programming

[sage crag]

[opaque wharf]

The jianghui has invaded again

[fast pagoda]

https://tenor.com/view/one-punch-man-opm-black-dance-dancing-gif-21561002

[sage crag]

i saw real disappointment in that message

you cannot be allowed to live

perish

[opaque wharf]

What the f is that

[real sierra]

wat

[fast pagoda]

golden

[real sierra]

I will probably form an iron core as most stars of my class do

[nocturne olive]

That things like a million times bigger than a NeuroSynth

[sage crag]

foundation establishment at the ninth tier and soon the qi in your dantian should condense into a golden core

right

[real sierra]

meow

[sage crag]

o

[fickle rain]

Get to Nascent Soul then we’ll talk

[opaque wharf]

https://tenor.com/view/god-why-ugh-nooo-ahh-gif-27454757

[olive sable]

Im not even home to defend myself

[sage crag]

if youre not at least in the void refining stage dont message me

[fast pagoda]

[real sierra]

it's ok I'm not home either

the timeline here is like

days or weeks

[fast pagoda]

galaxies form

[olive sable]

Oh

[fast pagoda]

black holes die

[olive sable]

Oh im fine then

[real sierra]

indeed my PC will be in a box until May after Wednesday

[sage crag]

black holes die on a much greater frequency than galaxies form

for example

[fast pagoda]

god bless this discord

will be gone

[sage crag]

a black hole just died in your nose

and

[real sierra]

what

where's it going

[sage crag]

probably in every other cell

[fast pagoda]

incinerated

[real sierra]

don't incinerate this place I work hard to avoid that

[fast pagoda]

for free

wait

admin

is nit free

?/?

[sage crag]

shiro what will you do if proton decay is real and the bonds between the quarks of this server reduce to energy and then via entropy dissipates slowly as heat

[mighty thorn]

Nuh uh

[mighty thorn]

Probably die

[fast pagoda]

fdkin hate when those raids happen

[tender river]

shiro will

make a wish

and save the server

[sage crag]

is the wish made of protons

is shiro made of protons

[mighty thorn]

Yeah I hate when the server gets raided so all of the protons in my body dissolve into quarks and I melt into a puddle of pure energy

[sage crag]

better question

[fast pagoda]

nothing is maed of proteins

[sage crag]

the

[opaque wharf]

the

[mighty thorn]

the

[glass jetty]

ThEEE

[opaque wharf]

Nay, thee

[warped narwhal]

THHHHH

-# e

[fast pagoda]

that right there

[amber fractal]

T he

[fast pagoda]

i gotta say

that

that right there

that stuff

that's the

that there

that's technical

[sage crag]

idiot

[fast pagoda]

wtf i did nothign

[fast pagoda]

https://tenor.com/view/thanos-avengers-infinity-war-mind-stone-last-stone-infinity-gauntlet-gif-15735729

[sage crag]

sentenced to listen to DADADADADADADADADADA 23 times

[fast pagoda]

Mini had been truly ymtc'd

[fast pagoda]

i will listen to it but my condition is

i must pull the entire music library from neuro karaoke site 6 times

as it is my favored speed test now

[sage crag]

https://www.evilkaraoke.com/song/cc4091ef-058d-402d-9c77-e8e707dcf2f2

[fast pagoda]

mmhjm you see 1 song i see a vast colection

[sage crag]

[fast pagoda]

ope

im on windows rn

it's over

[opaque sigil]

aight that's it, your toast privileges are being revoked

[fast pagoda]

NO

im restarting now

[opaque sigil]

[sage crag]

oh

filtered

do not screenshot, repeat, or bypass this automod

you will get warned

[tender river]

ohisee

[sage crag]

see rules, open a ticket if you have a question

[fast pagoda]

the rules:

https://tenor.com/view/palla-deserto-desert-hot-gif-6014273

[sage crag]

ye the only rule in the jianghu is the law of the strong

[fast pagoda]

based department?

