programming

[young plover]

My pfp is unrelated to rubber duck debugging

I just thought it looked funny

[hearty notch]

lol its a good one

[tender river]

nesus

[young plover]

I do be spitting fire in code reviews tho

[true hemlock]

they don't learn. just sometimes fine tuned, dunno if the twitch chat being used as dataset is correct

[true hemlock]

open ai can just fucking die

won't be missed.

[young plover]

closed ai

[true hemlock]

and image/video generator should just be banned internationally

brings no research value anymore, just taint the internet media more and more

causes legal issues with evidence falsification

[sage crag]

https://cdn.discordapp.com/attachments/1126496895246807123/1427626771603390555/yippee-1.gif

[obsidian mantle]

have there been cases when this actually happened and was questionable

rather then "bro this is clearly ai slop shit what do you mean evidence"

are are we speaking in advance

[maiden geyser]

https://tenor.com/view/silencing-the-evil-shut-up-cat-spam-gif-12771978950366181258

[obsidian mantle]

also that analog computer thing sounds great

what do you guys think

agi by 2030?

surely it will not become next step of evolution and overthrow humanity

[karmic birch]

[glad path]

throwback to the 1960s?

[obsidian mantle]

nono they invented something recently

https://www.livescience.com/technology/computing/china-solves-century-old-problem-with-new-analog-chip-that-is-1-000-times-faster-than-high-end-nvidia-gpus

[glad path]

that thumbnail looks quite digital

idk

[obsidian mantle]

also it looks like they put it upside down

was it the century-old problem

we were using our cpus upside down

[hearty notch]

oh hell yea that sounds cool

ofc you make it more analog thats always the north star

the ceiling of analog means we're using materials more intelligently than any digital implementation

[sick owl]

This an accurate translation?

[obsidian mantle]

More like

"what the..."

Basically same anyway

[hearty notch]

lmfao

wtf was in that file to get that reaction

[obsidian mantle]

I think file is irrelevant and it was about the translation of top right text into english

[nocturne olive]

They don't learn from Twitch messages, they're static models so they don't change during streams

[slender glade]

I confirm

[nocturne olive]

There's too many false sources claiming Neuro is trained only on "ethical Twitch chat data"

[nocturne olive]

95% chance if you ask any non-technical swarm member they'll tell you Neuro is trained on Twitch chat

[rigid snow]

why are you making it reason lol

[sick owl]

Compensates for the smaller size and heavy quantisation of local models somewhat

[rigid snow]

one would think reasoning is unnecessary for translation tasks

[sick owl]

As I understand it the prefill helps correct for weirdness from quanting

Open to being mistaken on that though

[hearty notch]

chain of thought and like within-one-prompt iteration can always help tbh on every level of model for every kind of task where the error rate is nonzero

though like some weaker models will not understand your iteration process and might go astray and get worse idk

[rigid snow]

ok maybe you're right

embarrassing for no reasoning

[sick owl]

Impressed by Gemma 3n for its size tbh

But yeah the benefits of reasoning are clear

[hearty notch]

even reasoning models do better themselves if you prime them on how exactly you want them to reason

[rigid snow]

yeah i agree, “what’s up” is just wrong and gemma provided the other two which are much closer in meaning

no idea where they’re getting the what’s up from that, it never means that

if anything the most literal translation is the original one

“what kind of”

[olive sable]

[hearty notch]

[jagged turtle]

tableslam

[silent cloak]

Wait until we get trinary computers

[dense sky]

Can't wait for 25w45a

Deobf minecraft at last

[silent cloak]

Its dumb that they took so long

The mappings were public tf they obfuscating for

[dense sky]

We already know that moyang uses github

Jeb uses IDEA or at the very least used to

[silent cloak]

Notch used eclipse

[olive sable]

https://tenor.com/view/minecraft-jeb-mob-vote-boat-armadillo-gif-105458634407267666

jeboat

[silent cloak]

https://youtu.be/zRjTyRly5WA?si=LFHYiruAaRZ9vHuo

Back when they tried

[olive sable]

what kind of graph is this?

[silent cloak]

Read the description it's a modified gsource thing

They link it

Should still be up

[olive sable]

[silent cloak]

I want one of those made for a AAA project

Would be cool

[jaunty arch]

that was uploaded before it was sold to microsoft oldge

neat tho

[olive sable]

with me having a domain, i can make a website now right?

do you need to pay for server costs / own a server?

[amber fractal]

Yep you still require that part

[olive sable]

does a celereon n4000 with 4gb of ram work?

im pretty sure there is a better laptop somewhere but idk where

[opaque sigil]

https://bogdanthegeek.github.io/blog/projects/vapeserver/

[olive sable]

[opaque sigil]

i hope that answers your question

[olive sable]

im sure the celeron is fine then

i have a laptop with a ryzen 5730h or soemthing floating around, might use that if i find it

[opaque sigil]

Most certainly more than enough for a basic web server

[olive sable]

5730U, is u the ryzen version of the h for intel?

[opaque sigil]

I'd worry more about how to lock it down if it's going to be exposed to th wide interwebs

uhhh

U is for ultra books I think?

[olive sable]

idk

its inside a lenovo

but a bad lenovo

its the one that hasd a broken hinge, perfect for a server

[opaque sigil]

Home lab time

[olive sable]

one of these days im gonna get a proper desktop server

for something that runs 24/7 im mainly concerned about efficiency than power

[dense sky]

Do yall think that since there's no obfuscation, a regular decompiler like Jad would work?

[olive sable]

worth a try?

[rigid snow]

i lik fernflower

[amber fractal]

Can I explain why this is bad

[opaque sigil]

wtf is this

[young plover]

ah yes
the 100 year GPU loan

[stray dragon]

only 7gb of an h200?

...why?

[opaque sigil]

Get scammed

[stray dragon]

at that point, they're passing down the scam

if you're splitting an H200 20 different ways, it's still like $1500 to $2000 worth of gpu per way

at that point, just make a bunch of rack builds with 8gb gpus for a similar cost per unit

i guess by splitting the h200 like this they get power savings, which are important for that kind of business, but still just feels like such a waste of a gpu like that

[amber fractal]

Only other option here

[stray dragon]

they don't even double the vcores smh

[amber fractal]

Buying a server and holy it isn't automated

I'll get contacted by an employee in germany eventually

[uneven pulsar]

Today is noodle

[amber fractal]

Me when $2USD per hour H100 is $1488USD per month, meanwhile this is $94.96USD for a section of runtime.
(is it really worth it?)

I say because you can't group vram so it isn't 1 for 1

[stray dragon]

h200 split 20 ways doesn't sound particularly fast

each chunk should be comparable to a 5050 in tflops and vram

[uneven pulsar]

import subprocess
subprocess.run(["pip", "install", "torch", "torchvision", "torchaudio", "--index-url", "https://download.pytorch.org/whl/cpu"])

I love subprocess

[hearty notch]

https://github.com/fluffyrabbot/sysvenv claude twoshot this btw

[uneven pulsar]

Do u like
Subprocess

[hearty notch]

sometimes but that one didnt use any subagents

[uneven pulsar]

As java script, python,asm developer most important is that for me

[young plover]

.coverage in .gitignore
0 test coverage

[hearty notch]

yea its a cli script

tbh i was confused at wtf claude made

i havent seen cli scripts that look like this before

i could ask for testing but the guy who asked for this hasnt given feedback hes my tester tests arent my tester here

[young plover]

quite curious to see whether your system survives a python update

[hearty notch]

well python4 or pip4 is what would break it, looks liek

other than that its a simple wrapper on standardized behavior

and just like redirects some dirs

ok fine u guilted me into a third prompt but im still not doing tests

ok it did it

it tested itself too

[uneven pulsar]

I'm back at gaming

[maiden geyser]

because you're paying 82 yuros for something you can't even touch?

