlin
Memory
Surely casting between fp16 for the inactive ones and fp32 for the active and next one cant be that bad?
mighty thorn
Get out your notes on Neurosynth
nocturne olive
This must be the stupidest thing to have been put in a Systemd service
I hate Python environment management
proven merlin
Surely casting between fp16 for the inactive ones and fp32 for the active and ne...
I mean, doesnt gguf cast int(whatever) to fp16/fp32 anyway?
nocturne olive
When it comes to compute itself, usually if you want to do high throughput you want to stick to the same dtype as what is in memory
In inference you can get away with the compute being worse due to having to do casts because the mem bandwidth is the limiting factor anyway
proven merlin
So.. it sounds like that casting between fp32/fp16 sounds like the thing to do then
Or ig Int4/5 to fp32 instead of 16 or whatever
nocturne olive
If you really want to do inference with FP16 weights, yes
If you're training though, better get an FP16 capable card
proven merlin
Just interested in cheap ways of running models
nocturne olive
The P40 is certainly a cheap VRAM thing
proven merlin
And kinda seems insane to me that pascal seems so capable but just not used at all due to bad fp16
amber fractal
Last time I ran a model on one of my p40s it was about 2-3 years ago
nocturne olive
This must be the stupidest thing to have been put in a Systemd service
I find something about this extremely cursed
proven merlin
I find something about this extremely cursed
The ./ Seems bad to me.. idk tho
amber fractal
Just fell out of interest and now I don't have a setup to use them in
nocturne olive
The ./ Seems bad to me.. idk tho
The service sets the working directory before that script runs
proven merlin
Ah
proven merlin
Just fell out of interest and now I don't have a setup to use them in
I mean.. 48gb for 500€? Hows that not interesting?
nocturne olive
48GB with so little compute it's useless and I'd rather use that 500€ on a 3090
proven merlin
I can barely get a 3090 for 750€
nocturne olive
I got my 3090 for 500 €
proven merlin
U lucky
nocturne olive
Nah I just asked the triangle to find one
proven merlin
The triangle?
olive sable
proven merlin
Oh ig
amber fractal
I mean.. 48gb for 500€? Hows that not interesting?
I ended up getting 96gb 700 euro give or take.
proven merlin
Thats two triangles tho
amber fractal
price of a single 3090
proven merlin
Yea
olive sable
I can barely get a 3090 for 750€
prices have risen since the ram crisis. there was a good week or so i could get them for 400
amber fractal
so I just traded in compute for memory
proven merlin
48GB with so little compute it's useless and I'd rather use that 500€ on a 3090
Im mostly interested in models like qwen3.5 35b a3b.. so compute aint a biggie
amber fractal
I have my 3060m for training, smol models sure but I still consider larger models to be a skill issue.
nocturne olive
I find something about this extremely cursed
The story behind this:
For whatever stupid reason, today Streamlink decided it would be a good idea to bork
So then I had to go and first figure out what the issue could be, then I ended up recompiling my stream capturer wrapper, and even then it still didn't work
Then I got it working in my regular terminal but not as the Systemd service
So then I found out that the stupid Streamlink in the service was the wrong version
So then I went to try and upgrade it, but Python is stupid and I could not figure out where in the name of Neuro-sama did I install it originally
So I ended up having to make a virtual environment dedicated to the program specifically
nocturne olive
Im mostly interested in models like qwen3.5 35b a3b.. so compute aint a biggie
The 3090 could handle that fine
nocturne olive
The story behind this:
For whatever stupid reason, today Streamlink decided it w...