[tender river]

zep is

pretty strong

[sage crag]

fellow daoist zeppelin

[fast pagoda]

that's wjhy the zepp has been able to exert unlimited power for years

[stark needle]

wrrrrrr

[sage crag]

zeppelin belong to what sect

or is it a loose cultivator

[tender river]

rwx

[sage crag]

[tender river]

zeppelin listen to

shiro orders

[fast pagoda]

clasdsic rwx sect

[tender river]

[fast pagoda]

rwby sect

[sage crag]

[fast pagoda]

https://en.wikipedia.org/wiki/Quanzhen_School

fotm

[sage crag]

this isnt like my webnovels

why arent they flying on swords

[fast pagoda]

there's definitely yuri

just look harder

[sage crag]

wrrr

this world isnt even a wuxia

let alone a xuanhuan or a xianxia

[fast pagoda]

it's so ober

I made an amazing discovery today

I had one box of unopened iems

Just waiting for me to discover

It's Christmas in July

Never forget

Booting is going greT

[tender river]

don't forget the original intention of using headphones is to enjoy music

[fast pagoda]

Do not

Forget

[sage crag]

do not not forget the not forgetting that you forgot to not forget that youve not not not forgotten that you will forget later

[fast pagoda]

It's supposed to continue to say "which is why these headphones are kind very similar to the other ones that look the same "

[sage crag]

this is my wisdom for you today

[fast pagoda]

Thanks

[sage crag]

pleas digest it carefully and comprehend the dao

[fast pagoda]

I am contemplating

[sage crag]

Thought for 3m 53s

I cannot boot your computer because it is unsafe and goes against my principles

[fast pagoda]

Soon as you said that I've entered a state of disenlightenment

That's what get for launching windows twice

[sage crag]

im replacing every call to any sorting functions in glibc and the linux kernel with

bogosort

in specifically your computer

[fast pagoda]

Please at least bogomips

[sage crag]

it parallelised bogosort on the gpu

highly fast

pleas do not fear

[fast pagoda]

I'mma be fucking cruising

[sage crag]

replacing every sort function with

[fast pagoda]

Bogoshart

[sage crag]

a gui that you have to manually sort

[fast pagoda]

Czkawka moment

Or krokiet as they like to say now

[sage crag]

replacing every sorting function with a call to memset(ptr, 0, len)

[tender river]

[sage crag]

it sorted

[tender river]

replace sort with llm call

[fast pagoda]

Reboot once and completely fine

I love computers

[sage crag]

replace every call to sort with

uh

uhhhh

something

[fast pagoda]

This boot rudely replaced my login screen with the joke windows XP one

[tender river]

replace computers with

girls kissing

[fast pagoda]

Let's go

[sage crag]

o

[fast pagoda]

Now that's public policy in action

[sage crag]

every call to sort has to mine some bitcoin before it allows you to use the sorted list

and the bitcoin is paid to me

[fast pagoda]

3d printer filament uses waste heat as it cools to mine btc

[rough bloom]

nop

no more sorting

embrace chaos

[tender river]

[sage crag]

replace every call to sort with

ud2

[amber fractal]

Sure I love stress testing my CPU with a sorting test. Wait where is all of my Bitcoin???

[fast pagoda]

Naive slime search

[tender river]

your

[fast pagoda]

Use slime

[tender river]

[fast pagoda]

But the slime is an idiot

[sage crag]

replace every call to sort with a fluid simulation where the density of each particle is represented by its value in memory

ye

[fast pagoda]

Mfw gotta use Lawrence Livermore to search

[tender river]

replace every sort with

optimization function

solve using

simplex

[sage crag]

z <-

[fast pagoda]

I'd love to search for a file but I have to get the contract passed first

[sage crag]

replace every sort with

yaccho

[tender river]

[rough bloom]

[sage crag]

https://tenor.com/view/yachiyo-gif-11851451255531159989

[tender river]

sorry i mean

bless you

[fast pagoda]

replace every search with a konii instance

[sage crag]

do not

[fast pagoda]

forced to search by hand with midnight commander

and also ranger

they swap randomlyu

[real sierra]

I like this idea

[sage crag]

i am highly inefficient to compute

do not sim

[real sierra]

but I already have Mito search

[fast pagoda]

vedal search

[sage crag]

@cobalt vessel

[fast pagoda]

he cancels it till tomorrow

[sage crag]

mito hi

[tender river]

mto

[umbral wigeon]

Jarvis

[real sierra]

@cobalt vessel please post the recording of the neurocord vc during the unban vc stream in 2023

[fast pagoda]

someone start calling vaedl green guy and shit until he posts for the first time in this chat since 1967

[sage crag]

how many times does a toast have to ping vedal before the whole programming channel gets banned

[fast pagoda]

1

[sage crag]

how do we test it

[real sierra]

he posted here semi recently, I remember it

[fast pagoda]

can ew post the vod when twitch unbanned nudity for no reasson

[real sierra]

he almost fired me for adding something to the faq

[tender river]

over 32

[fast pagoda]

lmfao

[sage crag]

no i mean

if every toast

right now

one by one

pinged vedal and got muted

how many would it require to cause a mental breakdown and hence

a removal of the channel

[real sierra]

mass mute

would not remove the channel

[sage crag]

yes but slowly

[real sierra]

just the chatters

[sage crag]

screeps discord js

[real sierra]

ringleader would probably vanish first for inciting it

[amber fractal]