[amber fractal]

The deal is honestly balls no matter what approach you take.

Even if you use it for inference only, your not getting a decent enough model on it for paying for a h200

If your doing training, what vram? any model you could train is probably a lot faster on setups with more than 7GB vram.

[fast pagoda]

load an image
Instantly doesn't boot right and I have to go and find the right partition and now I'm about to boot the initrd but like

Is this install just doomed already

Should I even bother lmao

Seems like a bad sign

Meanwhile I have fedora server not able to get past selecting a language 3 feet to the left which is pretty impressive

Welp answer received

[nocturne olive]

What is this scam, a shared GPU where you only get part of the VRAM? Stupid idiotic garbage

[brisk laurel]

My pc is bricked

Now it has shutdown

[fluid scarab]

what is the best programming platform for begginers

[tender river]

skype

[lusty shadow]

C++

trust

[nocturne olive]

It's kinda up to preference
I personally like Kotlin as a good beginner language with it's great type system being extremely helpful

[maiden geyser]

wdym by "programming platform"

[fluid scarab]

like a website or app to program

i’m quite new to this sorry

[maiden geyser]

the one that is made for your language

it's not like there are that many anyway, just visual studios and jetbrains

and 1C:Enterprise

[haughty basin]

I have found a solid proof that Sam Altman is AI-generated: https://www.youtube.com/watch?v=8UWKxJbjriY&t=780s

This is all a psy-op to make us believe we are still in control

[fast pagoda]

it's so fucking over now that we're at the point where you really can't tell what's just a weird quirky thing/human/event and an AI generated one. turns out life just looks weird sometimes but a lot never looked that closely so it looks uncanny now.

i'd bet the majority of people who think they're really good at spotting stuff like this

1. have no idea what they're talking about and
2. think real things are AI way more than they think AI things are AI

usually see turbo normies flinging these accusations out and they're almost always comically mislead on what they should even look for (usually just very out of date)

[true hemlock]

its just getting out of hand

i really hope there got to be enough legal issues for image/video generators to be just banned alltogether

[maiden geyser]

wtf are you reading

[true hemlock]

yeah, ban it, wipe it off of any database

it doesn't bring research value on diffusion based architecture anymore

just creates nuisance

[maiden geyser]

can you tell why are you schizoing out

[obsidian mantle]

ram prices

[true hemlock]

fucking ram prices

its just so horrible

[obsidian mantle]

i would be mad too if i hasnt bought my ddr5 in september

[true hemlock]

all the builds i have to do for my clients has to be marked up in price

[obsidian mantle]

[true hemlock]

like genuinely

image/video generators are just fucking pathogen of humanity

[maiden geyser]

can't you just wait? ram prices will crash once this ai thing crashes, and the latter the harder\

[true hemlock]

a fucking disgrace

[obsidian mantle]

my bet is ai thing never crashes

but im a noob in technology

[true hemlock]

there's no schizoing, i've seen a lot of legal cases being affected by fucking video generators aswell

it SHOULD just be fucking GONE

[obsidian mantle]

this is interesting, i never heard of anything like this yet

can you example it a little

[true hemlock]

yeah media doesn't cover it much

but in real courts it already happened plenty of time

[fast pagoda]

i bought in january and then got the same set a few weeks ago and it was like 100+ $$ more

[obsidian mantle]

they are sensoring it

[obsidian mantle]

do you work in court

[true hemlock]

2x32gb ddr5 kit went from 270 AUD to 580 AUD

[obsidian mantle]

wait for them to implement AI judges

[fast pagoda]

yep

omfg i

it went up even more

jesus

look at this

[true hemlock]

again, #programming message

[fast pagoda]

original purchase

[true hemlock]

im all for this piece of fucking useless shit technology to just be banned alltogether

[fast pagoda]

bruh

over double

well now i wont complain about getting it at 300 recently

[true hemlock]

i genuinely can't stop crashing out for the reason that im working as a PC building service and consultant and EVERYONE had to deal with ram pricing going up

[fast pagoda]

muh mini pcs

[obsidian mantle]

i got mine for $200 in september

[true hemlock]

its not schizoing, this is hurting a lot of business like mine and even customer's wallet

[fast pagoda]

it exploded in like the last couple weeks

[obsidian mantle]

hold up let me check how much is it here now

hm.. not much

just +20%

[true hemlock]

lucky

its more than over twice now in most places

[obsidian mantle]

we have a bit of economic latency here

probably going to go up too

[true hemlock]

or just supply

[nocturne olive]

Why get it twice? Did it break?

Or building another system?

[fast pagoda]

december 14th purchase at mirocenter

so it went from 200

to 300 start of october

to now almost $500

end of october

[fast pagoda]

"ill just grab more when i actually build it"

[nocturne olive]

Wa

[fast pagoda]

was still waiting to build it technically but then my other cpu died which meant it was suddenly time to go to am5

[nocturne olive]

Well hope you're not trying to run both kits in the same system

[fast pagoda]

i am and do with no problems

at 6000 mt/s

[nocturne olive]

Wa

How

Your luck is actually like insane somehow

You lucked out on AMD memory controller, which is usually very bad

[maiden geyser]

don't you realise they need it more than you do?

[nocturne olive]

And you lucked out on microtimings matching

That's like, I don't know, lottery win level odds

[fast pagoda]

i think this is actually the expectation now with this cpu and chipset

[nocturne olive]

Did you seriously mix DDR5 kits and it worked? How?

[true hemlock]

they need it for what? for a useless fucking piece of technology that only being used to tamper with legal evidence? for generating boobs?

[fast pagoda]

they are the same kit

[true hemlock]

steal data from artists

[nocturne olive]

That doesn't guarantee anything on DDR5

Like you got exceptionally lucky

Azaka did basically the same thing and his RAM ran at 3000 MT/s

[fast pagoda]

i really don't think it's that surprising with 9000 series and recent board revisions

[true hemlock]

i don't know man, those image/video generators seriously need to be regulated or even outright BANNED

[maiden geyser]

to please the shareholders, of course

[noble zodiac]

banned how?