At least it happened today and not tomorrow
That would have been baaaaadddddd
proven merlin
The 3090 could handle that fine
Well, for 750€ vs 2x250€ .. the p40 is considerably cheaper
nocturne olive
Mabe with your prices
proven merlin
And would handle more context in theory
nocturne olive
With high context comes high compute need
proven merlin
Mabe with your prices
I mean yea i dont see any decent prices in germany
nocturne olive
Ah of course, Germany, the prices there are messed up
tender river
how do yall survive without employment
extremely cheap rent
proven merlin
With high context comes high compute need
True, TTFT increases, but im happy with like 100k tokens and willing to wait like half a minute or something
olive sable
RTX 5000 for 785 popped up locally
rough bloom
Okay guys im very confused. Why is FP16 so much slower on this (or probably othe...
apparently the P40 does have native fp16, it's just horrible because there aren't many fp16 cores on the die
in the Pascal generation only the GP100 (in the Tesla P100) has the newer fp32 cores that support fp16 by running at half width
for the other chips there are still some fp16 cores to have any fp16 support at all, but only a few to avoid wasting space
proven merlin
Ah of course, Germany, the prices there are messed up
Yea. Its awful.. i was happy to get a mined on rx6700xt about 2-3 years for 205€ .. still going strong.
nocturne olive
True, TTFT increases, but im happy with like 100k tokens and willing to wait lik...
Meanwhile me running the higher quality 27B model at like 37 t/s with more CTX than that
olive sable
superbox stop flexing 
proven merlin
Meanwhile me running the higher quality 27B model at like 37 t/s with more CTX t...
Lemme just..
olive sable
i got scammed out of 428 bucks cuz you wanted another 3090
nocturne olive
I'm just doing my normal usual sillying
nocturne olive
i got scammed out of 428 bucks cuz you wanted another 3090
Hem
Hopefully you can get those back
olive sable
thanks for clarifying
nocturne olive
I may also be stupid as well who knows
Can't even get a stupid Systemd service to work without a cursed thing
proven merlin
Lemme just..
This is sorted cheapest first..
olive sable
i can get them from like 700 on a good day
olive sable
worth it
proven merlin
But, according to them the card is hella loud and runs insanely hot (apparently some crappy palit model)
olive sable
if you dont want it ill buy it from you
proven merlin
Repasted and repadded
nocturne olive
But, according to them the card is hella loud and runs insanely hot (apparently ...
Re-paste and it's good
proven merlin
It is repasted
nocturne olive
Hem
proven merlin
Ill be allowed to play with it but im not too keen on those 16gb vram
nocturne olive
16GB is kinda fine
I used to train vocal synthesizers on 12GB
proven merlin
Especially in combination with my 12gb rx6700xt and my (poor) 700W psu lmao
amber fractal
nocturne olive
That is indeed a good deal for a 4080
If you buy that 4080 can I buy your old 3090?
4080 has a decently higher gaming performance than 3090
olive sable
If you buy that 4080 can I buy your old 3090?
no
nocturne olive
Aw
olive sable
i need a 2nd gpu for the theatre room pc
proven merlin
thats a good deal
Yea, and normally id take it immediately but its not like i need it, and since its a friends deal idk if id feel fine selling it and stuff
olive sable
ia lready have a pc with a cpu, but no gpu
nocturne olive
4080 is 33% faster than 3090 according to this performance chart
nocturne olive
i need a 2nd gpu for the theatre room pc
Hem
I see
Though an Nvidia GPU will mean it'll idle at at least 20W no matter what
proven merlin
Yea, and normally id take it immediately but its not like i need it, and since i...
I mean tbh probably would be fine selling it.. but at the same time, id like to run bigger llms.. how does vulkan deal with an amd card and a nvidia card at the same time?
nocturne olive
If you were in Germany you'd care, with them having the world's highest electricity prices
olive sable
I mean tbh probably would be fine selling it.. but at the same time, id like to ...
i program in vulkan
so
not great
nocturne olive
If you were in Germany you'd care, with them having the world's highest electric...
(I know from my friend Xoda)
proven merlin
If you were in Germany you'd care, with them having the world's highest electric...
My server is tuned to sip 28W and i dont run it 24/7 LMAO
olive sable
also, why you using vulkan for llms?
proven merlin
also, why you using vulkan for llms?