Toast culling

[sage crag]

good thing i dont know

[fast pagoda]

he hasnt killed everyone yet

[sage crag]

anyone who would suggest such a thing

[tender river]

i personally wouldnt

but

ring is

cyclic linked list

supporting O(1) removal

[fast pagoda]

is this inaccurate yet

[tender river]

toast ring resilient against purges

[sage crag]

toast ring

[fast pagoda]

onion ring

funyun

what the fuck

[amber fractal]

?mute Toast Toast Toast Toast

[sage crag]

yes why ddid you think i was calling you

fonion

[sage crag]

operations upon toast ring remain within toast ring

[sage crag]

[fast pagoda]

this being a transformer exploding is truly a sign of intelligent design

a metaphor for me hiding in plain sight

[sage crag]

i thought it was a ring of intelligent communicating mosquitoes

[fast pagoda]

that's the 2nd episode of one punch man you're thinking of

[sage crag]

how can you punch mosquito ring

[fast pagoda]

that's the plot of the fight

[sage crag]

oh actually if punch speed exceeds speed of sound you can kill them with pressure wave

[fast pagoda]

in a sense

[sage crag]

should be trivial

[olive sable]

My h670 has arrived

So

Theatre pc is complete

Except i need to find a 7900xtx for my desktop with 3 DP ports, and then i can put the current 7900xtx in the theatre pc

[opaque wharf]

So it is not complete

[fast pagoda]

8 port vega 64

[olive sable]

It is complete

But my desktop isnt

[stark needle]

hi toast

[olive sable]

Hi non-toast

[stark needle]

what's letterswap lmao

[olive sable]

Oh

Thats still there

Letterswap is 2026 arc 1

You missed that arc

[stark needle]

damn

im just gonna set up my own kubernetes cluster, how hard can this be

😂

[reef ingot]

Day 4 or 5 making my own game engine
lerning bsp(binary space partition) no code yet

[fast pagoda]

btruh

[glass jetty]

i dropped this idea around second week, when actually learned opengl

a lot of work for no reason, i wouldnt add anything new

[fast pagoda]

yeah 4 making my own c++
no language yet

[glass jetty]

finally JUST banana toast

[reef ingot]

Oh mine is build for my own compter which I made on a os which is still under construction

[glass jetty]

year 4?

[reef ingot]

Mine is a bit weird it's built to run on a computer I made

[glass jetty]

human lifespan is not enough to make computer

what exactly is your

btw, any news about photonic(a)?

[fast pagoda]

jensen thought about photonic compue while walking the streets and trying to find a bathroom,

that's all

[opaque wharf]

I have a firm believe that k8s is done by people that doesn't want to deal with networking

[amber fractal]

Thoughts on Tailscale?

[opaque wharf]

Different purpose

But still, prefer wireguard

[amber fractal]

Fair, was something I wanted to roll out eventually

Was using wireguard before and while I'd be alight reusing it personally I have other reasons to be using the Tailscale.

[opaque wharf]

If you need better management then sure, tailscale or any other vpn provider would prove to be more easy to manage and deploy at scale

[amber fractal]

In my case, I'm wanting to expand from my local net and link up with a few other people

So Tailscale makes sense in that regard

[olive sable]

[amber fractal]

I feel too tired to do logic, so let me slave away at the setup

[opaque wharf]

Well, that depends on your ISP. If your ISP provides a public IP then tailscale may not be needed. Unless you also don't want to expose that IP

[amber fractal]

I know I personally have a public IP but I have to agree with a few others who aren't as willing

[opaque wharf]

Well, the others can just disclose their IP only to you

[amber fractal]

I do want to make it clear, every other person I'm linking up with are in this chatroom

Either way I don't mind much

I think it is fun at least

[opaque wharf]

Because after every peer are connected to you, and you configure things correctly, only you will be able to know the incoming IP

[olive sable]

i think tailscale is jsut a nicer solution than having to deal with portforwarding or whatever

[amber fractal]

outing youself as part of the cluster

[opaque wharf]

Ehh, depends on the requirement really

[olive sable]

[opaque wharf]

Also, you don't have to port forward everyone

But again, depending on the design requirement

[amber fractal]

I think the management outweighs the overhead personally. Might as well be flexible with permissions as some people defently need them

including myself

[opaque wharf]

Yeah, as I said, depends on the design requirement

[olive sable]

basically all i need is a free acces mode when my pc is on, and a locked mode when i need all of my own hardware dedicated to me

[opaque wharf]

Yeah, maybe you're better off with tailscale

[olive sable]

btw

what is a finti?