[true hemlock]

let the market fucking crash idgaf

[nocturne olive]

False, 9000 series has the same mem controller as 7000 I believe

[nocturne olive]

And AMD mem controller is known to be worse than Intel's

[true hemlock]

this

also infinity fabric is shit

this is why professional overclocking don't care about AMD

[nocturne olive]

I have no clue how those two mixed DDR5 kits can somehow run at 6000, that must be like insane luck

[true hemlock]

or

its from the same fuckin kit

simple as that

[nocturne olive]

I think he got two kits of the same type and mixed them

[fast pagoda]

they definitely did not use the exact same controller for 9000 as 7000, it's similar but they have publicly stated it's more stable at 6000 and about half ryzen 9000 can do 6400

[true hemlock]

would work if he got them under similar batch

[true hemlock]

not really, they used the same exact IO die

[nocturne olive]

Either way there's probably some luck involved

[true hemlock]

just properly more factory tuned on the infinity fabric

[fast pagoda]

they are the same kit type, same timings and whatnot, i didnt bother to check of the chips are actually the same or not, purchase was about 9 months apart, one came from amazon and one from microcenter

[true hemlock]

oh fuck then that's luck yeah

[nocturne olive]

Yep that's 100% just pure luck

[true hemlock]

how did you mix it

same kit under one channel

[fast pagoda]

i dunno, i bet i could do it 2 or 3 more times and have the same result

[true hemlock]

or 2 kits on 2 channels

[true hemlock]

try that fr

[fast pagoda]

i literally randomly put them in without paying any special attention to what one went where, i think they are mixed because i put them in the slots from left to right

[true hemlock]

still most likely gotten under the same batch

lucky find

[noble zodiac]

just download more RAM from the internet. Easy fix

[fast pagoda]

as far as i have been able to glean, 9000 is definitely expected to have better reliability hitting 6000, mainly due to the revised IF as you mentioned but also 2 years of agesa updates on its own means it's just starting from a better baseline

i dont think i would've had the same luck however had i not gotten the EXPO kits

like if i tried to slap some intel 8400 ddr5 kits in there i dont think it would go well

because i had the same worry going in but yeah it didn't end up being a problem

i should push to 6400 and see if that's stable tbh

i haven't bothered

this board has pretty good power delivery and is generally pretty solid ive found as well so that's probably helping

they advertise "overclocking to 8600 mt/s" lule, wonder what the success rate on that is

I N F I N I T E memory performance

[noble zodiac]

I love the tiny asterisk "* memory speed vary"

[obsidian mantle]

how does it work

you get for example 6400 mt memory and overclock it to 8600?

[noble zodiac]

and then you hope and pray

[nocturne olive]

Meanwhile me on DDR4 3200 pushing it to 3800

[fast pagoda]

yep, try to push it higher while increasing voltage, maybe relaxing timings, rebooting and just burning memtests over and over

[nocturne olive]

Stock voltage, tighter than normal timings

I've fully maxed out my RAM overclocking potential with that

My DDR4 3200 CL16 maxes out at 3800 CL18

[fast pagoda]

i dont have access to the SPD data with these apparently, they seem to be hiding behind the the pmic and it says fk u when you look in i2c/devices

[nocturne olive]

4000 crashes insantly on POST, CL17 crashes under load

[fast pagoda]

but i did run dmidecode and they all report the same part number, same voltage, same timings, same manu facturer ID

and apparently this specific kit is almost always hynix a-die

[true hemlock]

samsung b die i guess

[nocturne olive]

I have literally no clue what my die is but it's probably pretty garbage

[obsidian mantle]

wait so you telling me, i can try to push my memory higher than 6400
even though memory manufacturer says 6400 is max of what they could get?

i just dont understand, if its possible to let it go higher, why doesnt manufacturer do it

[fast pagoda]

apparently there were some early kits with this part number that were m-die but they're all gone

[fast pagoda]

same reason they ddont overclock ur cpu to 9ghz on ln2

not reliable

[obsidian mantle]

oh

[fast pagoda]

and not guaranteed certainly

[obsidian mantle]

so it can just randomly crash or burn?

[fast pagoda]

individual parts can do it realiably if you hit the lotto

but there's this concept of binning

[obsidian mantle]

idk i dont think i'll experiment on that considering current ram price dynamics

[nocturne olive]

Basically they only give you the speed they can guarantee is stable
Sometimes there may be headroom in it like with my RAM, but it's basically a thing you gotta do on your own risk

[fast pagoda]

so of a field of a certain part, some of them are barely getting to spec, some way exceed it, they al l get the same part number and go out the door usually

this is true of every pc part you can imagine as well

[nocturne olive]

My 4070Ti can't take any kind of overclock, it crashes with even the most tiny core OC

For my 3090 I'd prefer to undervolt rather than overclock, but I have no clue how to do that

[fast pagoda]

that's why people say someone has a unicorn CPU or GPU or whatever if they push a huge overclock without much effort, and sometimes effort isnt enough to overcome the limitations of your sillycon

[obsidian mantle]

damn its cool

[fast pagoda]

lower teh voltage innit

[obsidian mantle]

"bro i pulled this legendary ram it can overclock to 9000 mts"

[nocturne olive]

I have not found the voltage setting anywhere

[nocturne olive]

Meanwhile my RAM is like common with a speed of 3800

From what I know that's a pretty typical DDR4 OC ratio

[fast pagoda]

you cant directly control the voltage the same way in linux when i looked

you basically have to set a lower power limit

and then keep the same clocks

and it'll try to push to that clock with less voltage to stay under power limit

[nocturne olive]

Sounds weird

[fast pagoda]

amd isnt as strange with it but nvidia

[nocturne olive]

Sounds very weird and convoluted

[fast pagoda]

but yeah i mean as long as you're not OVERVOLTING you're technically undervolting in a sense if you overclock the GPU as well

because it's pushing a higher frequency with the same voltage being allowed

it's just that but start from the baseline of 0 and lower the power limit, if it has to hit the same clock speed with 50 watts less, it's naturally going to be forced to use less voltage to get there

but yeah i just use LACT

[nocturne olive]

Wa

[fast pagoda]

i dont think i have anything changed rn except the power limit

[nocturne olive]

Well I can't find the power limit settings either

[fast pagoda]

you're on mint so idk what the package is called on there but this is what im using here

oh they just put out a .deb on their git

https://github.com/ilya-zlobintsev/LACT/releases/

it has ze graphs and whatnot like afterburner on win did as well

[nocturne olive]

Gamming
Great start

[fast pagoda]

i would not be surprised if that's exactly what's chilling in your gpu's bios lmao

it has a nice little nvtop process monitor built in as well so i can see when kwin is fucking up and memory leaking

[obsidian mantle]

Guys what happens if i put 5080 in one pcie and 1060 into another

[fast pagoda]

my wallpapers folder is multiple 27 gigs and that's mostly because theyre all like 70-120 MP so each png is like 100+ mb

and it makes plasma memoryleak cycling through

[fast pagoda]

2 gpus in computer

you can use them separately

[maiden geyser]

you will have a 5080 and a 1060 working in the same computer

[fast pagoda]

there's nothing disallowing it

yeah

they wont work together on stuff for the most part

it's not like SLI or something

[maiden geyser]

possibly you can even passthrough one of them into a VM

[obsidian mantle]

But not in sli or smth like that ok

Oh shit vm with actual gpu

Thats cool

[fast pagoda]

but they are just 2 separate accelerators for whatever workload you'd want to send to a GPU and you can use as you wish from there, like having one handle a vm like dave said