Rx6700xt no likee rocm
nocturne olive
Nvidia drivers are funny, no matter what you do there will be a default power consumption of 20W/GPU
proven merlin
And also, if i have the rx6700xt + rtx4080 in one system, don't i have to use vulkan?
olive sable
the graphics api?
proven merlin
Yea
olive sable
for llms?
nocturne olive
Vulkan is also used for compute
proven merlin
On my rtx2060 laptop vulkan vs cuda is like 2t/s difference (at a scale of 30t/s)
azure lynx
you're just running the same code as a shader
nocturne olive
llama.cpp for example has Vulkan compute capabilities
obsidian mantle
why can i not shrink it like this 
RE::TESContainer* container2 = ((RE::TESContainer*)((RE::TESObjectCONT*)((RE::TESForm*)((RE::TESObjectREFR*)((RE::TESForm*)a_ref))->data.objectReference)));
RE::TESContainer* container3 = (RE::TESContainer*)(a_ref->data.objectReference);
visual studio does 1st line and it works
i thougth these casts are excessive and did 2nd line
but its different
olive sable
if vulkan is used for ai the same way it is used for gamedev. then it will probably run fine with both, but you need to have it work on different things, or it will need a copy of the data on each gpu
nocturne olive
And also, if i have the rx6700xt + rtx4080 in one system, don't i have to use vu...
Presumably if you want to use both GPUs
Or mabe if llama.cpp lets you compile both GGML_CUDA_ON and whatever AMD needs...
Who knows
proven merlin
I dont particularly want tensor.. what was it called? Im fine with layered offloading, since aslong as it stays in vram, its comfortably fast (30t/s) with my rx6700xt and id assume the rtx4080 will have 0 trouble keeping up
nocturne olive
It will definitely be speed limited by the slower GPU
proven merlin
I also was looking at the intel gpus, but idk.. >1000€ for 48gb vram of vulkan only seems iffy
proven merlin
It will definitely be speed limited by the slower GPU
Yea
nocturne olive
I also was looking at the intel gpus, but idk.. >1000€ for 48gb vram of vulkan o...
The Intel consumer series isn't that bad price wise though?
12GB GPUs you can stack a bunch of for cheap
proven merlin
Yeah, consumer only sadly really goes up to 12-16gb tho..
Tho im curious what intels are you talking about?
nocturne olive
The B580 is really good value wise
proven merlin
4080 also "only" has 16gb
Which is why im not too stoked about it
Even if its hella value
nocturne olive
If one doesn't want a 4080, one sells the 4080 to someone who wants the 4080
olive sable
id gladly trade you a RX 7900 XT for a 4080
proven merlin
If only they werent prone to die, thats a hella offer.. if shipping doesn't kill the deal
nocturne olive
That seems like a great trade in this situation
24GB for 16GB
AMD so it works with your existing GPU too
gentle haven
Is 4 tb’s enough storage.
nocturne olive
If only they werent prone to die, thats a hella offer.. if shipping doesn't kill...
If Belgium to Finland is cheap, Belgium to Germany gotta be even cheaper with no lake in the way
olive sable
That seems like a great trade in this situation
24GB for 16GB
AMD so it works wi...
the XT is only 20, but i can probably get my hands on a XTX too
nocturne olive
Is 4 tb’s enough storage.
No
nocturne olive
the XT is only 20, but i can probably get my hands on a XTX too
Wa
Since when?
olive sable
If Belgium to Finland is cheap, Belgium to Germany gotta be even cheaper with no...
to germany i would just save costs and meet in the middle
gentle haven
Dang need more
olive sable
Wa
Since when?
since always
nocturne olive
Huh
I did not know that
proven merlin
Well, if that could work out, that would be hella cool.. ill keep it on mind and keep you updated
olive sable
azure lynx
electricity in germany is expensive coz they closed all their nuclear coz they panicked after Fukushima iirc
proven merlin
Now next issue: will my 700W bequiet system power 9 blow up if i hook both cards up lmao
proven merlin
electricity in germany is expensive coz they closed all their nuclear coz they p...