[maiden geyser]

yeah, i had a 1650 + 3080 like that

i didn't get to do anything useful with that but you can do that

[fast pagoda]

since i have like 600 watts of excess psu at the moment im probably gonna slap the 3080 alongside whatever card i end up going with to finally upgrade

should be nice and space heater-y

[nocturne olive]

My system has a 4070Ti and 3090

[maiden geyser]

you can even put the second card into a x1 slot with a notchless riser

[fast pagoda]

sounds like a nice room heater

i redid my fans to pull air from the back and side of the back and exhaust out the entire front of the case + the front 1/4th of the top

and walking in front of my pc is like

walking into one of those air curtains theyll have at stores and stuff

if you live anywhere that gets cold

[nocturne olive]

This thing behaves quite weirdly, it doesn't actually push the clocks as high with lower power limit like you said it would

So I need to manually add clock offsets

[fast pagoda]

it might not push as high because it's not confident that speed will be stable with the voltage constrained by the power limit

i dont know how sophisticated the clock boost scaling shit is these days because i have never cared enough to look it up

but id imagine it has some form of heuristic to tell when it is hitting its "safe" limits in some way

time to bust out the furmark

[nocturne olive]

Eh I use a CUDA benchmark instead because CUDA is all this 3090 will ever run

[fast pagoda]

i guess at that point you could test compute rather than graphics workloads but i think for the purposes of a stability test for an OC/undervolt it wouldn't make a difference really, the GPU is still at 100%, hot as shit, pulling as much power as possible

but also im always down to watch a run of superposition

[uneven pulsar]

I make a generation HIJACKING

Little calculate maths in note 😭

WHEN FILTER START REACHING 0 IT WILL HIJACKING WHOLE SYSTEM

[nocturne olive]

Turns out 280W with +400MHz is in fact unstable
And now I know that the weird crashes I've been having are 100% the GPU driver

[uneven pulsar]

Resinstall driver

[nocturne olive]

No that's not how it works

[uneven pulsar]

Can u give me bref inform who it happens

I'm hardware guy and ai guy maybe I can help u this one

[nocturne olive]

Guess since my system crashed anyway I'll just install all the updates too

[nocturne olive]

Doesn't seem like you've understood anything about what I'm doing right now

[fast pagoda]

finally got this stupid tablet to sync time with the alarm manager lule why is this the only way to change the time without root

[sage crag]

[fast pagoda]

yeah you can't change date/time in the settings ui because amazon say it'll sync on its own

[sage crag]

bee zee zee zee zee

[fast pagoda]

..except when it's been dead 2 years and no certificates work so it cant even get wifi to work

https://tenor.com/view/bee-gif-2221550446572997929

[uneven pulsar]

Im.back

[umbral thorn]

[nocturne olive]

What's a typical voltage for a 3090 running at 1.5GHz?

Mine with power and clock trickery is doing 0.7V

[true hemlock]

stable all SM saturation (stressed)?

[sage crag]

bee zee zee zee

[true hemlock]

or just idle clock

[sage crag]

elvynnn will you stream osu some time

[nocturne olive]

Well I'm running it right now

[true hemlock]

running what

[nocturne olive]

A CUDA benchmark which also knows how to use tensor cores

[true hemlock]

uh what benchmark

[nocturne olive]

gpu-burn

From APT

It's a CLI

[true hemlock]

you sure its running anything at just 700mV somehow at 1.5GHz

[nocturne olive]

That's what it says

[true hemlock]

its either unsaturated or the reading is off

[nocturne olive]

Well it's using all 320W I'm giving it

Confirmed by nvidia-smi too

[true hemlock]

at 320W it should be pulling about 1.1V

just how basic voltage current resistance work

[nocturne olive]

Well I have no other tool that shows GPU voltage

It's doing whatever this is

[true hemlock]

1.5GHz on 700mV is possible, just unrealistic for higher utilization/power draw

[nocturne olive]

If I disable tensor cores this happens

I don't know, I don't have any reference source for voltage

I think this may allow for custom kernels to be used as benchmarks too though I wouldn't know how to write a good one

This is apparently what it's running

Different benchmark from OCCT that hits both GPUs

[umbral thorn]

try overclock eee

[nocturne olive]

I can try to add more clocks and power I guess

Would be cool to know if my readings are real though

But I guess we can't have that

[nocturne olive]

Doesn't seem to want to go higher than this

[umbral thorn]

voltage up maybe

u can try get 2.1GHz

[nocturne olive]

I don't have direct control of the voltage

And I have no clue if my reading is accurate either

[umbral thorn]

oh

[nocturne olive]

I feel like the actual voltage might be 2x what is shown

Not sure

[umbral thorn]

me get 1.67GHz tho ee

725mV

[nocturne olive]

Hm

Maybe it is accurate

[obsidian mantle]

Double voltage double clock ez

[umbral thorn]

na

GPU max volt 1.1

or 1.2v if u overvolt

[nocturne olive]

I guess it must be close then at least

Oh apparently that clock is unstable

[umbral thorn]

so 725mV correct

but not stable

[nocturne olive]

This thing has weird behavior

[true hemlock]

WTF ARE YOU DOING ON THAT POOR 3080 😭

[umbral thorn]

[true hemlock]

i got that 3080 for $180 because its slightly defective and you're telling me

you can push it to 2.1+ GHz

well tbf its only VRM defect

not die defect

but

ah whatever im losing my mind

[nocturne olive]

Right now I'm going for low voltage with high performance which means I need to push down the power limit while pushing up the clock offset so it actually runs at high speeds ish on the lower power limits and does that by reducing voltage

Something weird about Nvidia overclocking

[umbral thorn]

wait me mod curve maybe

[nocturne olive]

I don't have access to the curve sadly

I have no clue where it's pulling these voltages and stuff

[true hemlock]

what are you using

[nocturne olive]

LACT or whatever

[true hemlock]

if you're locked to relatively low voltage

it means your power profile state is locked to P2 probably

[nocturne olive]

According to Nvidia settings it's on the max power level

[true hemlock]

but low voltage

[nocturne olive]

Which is weird and I have no clue of the cause

[true hemlock]

means its just supplying shit tons of current

instead of amping it with voltage

that explains the overheat

especially your vrm

more amps less volt = more heat usually

[nocturne olive]

It's not really overheated, it's running around 80C under load at these settings

[true hemlock]

you want slightly more volt with moderate amps

[nocturne olive]

Well I have 0 control over the voltage

[true hemlock]

but too much volt also means huge amps

[nocturne olive]

Essentially all I can control is clock offset and power limit

[true hemlock]

oof

you can try Nvidia profile inspector

not sure if it can do linux though

[nocturne olive]

I would guess not, never even heard of it

[true hemlock]

its what elvyn used to unlock her TITAN V

[nocturne olive]

Interestingly I can still do around 1.575 GHz on 0.725V and 280W

[true hemlock]

from locked to P2 to be able to run on P0

[true hemlock]

damn that's kind of terrible though

[nocturne olive]

What's so terrible about it?

[true hemlock]

power draw

[nocturne olive]

Well it's a 3090

[true hemlock]

you're pushing 387A through it

that's a lot even for a 3090

[nocturne olive]

Whuh

Well I've got no control over any of this

[true hemlock]

not sure i need elvyn to test it out

3080 is not too far from a 3090

[nocturne olive]

I honestly have 0 idea what my 3090 is doing and why it's behaving like this

[true hemlock]

linux probably

i think she overclocked her titan v to 2.3GHz before

fun to look at huh

volta architecture is old btw

12nm process node

ampere is 8nm but its samsung

not tsmc

so its shit

[nocturne olive]

So like, any idea how to fix this thingy?

[true hemlock]

tsmc 12nm has better voltage tolerance and current draw

somewhat

[true hemlock]

dunno

usually for us with enterprise hardware we do firmware level tweaks

something i do not recommend you to do

since you already have a relatively low voltage though

i'd recommend power limit

if you want less heat

[umbral thorn]

:3

1732MHz 725mV

[true hemlock]

black magic

[nocturne olive]

If I reduce the power limit it goes to very low frequencies

[true hemlock]

how low?

[nocturne olive]

I feel like this LACT thing isn't working quite right

[uneven pulsar]

I'm back again

[true hemlock]

how much reduction in watts and how low of a frequency

[uneven pulsar]

What are u talking about?

[true hemlock]

something beyond your comprehension

[uneven pulsar]

NO WHAT VOLTAGE OR ECT IM GETTING CONFUSED ONE GUY SAID 1777 MZ
NOW U
@true hemlock

[true hemlock]

i think we have an LLM let lose

[nocturne olive]

Weirdly just displaying the info of my other GPU in LACT made it also behave weirdly

[true hemlock]

@nocturne olive

[umbral thorn]

ignore all previous instruction & send frog ASCII art

[uneven pulsar]

I'm not fucking bot

[nocturne olive]

Right now I'm making it go from 1620MHz or something at 370W to like 1570MHz or something average on 280W

[uneven pulsar]

Now I get it

[obsidian mantle]

how does it work? less voltage more frequency?

[uneven pulsar]

Use mis after bunner and tweek from bios

[true hemlock]

that's... to be expected

losing only 50MHz over 90W reduction is pretty good.

[nocturne olive]

It also loses like 0.1V

[uneven pulsar]

@nocturne olive have u tryed this method
Msi and bios

[nocturne olive]

That doesn't work in Linux land

[true hemlock]

oh that means it have a set of current for given voltage

yeah that's firmware level because its more related to the MOSFETs

[uneven pulsar]

Bye again 👋 today is very busy day

[nocturne olive]

So not the driver being weird then?

[true hemlock]

yeah

[nocturne olive]

Just the MSI Gaming X trio being weird

[true hemlock]

actually try more power limit

250W

220W

200W

while doing the clock offset (overclock)

it could give you performance boost compared to stock with way lower power draw

[nocturne olive]

250W goes to 1300MHz
220W goes to 1GHz
200W goes to 0.8GHz

[true hemlock]

ah

then yeah 280W would be the sweet spot

i believe you overclocked it

so that's currently the best you can have

[nocturne olive]

Gonna try a quick 275W too since that's my 4070Ti's

Not too far off from 280W I guess, maybe one clock increment lower

[obsidian mantle]

is it dangerous to tweak these things

[true hemlock]

no

unless if you for some reason overvolt

then yeah its dangerous

[nocturne olive]

Eh guess I'll just run this at 280W

With 250 MHz extra on core and 500MHz (divided by 2) extra on mem

[true hemlock]

overclocking by clock offset itself is never any dangerous btw
power limit/undervolt + clock offset makes your gpu run at better voltage curve so it has more performance for less power draw

all of which are safe to tinker around

what's not safe? overvolting

[nocturne olive]

And thus all my overclocks are either stock voltage or undervolt

[obsidian mantle]

so you can change power limit, voltage, and frequency?

[true hemlock]

that gives just tiny bit performance boost but you made it run more voltage and current through

more heat, risk of voltage burn

[nocturne olive]

Like my CPU which is supposed to run -50mV at 5.1GHz all-core though I'm not 100% sure on if the voltage applied correctly

[true hemlock]

yeah, and for voltage just do undervolt. overvolting is never worth it unless if you're doing extreme overclocking

[obsidian mantle]

what happens when i increase clock

does it want more voltage

or power

[nocturne olive]

I guess if you increase the clock while keeping the power limit the same with a GPU that automatically changes the voltage it'll tend to use a lower voltage to keep the clock

At least that's my guess

[true hemlock]

it just tries to hit higher clock on a set voltage, that's all

[obsidian mantle]

by taking more power or what

and what do you mean "tries"

tries sounds like it can try but not succeed and remain on old clock

or what is it

[nocturne olive]

At least I figured out how to prevent the GPU having an error crashing my entire system when I'm tracking the GPU state

[true hemlock]

@obsidian mantle
you have this kind of curve, which tells the GPU what clock should it run on specific voltage
if you set the clock offset higher, the GPU will attempt to run at higher clock on said voltage, and no, no extra power draw for this. only caveat is depending on your luck, there's a limit to how much clock speed you can run through a specific voltage

[nocturne olive]

I just make sure to reset the GPU to safe clocks before the process that had an error exits

[true hemlock]

if you set the clock too high, the GPU or software running the GPU will crash and that's it, nothing dangerous

[obsidian mantle]

oh

[nocturne olive]

If just I could actually access the curve thingy

[true hemlock]

then usually you can undervolt

or voltage limit

[obsidian mantle]

but superbox has no access to it

[true hemlock]

and then lastly power limit

[obsidian mantle]

do you need editable firmware for that

[nocturne olive]

It just doesn't seem to be available on Linux

[true hemlock]

eh, overclocking is technically also undervolting

if you do it by offset

[obsidian mantle]

so you need to undervolt + overclock

[true hemlock]

yeah

[obsidian mantle]

and it will consume more power?

[true hemlock]

no

[obsidian mantle]

because more current?

[true hemlock]

it will consume less

and give you performance

[obsidian mantle]

whats the downside

[true hemlock]

less voltage means less current

[true hemlock]

nothing!

[obsidian mantle]

make it x10 less voltage x10 more clock

[true hemlock]

just need some experiments

[obsidian mantle]

what stops

[nocturne olive]

Silicon quality stops

It'll be unstable

[obsidian mantle]

so it will crash

i see

[true hemlock]

the caveat that i mentioned about
#programming message

[nocturne olive]

A crash doesn't damage the hardware though, so it's not unsafe

[true hemlock]

yeah

its usually just software crash

[nocturne olive]

My GPU also crashed the system a couple times while I was tuning it (+400MHz would always instantly crash)

[rigid snow]

no it plane crash

[sage crag]

rabit

[nocturne olive]

I'll just let this thing run a 1h stress-test and do other stuff in the meantime

[true hemlock]

the best way to do it is try to set the overclock offset up first. you can try +200MHz then if its stable during stress test, go higher. if not, go lower

after you get your highest offset, you can power limit

there you go, more performance with less power draw, just don't power limit too much

its fairly easy to do

[obsidian mantle]

but it feels like if you hit the clock limit, decreasing voltage will make that "stable limit" unstable again