More or less, tho i think Chernobyl was the turning point or something..
gentle haven
electricity in germany is expensive coz they closed all their nuclear coz they p...
Is this programming 😭
gentle haven
Everything is code at this point
olive sable
im ngl, the rx 7900 xtx seems to be a lot better than the 4080 in everything besides RT
this is a terrible trade
nocturne olive
Now next issue: will my 700W bequiet system power 9 blow up if i hook both cards...
355 + 230W is a bit tight
olive sable
except the 7900 xtx is cheaper
proven merlin
this is a terrible trade
Im not even gonna lie, i expected that, but i haven't kept up to date at all
nocturne olive
im ngl, the rx 7900 xtx seems to be a lot better than the 4080 in everything bes...
Seems like in this performance chart I'm looking at it's exactly equal
proven merlin
355 + 230W is a bit tight
Plus a ryzen 7 3700x xD
nocturne olive
Yeaaah you may want to invest in a PSU upgrade
proven merlin
Seems like in this performance chart I'm looking at it's exactly equal
I mean the 7000 series did target the 4080
proven merlin
Yeaaah you may want to invest in a PSU upgrade
I mean rn im running a 5 2600x.. but that aint better
I dont mind undervolting and wattage limiting but yea
Its.. tight
nocturne olive
So basically with a 7900 XTX you'd be trading a 4080 for a 4080 level GPU that works more efficiently and better with Linux and has more VRAM
proven merlin
Yup, giving up cuda, gaining vram, "gaining" actual rocm
And "good" raytracing.. which i couldn't care for at all
nocturne olive
I specifically need CUDA because NeuroSynth runs only on either CPU or CUDA on Linux
On Windows it'd run on more but I don't touch slop OS any more
proven merlin
Only bummer is no dlss but.. oh well, i cant even max out my rx6700xt in games
proven merlin
I specifically need CUDA because NeuroSynth runs only on either CPU or CUDA on L...
Have you tried zluda and the other thingy?
I haven't yet and i wonder how it performs
nocturne olive
I do not have any AMD GPUs that I could test on
proven merlin
Ah fair enough
nocturne olive
If I had AMD GPUs I could probably write a patch for NeuroSynth to run on that too
olive sable
storage isnt an issue
nocturne olive
Holy a Quadro 48GB??
proven merlin
Its very tempting to throw out my money at a DGX but i know better.. and i need it for a driving license lmao
olive sable
i know a guy with one, it sucks
proven merlin
Oh?
olive sable
according to them
proven merlin
It seems like a nice machine to run 70b models on
azure lynx
yeah. don't get a driving license ;P
nocturne olive
It seems like a nice machine to run 70b models on
If you mean that tiny thingy with 128GB of slow unified memory, it's not gonna be very nice
azure lynx
except for going places, what use is a car anyway?
proven merlin
If you mean that tiny thingy with 128GB of slow unified memory, it's not gonna b...
Yeah not 4k€ nice
proven merlin
except for going places, what use is a car anyway?
Job, to buy more car
To do more job or something
Idk
nocturne olive
Yeah not 4k€ nice
I'd rather take 4K€ worth of 3090s than that thing
proven merlin
Yea
proven merlin
If that thing was like 1k€ it would be a different story but like
olive sable
It seems like a nice machine to run 70b models on
supposedly thermal throttles so only reaches half of its tflops. the gpu is ass, and the software support is abysmal
arm stuff
nocturne olive
That'd be 192GB VRAM
Just casually loads up a 200B model on that
proven merlin
That'd be 192GB VRAM
Well, 96 or so for me but sure
proven merlin
supposedly thermal throttles so only reaches half of its tflops. the gpu is ass,...
Nothing a bit of modding and a dremel dont fix ^^
I wouldn't hesitate modding it lmao
nocturne olive
Just casually loads up a 200B model on that
Or alternatively NeuroSynth with 4096 batch size
opaque sigil
oh yeah nvidia's runs on the hotter side, though putting it on the side apparently helps a lot
nocturne olive
Or alternatively NeuroSynth with 4096 batch size
In 8-way parallelism
olive sable
oh yeah nvidia's runs on the hotter side, though putting it on the side apparent...
laptop cooler
proven merlin
oh yeah nvidia's runs on the hotter side, though putting it on the side apparent...