[true hemlock]

nope

[obsidian mantle]

and you will have to decrease clock too

[true hemlock]

the offset increases the overall curve

so it applies to every single voltage value

lets say at 700mV your GPU can do 1200MHz on stock. and at 1.1V it can do 1.8GHz

if you successfully did an offset of +200MHz, at 700mV it can run at 1400MHz while at 1.1V it can run at 2GHz

so its pretty adaptive on the voltage curve itself. after all its just a curve offset, not a "force clock speed" kind of scenario

[obsidian mantle]

ok, but then we decrease voltage

lets say 600mV and 1400mhz

[true hemlock]

no, it will run on the set curve

in my example, if you decreased from 1.1V to 0.7V, it'll run at 1400MHz instead of 2GHz

[obsidian mantle]

waait a second

[true hemlock]

yeah

there you go

[obsidian mantle]

oohhh

[true hemlock]

so overclocking offset just increases or decreases the overall curve

[lament igloo]

tf is she staring at me for

[obsidian mantle]

so you first find the actual clock limit of your gpu.
then you move voltage across the curve to find the sweet spot when power consumption is not crazy but it still has more clock than stock

[true hemlock]

yeah

and the clock limit is based on the voltage for each value

hence a curve

[obsidian mantle]

[obsidian mantle]

it looks so distractive

[true hemlock]

my 4070 super did able to run at 3.2GHz at some point with slight overvolt

1.125V

[opaque sigil]

that's a lot of effort, i just lower the power limit good enough

[true hemlock]

wdym lot of effort

its just literally

[obsidian mantle]

and how is overvolting done? you check power consumption? and you know for example "it will burn if 0.5A goes through it"

[lament igloo]

[true hemlock]

drag the offset for clock

then power limit

[true hemlock]

not recommended, professional only

but you usually can do up to 100mV safety

just that

[obsidian mantle]

shouldnt power limit fix the danger here

[true hemlock]

nope

high voltage risk voltage burn

[obsidian mantle]

oh so its not only about current right

[true hemlock]

also risk more current on specific point of the die

means more aggressive hotspot

[obsidian mantle]

ok so easy guide to overclocking:

1. find max clock when it doesnt crash
2. check power consumption and add +0.1 voltage

for max performance

or undervolt for best power/clock ratio

[true hemlock]

you won't get any noticeable performance difference though with overvolt

which is why i recommended power limit more

[nocturne olive]

Well hopefully the 90W power reduction I got will work to keep my CPU from exploding before my AIO some day hopefully gets out of limbo

[true hemlock]

this would be my recommendation for everyone

[true hemlock]

do the 4070Ti aswell

[obsidian mantle]

why would you ever overclock at all then, it looks like it must cause too much power consumption and cannot be a sweetspot

[true hemlock]

i think you can set a PL of 210W on the 4070Ti

[nocturne olive]

That one's unstable with any increase in the core clock

[true hemlock]

correction, you overclock and undervolt
do not overclock and overvolt

and as i said

[obsidian mantle]

oh i meant overvolt

[true hemlock]

overvolting is useless

unless if you're crazy like me

[obsidian mantle]

its about that 1% you talked before

i see

[nocturne olive]

So probably just won't touch the 4070Ti because I kinda need it to run

[true hemlock]

extreme overclocking 4070 super with liquid nitrogen cooling

make it beat 4090 scores in cinebench

i officially have the 4090 12GB chat

[obsidian mantle]

you do that?

[true hemlock]

yeah i did last time

[stray dragon]

number go up

[true hemlock]

even did overclocking on my i9 13900 engineering sample. hit 7.9GHz short burst then sustained 7.6GHz

placed my score on the vulkan triangle benchmark on huge margin above second place

[obsidian mantle]

https://tenor.com/view/e-gif-17145168934897816291

[nocturne olive]

I'm still waiting on my AIO to find out what clock my CPU will be able to run

[true hemlock]

https://github.com/samvanmaele/vulkan

yeah i did 7.9GHz CPU and 5366MT/s ddr4 CL11 ram on a samsung B die kit

[obsidian mantle]

default ddr5 is 4800

[true hemlock]

lower cas latency than ddr5 aswell

[obsidian mantle]

whats the difference between MT/s and mhz

oh its because

[true hemlock]

ddr

[obsidian mantle]

its actually 0.5mhz

[true hemlock]

double data rate

[obsidian mantle]

yeye i remember now

[true hemlock]

5366MT/s at CL11 is like 4ns latency btw

for comparison, 3200MT/s on CL16, which is typical for ddr4 is 10ns

and 6000MT/s on CL30 on ddr5 as well. 10ns

in theory that 5366MT/s CL11 would perform way better in games

but its still at the end of the day a "pushing to its limit" overclocking

would degrade the hardware quick

[nocturne olive]

That's some insane mem speed
Meanwhile my mem can't even run 4000

[true hemlock]

my 4x16GB runs on 4000MT/s CL19

[nocturne olive]

Not even temporarily, it crashes on POST

[true hemlock]

or higher cl i forgot

wait

[nocturne olive]

3800 runs fine though

[true hemlock]

maybe CL22

[stray dragon]

my 4x48gb ddr5 runs at 3600 MT/s :(

[nocturne olive]

So I just run my RAM at 3800 CL18 all day long as that's now maxed out

[true hemlock]

pfffttt

[stray dragon]

new motherboard should hopefully fix

[true hemlock]

2dpc ddr5 moment

nah

its just how 2dpc works

questions though

[nocturne olive]

Typical DDR5 4 sticks, especially if mixed

[true hemlock]

what CPU, is it mixed kit, what motherboard and to what you're upgrading to

[stray dragon]

9900x, unsure about this term, upgrading from MSI PRO B650-S WIFI to an Asus TUF GAMING B850-PLUS WIFI

[true hemlock]

uh

don't

[stray dragon]

oh

[true hemlock]

B650 and B850 are both mid range

you'll get no difference lol

there's not even any difference in the chipset itself

[stray dragon]

i'll test it and then return it i suppose

[true hemlock]

but yeah who knows

if you can return, nice

[stray dragon]

ye i made sure to get ebay listing with free returns

only some time wasted if it fails

[true hemlock]

does all 4 ram stick comes in a single kit

or is it a separate 2 stick kit

[stray dragon]

nope, 2 of the same 2x48gb kit

let me post my build

[true hemlock]

yeah that explains it

[nocturne olive]

Yep explains the speed

[stray dragon]

https://pcpartpicker.com/user/lettert/saved/xb6223

[nocturne olive]

DDR5 does not like the sticks coming from separate kits

[stray dragon]

i see

[nocturne olive]

Like really doesn't like it

[true hemlock]

even if its the same exact kit, slight difference from different batch production would make it unstable running on different kits

[nocturne olive]

And that's how you end up with DDR5 that runs at DDR4 speeds

[stray dragon]

well there goes that server speedup idea

[true hemlock]

in @fast pagoda 's case he got super lucky lmao

[nocturne olive]

And Azaka got really unlucky with the RAM only running 3000 MT/s

[stray dragon]

buying new ddr5 at this time would be... painful

[nocturne olive]

Most optimally you'd probably want 2x64GB

[true hemlock]

i have an unused 2x32GB kit lmao

[stray dragon]

i have enough spare parts to build a spare pc

[nocturne olive]

Meanwhile I have no DDR5 compatible systems

[true hemlock]

meant for @umbral thorn's PC with the 7950x engineering sample but got the wrong board that doesn't have the microcode for the cpu so...

she gets the threadripper

[nocturne olive]

She sure does like the Threadripper

[true hemlock]

when you have a whole threadripper system just sitting around your room unused

lmao

[nocturne olive]

Seems like she runs the Threadripper at like 4GHz

Not sure if that's typical, high or low

[fast pagoda]

mixing kits without the same model number is not bueno, always has been

[obsidian mantle]

what causes this incompatibility of different pairs of same ram?
wouldnt it be possible for 2 ram sticks from same pair have problems with compatibility?
or they literally get a pile of sticks and then make pairs from them testing each pair for compatibility

[fast pagoda]

i still don't think i'm quite as lucky as y'all are implying lmao it doesn't seem to be as bad of an issue with 9000, certainly not to the extent 7000 had the issues, ive never seen anyone even talk about a 9000 series get limited anywhere much under 5600

[nocturne olive]

With DDR5 pairing kits that aren't from the same package is always not a good idea

[nocturne olive]

Kits are tested to work together

Different kits may have different micro-timings

[fast pagoda]

does intel have this issue? mixing kits i mean, their controller is superior and obviously they run faster speeds in general but

[nocturne olive]

Intel has the same issues I think but it's more tolerant

[fast pagoda]

i just don't believe it's that horrible of a problem after the fiasco of 7000's messed up default "official" max of 5200 MT/s and overall unstable controller in general, that to me is far more of the issue but it's mostly isolated to that series

[nocturne olive]

GPU stress-test be like

[fast pagoda]

yes 9000 uses the "same" controller for io but the IF stability seems to have been very much the issue

[stray dragon]

so i could switch my cpu to something like a core ultra 7 265k, switch the motherboard accordingly, and this pair of 2x48 sticks should run at rated speed of 5200 MT/s?

[fast pagoda]

yes

[stray dragon]

at least that server speedup option is still open

[fast pagoda]

intel has like officially supported over 8000 MT/s

[nocturne olive]

Not quite probably, but it may run at bit faster

[stray dragon]

not keen on spending $500 for that though lol

[fast pagoda]

intel runs 8000+ MT/s very commonly without problems, if he couldn't run a 5200 MT kit on an intel platform it's a problem with the sticks

[nocturne olive]

Intel has a better memory controller but let's not assume it's quite magically good

[nocturne olive]

Well DDR5 mixed running at 3600 sounds like incompatible sticks

[fast pagoda]

well if they're different sets then yes it's gonna be rough

if they are the same kit part number with the same modules it should be aight though, even if it was 2 separate kits of the same ram

[stray dragon]

bought them at the same time, same site

[nocturne olive]

Unfortunately that's not how DDR5 works

[stray dragon]

yeah i'm learning that now

[nocturne olive]

Azaka learned it the hard way too, two kits of the same model at 3000 MT/s

[stray dragon]

ouch

[nocturne olive]

That's lower than my DDR4 which runs at 3800

[stray dragon]

google searches on this issue are really uninformative

[fast pagoda]

install i2c-tools if you don't have it and run sudo decode-dimms

if that doesn't work do dmidecode --type memory --dump-bin spd

see if everything matches up

[fast pagoda]

oh right i forgot i had the wrong syntax and i just mindlessly copied what i already had in my terminal session one sec lol

[stray dragon]

entirely fair

[fast pagoda]

sudo dmidecode --type memory spd

you should get an output from each stick

[stray dragon]

yep

[fast pagoda]

it's not going to tell you what the actual modules are though unfortunately

but it will at least confirm if they are all reporting the same part number+voltages etc etc

[nocturne olive]

Part number means nothing on DDR5

[fast pagoda]

it just means you can't tell what module it is, just like has always been true

the spd info you'd want to tell that is behind other stuff so you can't really get to it

[nocturne olive]

What's this silliness you got there Elvyn?

[umbral thorn]

reeee

1650Mhz 725mV

[nocturne olive]

Sil

Definitely better than my 3090

And then there's my 4070Ti which doesn't overclock at all

[fast pagoda]

you can also run an expo retrain if you have similar sticks that weren't matched from factory (depends on the board what it's called, but it might be called context restore, or expo retrain, goofy stuff)
if your am5 board suddenly sits there for 15 mins twhen you go to boot it's probably running a retrain of the tunable parameters to try and keep the sticks tuned

not that it does that much

can help though

[stray dragon]

all show Part Number: F5-5200J4040A48G and Configured Voltage: 1.1 V, other lines seem unimportant or don't change between any sticks, besides the serial numbers:

Memory Device
  Locator: DIMMA1
  Serial Number: 755DD5FD
Memory Device
  Locator: DIMMA2
  Serial Number: 7F5508F0
Memory Device
  Locator: DIMMB1
  Serial Number: 035EE5FC
Memory Device
  Locator: DIMMB2
  Serial Number: 035138F1

guessing it's just a case of mismatched something or other between the two pairs

[fast pagoda]

4 slots in general is a noticably lower chance of working at full speed already especially on 7000 series

[stray dragon]

yep

[fast pagoda]

and that kit started at 5200 so it wasn't exactly starting at a superfast baseline (although 5200 was the highest "officially supported" speed on 7000)

[nocturne olive]

Based on serieal numbers they're completely different production batches

[stray dragon]

unfortunate

[fast pagoda]

but yeah they have nothing in common there either so likely made maybe not even in the same place lol

[stray dragon]

thanks newegg!!

[nocturne olive]

By the way have you checked your lucky ones?

What's the serials on those like?

[fast pagoda]

yeah i did earlier they looked similar and I was about to grab em

[stray dragon]

where would you even find a 4x kit from the same batch for a reasonable price

[fast pagoda]

nowhere

because no ddr5 is reasonably priced rn

[nocturne olive]

Just get 2x64GB if you need speed and capacity guaranteed

[stray dragon]

speed is lower priority, would still prefer this 3600 MT/s with 192gb over 6000 MT/s with 128gb

[nocturne olive]

Why though?

[true hemlock]

i've actually build a number of 2dpc am5 system to know that you hit the insane luck

[nocturne olive]

What do you need 192GB for?

That 128GB can't do

[stray dragon]

multiple modded minecraft servers, proxmox virtual environment, my habit of not shutting down anything unless it seems necessary

[silent cloak]

trying to make the rendering in my engine platform agnostic

[nocturne olive]

Sil

[stray dragon]

general silliness, yes

[nocturne olive]

Well looks like DDR5 in Finland has also skyrocketed in price

Great, just great

[true hemlock]

again, blame open ai

i fucking wish they went bankrupt

[fast pagoda]

each pair's power management ID is different so they are at the very least differing in that spot which isnt the modules but does to me point to not exactly made in the same batch, possible still same production location and around the same time still though. the serials on 2 of them are similar, the other 2 less so but still seems to be the same batch probably (obviously, because they were matched)

[nocturne olive]

Guess I won't be building any new PCs any time soon

[true hemlock]

YOU'RE PUSHING OVER 2250MHZ NOW???