I mean yea.. the cooler slot is wide, has input and exit next to each other iirc.. so putting it on its side and letting it rise at the back makes sense
opaque sigil
well yeah what did you expect
nocturne olive
laptop cooler
That thing looks like a lunch box, not a high power computer
opaque sigil
and where is big fan going to go
proven merlin
Just waterblock it
proven merlin
Just waterblock it
Honestly.. how bad would this be
opaque sigil
good luck
nocturne olive
Well the cooling would be like way bigger than the computer itself so you'd basically just be redesigning the whole thing
nocturne olive
What's the specs on that?
nocturne olive
Hm, I see
opaque sigil
beefed up 2080ti
nocturne olive
Looks like mem is 2/3rds 3090 speed
And the compute is not that far behind
olive sable
half the tflops of a 3090 too supposedly
proven merlin
i can get a titan RTX for 475
They start at 700€ here lmao
nocturne olive
half the tflops of a 3090 too supposedly
Oh you're looking at FP32 I guess
nocturne olive
3090 does have nearly double the FP32
But in FP16 they're like 3TFLOPS apart
proven merlin
Does lamacpp and such care about fp16 or fp32 when running quantizisied models?
nocturne olive
I'd assume it uses FP16 if the card supports it
olive sable
it supports both iirc
proven merlin
And no way to force fp32 with on the fly casting i assume (is it really that costly?)
olive sable
fp16 should be the standart mode???
nocturne olive
Titan RTX doesn't seem that bad
olive sable
look
azure lynx
some parts of the model are still fp16 or fp32 even when quantized (as far as i understand it)
olive sable
its 475, 500 with shipping, and i need to recoup some of the money i lost on that 3090
so ill do 550 for it
nocturne olive
Hmm
olive sable
preferably more if you can spare some change
olive sable
"used in a research environment" supposedly
nocturne olive
I assume it's the base model cooler?
olive sable
ye
nocturne olive
Are you capable of visiting its location to validate that it's real?
Or is the seller reputable looking?
olive sable
yes, but it is quite far
shipping shouldnt be an issue, but ill sjut be less lenient about the payment
nocturne olive
Hmmmmmmmm
olive sable
the seller hasd been on this website for 18 years, so they're decently reputable
nocturne olive
Yeah I think I can trust the seller this time at least
olive sable
oh actually, the highest bidder is at 500
nocturne olive
That I definitely can't afford
proven merlin
I was about to say that aint bad, but then it was "bad"
nocturne olive
You could probably sell the RAM for like 2K
proven merlin
Thats good for moe
proven merlin
You could probably sell the RAM for like 2K
Wouldn't be out of the ordinary
olive sable
- 2 x Processor Xeon E5-2683 V4 (FCLGA2011-3)
- Total 32 cores / 64 threads
- Memory 128GB DDR4 ECC RDIMM 2133Mhz
- ASUS GeForce GTX 1080 8GB (for the video outputs)
- 2 x Nvidia Tesla P40 24GB ECC (Total 48GB)
- Total 56GB NVRAM (slightly less by ECC of the P40s)
- Total 10240 CUDA Cores (for comparison Nvidia GeForce RTX 3090 there has 10496 but much less VRAM)
- Tailor-made cooling system for the Tesla cards, high air pressure fan with shrout, but less noise pollution than 40mm server fans. Under full load, the P40 cards remain neatly below 70°, the GTX at 88°.
- 250GB Sata SSD Windows 11
- 480GB Sata SSD Debian Linux
- Currently not dual boot, you need to physically swap the SSD.
- SilverStone SST-FS202B for two 2.5" SAS/SATA drives.
- LM Studio, Ollama etc.