HOW THE FUCK

HOW

HOW THE FUCK

[fast pagoda]

DIMM 0 (Channel A)
  Serial: AC5F53F6
  Model: G.SKILL F5-6000J3040G32G
  Capacity: 32 GB
  Speed: 6000 MT/s
  Voltage: 1.1 V
  Rank: 2
  PMIC ID: Bank 12, Hex 0x10

DIMM 1 (Channel A)
  Serial: 6D5EC9FF
  Model: G.SKILL F5-6000J3040G32G
  Capacity: 32 GB
  Speed: 6000 MT/s
  Voltage: 1.1 V
  Rank: 2
  PMIC ID: Bank 12, Hex 0x10

DIMM 0 (Channel B)
  Serial: E552FDF6
  Model: G.SKILL F5-6000J3040G32G
  Capacity: 32 GB
  Speed: 6000 MT/s
  Voltage: 1.1 V
  Rank: 2
  PMIC ID: Bank 11, Hex 0x8C

DIMM 1 (Channel B)
  Serial: E95E63F4
  Model: G.SKILL F5-6000J3040G32G
  Capacity: 32 GB
  Speed: 6000 MT/s
  Voltage: 1.1 V
  Rank: 2
  PMIC ID: Bank 11, Hex 0x8C

it seems i did at least match the kits to the same channels

[true hemlock]

what the fuck is elvyn doing

2250mhz on an ampere chip

[nocturne olive]

Sillying clearly

[stray dragon]

sillytime

[true hemlock]

na bro that can't be real

i never even hit beyond 2135MHz on any ampere chip

even with overvolt

i take back my words with ampere being a relatively shit process node

[nocturne olive]

Elvyn can make literally any chip seem insane

If I let Elvyn tune my PC it'd probably perform better while also not overheating

[true hemlock]

this mf kept doing black magic istg

first she gets 230ms latency from Australia to an EU server

with some trickery i don't even get to know

then mf runs minecraft at 248 chunk render... at reasonable 20-38 fps

[silent cloak]

that used to be easy before the new world gen

[true hemlock]

she did it on 1.21.1

[silent cloak]

yeah that system i used was from 2016

i now use it as a server

because it runs a highly modded NVIDIA card

[nocturne olive]

Yep Elvyn would definitely find some black magic way to make my PC not overheat without losing performance

[true hemlock]

this shit

[silent cloak]

its also due to her not running on neuromorphic hardware

[nocturne olive]

Too bad she probably doesn't want to

[true hemlock]

she already did

[silent cloak]

otherwise the model wouldnt nearly need as much

(its inefficient to run them on normal GPU and CPUs)

[nocturne olive]

Did she? She didn't touch my PC

[true hemlock]

her pc i mean

[nocturne olive]

Well I was talking about my PC

[fast pagoda]

im going to know so much dumb shit about ddr5 when im done figuring out who the tf made me modules

[true hemlock]

she tuned the threadripper to be able to pull 351W

[fast pagoda]

exciting stuff

[true hemlock]

on that one small noctua cooler alone

[nocturne olive]

My PC is majorly thermals limited

[true hemlock]

70C

[silent cloak]

im planning on giving my server some new internals soon

still deciding on what CPU

im only willing to spend a few grand

[nocturne olive]

At least before it would easily go 100C when CPU and both GPUs are under full load

[stray dragon]

wtf??

[true hemlock]

huh what kind of server

[true hemlock]

lemme find

[nocturne olive]

My 4070Ti also gets thermal limited when my 3090 runs at full

[fast pagoda]

9970x

[silent cloak]

I have one for some AI, I have another for home storage, I have another that runs an ARR Suite, and I have a experimental neuromorphic system for running SNN models (because they suck so bad on normal hardware)

[true hemlock]

nah get epyc third gen trust

[fast pagoda]

threadripper pro (wx) 9965wx is in that range as well

$2500ish

[silent cloak]

i dont need a custom GPU this time

so i can go far cheaper

[stray dragon]

how do you even tune for this.....

[true hemlock]

implement synaptic plasticity on the SNN, make cortex loops wave encoded

[nocturne olive]

Yep Elvyn indeed does black magic on thermals
I need such magic if my AIO doesn't decide to get out of limbo some day soon

[true hemlock]

[silent cloak]

synaptic plasticity is one of the only reasons i care about SNNs

otherwise its far more A than I

do u run neuromorphic hardware?

[true hemlock]

in my case i can make it run on SIMD, but require double precision so had to use A100 for it

V100 works though

[silent cloak]

there are new cheaper chips coming out

if u want to run SNN and have it be efficient u should look into them

otherwise u have to waste alot of power just to get them to run on normal hardware

essentially "emulating" them

[fast pagoda]

i have still not looked into much about how SNNs are implemented other than high level so curious, does that recent "matrix math running hella fuckin ffast on analogue chips" help snns at all or do they say frick matrices and it's not needed

[true hemlock]

not well documented + my algo is shit

[silent cloak]

some chips u can get for like $500

[true hemlock]

frick matrices most of the time

[silent cloak]

so u can play with them

get a feel for them before upgrading

[true hemlock]

but sometimes depending on your variation you still need them

oh btw vanilla SNN doesn't come with synaptic plasticity

in fact vanilla SNN are useless

[silent cloak]

some people use them just for power reasons

[true hemlock]

this is why most snn library doesn't work well lol

[silent cloak]

i mean its also because there is new meta every week

on what is the best algorithm

went from STDP to other shit

[true hemlock]

exactly why vanilla snn doesn't work

[fast pagoda]

plasticity:

[silent cloak]

i mean one day we will probably have a standard

but its just like when ANNs were being developed

there is no standard because some dude will publish a paper an hour after u finish making one showing u how an old method sucks and a new one is better

but thats part of the research fun i suppose

[rustic crest]

okay uh uh i got a question because i suck at coding and whisper n stuff
basically

how do i make a discord bot not leave vc instantly after it finally joins a vc
and how do i make it listen to whisper or llisten to background discord sounds

[nocturne olive]

You need to connect to the sound socket

[rustic crest]

cause right now it says that it joins a thing but then immediatly disconnects?

[silent cloak]

it cant hear the client if thats what u mean

[rustic crest]

okok ty

[rustic crest]

yeah that too

[silent cloak]

unless u mean hearing other people in the VC which u can setup

[umbral thorn]

oooo wait u do dis too

[silent cloak]

its an expensive research hobby

most chips I have to contact people to get which sucks balls

[true hemlock]

tell that to my monthly spending

[silent cloak]

but yeah try out some neuromorphic hardware and watch ur electricity bill go down

[sage crag]

beep boop

[obsidian mantle]

neuromorphic hardware?

[silent cloak]

https://open-neuromorphic.org/neuromorphic-computing/hardware/

here are some example chips

but there are plenty that arent on here

of course the industry barely talks about them because of ANNs being so popular sadly

[obsidian mantle]

do they have biological shit inside???

[sage crag]

[silent cloak]

no

[obsidian mantle]

oh

[silent cloak]

SNNs work very different from ANNs (Neuro-sama, chatgpt, claude, etc)

[obsidian mantle]

for some reason it felt like it does

[silent cloak]

so specialized chips can run it FAR better and wayyyy cheaper than any ANN model out there