- Power supply PSU 1450 Watts (ample sufficient)
That storage is so tiny
olive sable
the storage is upgradeable
proven merlin
For storage id have tons of semi bad hdds lmao
olive sable
id jsut pop a 16tb hdd in it and be good to go
proven merlin
"just pop a 16tb in it" this hurts..
olive sable
why?
proven merlin
The 16tb hdd is like
olive sable
250 euro
rough bloom
id jsut pop a 16tb hdd in it and be good to go
are there any 16 TB HDDs in the 2.5" form factor
nocturne olive
I kinda wonder
What's your HDD pricing like these days?
I have a lack of HDDs
amber fractal
Doubt it
olive sable
are there any 16 TB HDDs in the 2.5" form factor
probably not
the big ones are still close to new
olive sable
```
- 2 x Processor Xeon E5-2683 V4 (FCLGA2011-3)
- Total 32 cores / 64 threads
...
iggly
1.2K
not a bad deal
amber fractal
I have 1k in my account
azure lynx
adding transparent restarts for my agent. i shouldn't have to tell them I'm shutting them down coz it'll be transparent when it starts up coz they'll just reload their context.
olive sable
128GB DDR4 ECC RDIMM 2133Mhz
nocturne olive
Well anyway
I am uncertain if I would have the capability to afford the RTX Titan
It would be kinda nice if I could, not sure
proven merlin
I have 1k in my account <:enub:1163912230585237655>
5k here but at MOST 1k available to spend.. driving license
rough bloom
probably not
ye, the biggest one I can find quickly is 5 TB
so you're not getting to 16 TB with those 2 2.5" bays 
olive sable
ye that ram alone is already 1.2K
olive sable
ye, the biggest one I can find quickly is 5 TB
so you're not getting to 16 TB wi...
thats fine, ill put it sideways
or jsut put it somewhere inside
nocturne olive
Keep an eye on that RTX Titan I guess
I'll see if I suddenly get extra money from somewhere
rough bloom
or jsut put it somewhere inside
the inside of these tends to be quite full already 
but I guess worst case you could just... remove the drive bays 
proven merlin
Usb connected sata adapter: hehe
(For the love of god, do not)
Although.. a pcie slot is free.. you could do usbc..
olive sable
i love how this has a 8 bit display thats coupled ot he outside of the gpu and not the actual gpu temp
olive sable
they put the gtx1080 inbetween the p40s
rough bloom
big fan
amber fractal
it'll be fineeeeee 
rough bloom
this might be not horribly noisy
olive sable
any buyers in chat? 1.275 euro + triangle fee
rough bloom
#ad
olive sable
all jokes aside, not a bad deal, but i cant afford to
proven merlin
What can one do with 6,5 p106-100 btw
I have em, dont use em..
(essentially mining 1060 6gb)
olive sable
6gb of vram, not much i guess
opaque sigil
6.5?
opaque sigil
olive sable
so 7 x 6gb
proven merlin
Well
olive sable
42gb of vram
proven merlin
8x i do have a 1060 6gb
olive sable
48gb of vram
proven merlin
I also have a 1650.. and a 1050 (latter one im not selling/giving up/using for sure, first gpu, its my baby)
opaque sigil
that sounds horrifying but i guess if you already have the hardware
olive sable
you could use it for ai if you want. more vram than that 4080. just a lot slower
olive sable
azure lynx
is it faster than CPU?
proven merlin
1 token per 1-3 business days?
nocturne olive
48gb of vram
48GB of horribly inefficient to use VRAM
opaque sigil
depends on what cpu
azure lynx
if you have AVX512?
opaque sigil
then probably not if i had to guess
proven merlin
you could use it for ai if you want. more vram than that 4080. just a lot slower
Also, also, poor pcie bus lol
nocturne olive
1 token per 1-3 business days?
Nah it'll be a bit faster than that
It'll go at the speed of one of those GPUs
proven merlin
Nah it'll be a bit faster than that
It'll go at the speed of one of those GPUs
Thats not bad..
proven merlin
A 1050ti can do 30t/s in qwen3.5 2b...
proven merlin
Only one GPU can work at a time
Ah, so it has to wait for the last gpu to start the next token?
azure lynx
hmmm.. i have a 1650 sitting in a box
nocturne olive
Ah, so it has to wait for the last gpu to start the next token?
Yeah it has to wait for the final GPU in the chain to generate the token before it can begin predicting the next one
azure lynx
that could do my TTS
proven merlin
Yeah it has to wait for the final GPU in the chain to generate the token before ...
That would mean.. 8x slowdown
nocturne olive
No it would mean 1x speed
No speedup
You can load a much bigger model, but it won't go any faster than if it was run on a single one of those GPUs
proven merlin
Why so? If every gpu needs like, idk, 20ms for their own step?
azure lynx
it's the communicating intermediate steps between the GPUs i think
proven merlin
You can load a much bigger model, but it won't go any faster than if it was run ...
Thats fine if they are as fast as a model that fits into a singular vram?
proven merlin
Why so? If every gpu needs like, idk, 20ms for their own step?
Wouldn't it be 160ms per token then, slowing it 8x?
azure lynx
if you keep the model on one GPU it'd be fast.
nocturne olive
Why so? If every gpu needs like, idk, 20ms for their own step?
If you load a big model, it'll be spread out so that each GPU has different layers
Each layer still operates at the same speed, so it's essentially as if the GPUs were a single GPU with the core and bandwidth of a single one but the VRAM capacity of all of them
amber fractal
You only pay in latency between GPUs
proven merlin
If you load a big model, it'll be spread out so that each GPU has different laye...
Yes but since you have to wait for each layer to finish before you can do the next layer, and have to wait till the last layer is done before feeding the first layer again...
azure lynx
the trick is to use multiple models and spread them across the GPUs in such a way as you tend to be using all of the GPUs efficiently
nocturne olive
Yes but since you have to wait for each layer to finish before you can do the ne...
That's how inference works in a single GPU operation too
Just that it happens on a single device
proven merlin
Yes but since you have to wait for each layer to finish before you can do the ne...
You pay the calculation time between the first and last gpu, no?
amber fractal
the trick is to use multiple models and spread them across the GPUs in such a wa...
This is true, however most single models only give you MoE which sucks
proven merlin
Just that it happens on a single device
Yeah, but since you parse only a few layers at a time, it would still be slower than a model that fits into just one gpu, no?
nocturne olive
RTX 6000 quadro for 750
Definitely out of my budget unfortunately
proven merlin
Yeah, but since you parse only a few layers at a time, it would still be slower ...
U cant streamline/pipeline it, no?
Aka feed the first gpu before the last gpu has finished the current token
nocturne olive
There's no pipelining in inference though, you need to compute the previous layer before you can start the next one
rough bloom
Yeah it has to wait for the final GPU in the chain to generate the token before ...
this is only true for batch size 1 btw
if you have a batch of multiple requests you can parallelize
azure lynx
I'm currently cycling between inference tasks, checking if each one is running and if so stepping it and going to the next one. two GPUs would make that code much easier one dedicated GPU to the main task and everything else on the 2nd
nocturne olive
You can't have multiple tokens in-flight
olive sable
highest bidder is at 510
(not the ada variant) its the other one
nocturne olive
this is only true for batch size 1 btw <:nub:1163912091569225779>
if you have a...
That is a true exception
proven merlin
There's no pipelining in inference though, you need to compute the previous laye...
Yea so each of the computing time of the previous token per gpu gets added.. its a slowdown
nocturne olive
(not the ada variant) its the other one
Turing?
olive sable
they didnt mention
amber fractal
lovely, this patcher loss is so low that I need to be able to see more decimal places on the loss to make sure it isn't walling.
olive sable
just "not ada"
nocturne olive
TU102
opaque sigil
Yea so each of the computing time of the previous token per gpu gets added.. its...
it's a slowdown yes, because each gpu ends up sending their output to the next gpu, which you don't need to do on a single gpu 
nocturne olive
Yea so each of the computing time of the previous token per gpu gets added.. its...
I don't get what you're saying
opaque sigil
there are ways of working around it to an extent by doing batching and async shenanigans so the hardware doesn't stall and do nothing
nocturne olive
it's a slowdown yes, because each gpu ends up sending their output to the next g...
That's latency, idle time instead of compute time
opaque sigil
same end result, it's slower
azure lynx
the slowdown part is the inter-device communication overhead as well as the difference in speed of the device
proven merlin
I don't get what you're saying
Ig my miscommunication was a bigger model has more layers which means more compute time, while you assumed running the same amount of layers on a single gpu?
While i meant less layers on a single gpu
nocturne olive
Ooooooo
Yes of course a bigger model means slower
azure lynx
also it's kind of obvious that if you put 1/2 the layers on one GPU and 1/2 on the other it would be slower than putting them all on the faster GPU.
proven merlin
Ooooooo
Yes of course a bigger model means slower
Yea.. would be cool if you could start feeding the first gpu before the last gpu finished but thats not possible
nocturne olive
With a bigger model you need to transfer more data from VRAM
nocturne olive
Yea.. would be cool if you could start feeding the first gpu before the last gpu...
Yeah unfortunately not because the tokens are a sequence
The previous token is needed to predict the next token
opaque sigil
can the bubble pop so i don't have to do ML stuff
rough bloom
not yet
nocturne olive
The bubble can wait a little longer so I can have a bit more money to work with to then grab some cheap GPU
proven merlin
Yea.. would be cool if you could start feeding the first gpu before the last gpu...
Ig multiple prompts works for well many people but yea not exactly what im going for.. not many people will want to run an 48gb ai at 5t/s (if im lucky) soo.. not worth investing into a mainboard for.. i might get my hands on the mining rig theyre out tho, idk
azure lynx
i'm pretty sure the AI coding genie won't go back into the lamp though.
olive sable
tesla m40 24gb for 200
i feel like the m40 isnt worth it tho
amber fractal
lovely, this patcher loss is so low that I need to be able to see more decimal p...
me staring at the loss terminal output as loss is starting to go under 5 decimal places
proven merlin
If you think about it, claude code and such are just slotmachines..
azure lynx
me staring at the loss terminal output as loss is starting to go under 5 decimal...
that's what %g is for. scientific notation.
(or %e)
proven merlin
%.20f my beloved
amber fractal
I'm just at %.10f atm
azure lynx
├───────┼───────────────────────────────────────────────┤
│g │ Prints a signed value in either f or e form, │
│ │ based on the given value and precision---an │
│ │ exponent less than -4 or greater than the │
│ │ precision selects the e form. Trailing zeros │
│ │ and the decimal point are printed only if │
│ │ necessary; takes a double. │
├───────┼───────────────────────────────────────────────┤
doesn't that sound more sensible?
it's nice knowing if it's 1e-9 or 1e-10 at a glance i think.
proven merlin
Fair, but math
opaque sigil
funni numbers go brrr
proven merlin
Exactly
opaque sigil
log scales are boring
i don't wanna use them but then you can't see anything 
(suffering from success)
azure lynx
0.00000000000000000000 isn't the number you thnk it is though
proven merlin
Whats life for if not seeing a ungodly amounts of 0's with every print
Im making too many typos.. i should sleep
amber fractal
I'd consider it once I'm that low
azure lynx
ungodly numbers of digits should only be used for exploits in CTFs
proven merlin
"print pi, the right values will be in there eventually, the grep will catch that"
opaque sigil
tbh you should just print in binary
olive sable
@nocturne olive
if adding a 2nd gpu is possibel, what stops you from getting a shitton of 8gb cards?
azure lynx
"print pi, the right values will be in there eventually, the grep will catch tha...
there is a similar method of finding primes using a regex.
opaque sigil
sam have you considered there's this thing called physical space