[opaque wharf]

Skydiving is back on the menu bois

[amber fractal]

[olive sable]

[amber fractal]

I do usually check for this kind of thing, depends on the lanuage

Nothing in English but in some others it has a defenition

[olive sable]

i think locally its a brand for wood stuff

carpentry or something

[kind nimbus]

Fml i missed the stream start again

[olive sable]

Finti is de fabrikant voor uniek gemodificeerd Noord-Europees hout voor duurzame kozijnen, ramen, deuren, boeidelen, gevelbekleding en tuinhout-toepassingen.
Finti is dé duurzame en milieuvriendelijke vervanger van tropisch hardhout. 

oh

uh

englsih

Finti is the manufacturer for uniquely modified Northern European wood for sustainable window frames, windows, doors, buoys, cladding and garden wood applications.
Finti is the sustainable and environmentally friendly replacement for tropical hardwood. 

[amber fractal]

Out of google translate I've seen Fake, Flax, and Male

[nocturne olive]

NeuroSynth sings "Machine love" but it's SynthV

[glass flower]

i can hear the teto accent in this somehow

[nocturne olive]

Well it's literally the Teto files ported directly to NeuroSynth including all tuning parameters

[glass flower]

[nocturne olive]

Thus, NeuroSynth but it's SynthV

[glass flower]

makes sense

[trim valve]

i am going insane

[nocturne olive]

[olive sable]

i am sam

[nocturne olive]

You are Toast 🔺

[trim valve]

I have been trying to make a userscript play nicely for the last 4 hours

for some reason onnxruntime-web insists on trying to fetch an mjs file

[nocturne olive]

OnnxRuntime is pain

[trim valve]

which would work if it wasn't ignoring my cache of blobs and instead sending a request to https://discord.com/channels/1256420997725290556/ort-wasm-simd-threaded.mjs

[trim valve]

it's about my only option here though

[nocturne olive]

[trim valve]

I think theoretically I could diy a wasm module in rust myself

but that sounds like a massive pain in the ass, and less performant

maybe I should write an app myself so I can debug this better and actually figure out whats doing this

[vague sonnet]

chat should I draw, do ML research, lock in on the vedal stream or engage in self-destructive behaviours?

so many options so little time...

[nocturne olive]

Draw

[vague sonnet]

[amber fractal]

Wheel

[trim valve]

chat is it a bad idea to make my userscript use the gpu

[vague sonnet]

we should abandon CPUs and do everything on GPUs

[trim valve]

wrr

its still fetching wasm

mr library developer

kindly consider the don't

I wish there was a nice way to debug this userscript issue

but I am sadly not a js nerd

[opaque sigil]

apparently my code runs 50x faster on a 5070 ti than a b300, seems fine to me

[nocturne olive]

B300 is not really good for anything because it's all in on FP4 and such

[opaque sigil]

it's all integer math and severely memory bound so i was hoping the latter would at least carry but i guess not

it's still 4x more SMs

who knows, i don't think i feel like spending the money to figure out why this happens

[amber fractal]

another Blackwell... Wait how in the frick

Good to know, do not the b300 for int opts

[opaque sigil]

i have a suspicion that it has to do with the fact that there are two separate dies but idk

it's also very atomic heavy

[nocturne olive]

Just don't use the B300, it's an LLM inference only accelerator

[amber fractal]

gotta train that fp4 native somehow

[fast pagoda]

Theoretical Performance
Pixel Rate
48.77 GPixel/s
Texture Rate
1,202.9 GTexel/s
FP16 (half)
1,231.8 TFLOPS (16:1)
FP32 (float)
76.99 TFLOPS
FP64 (double)
1,202.9 GFLOPS (1:64)

x2

wayyyyyyyyyyyyyy beyond A 5070 ti in actual compute EXCEPT not very far off on fp32 and and over 2 die, it's insane fp16

[opaque sigil]

also has like 10x the memory bandwidth

[fast pagoda]

2.46 petaflops of fp16

with 2x 4.10 TB/s mem bw

jesus

[nocturne olive]

B300 LLM inference card

[fast pagoda]

unrelated, i'm assembling a crew for a job, anyone experienced in . pentesting the security of the largest corporation in the world

[amber fractal]

yes it is

[trim valve]

ok incredible

I don't know how I did it, but it works

[amber fractal]

true, that isn't a fun and satifiying answer to what is happening

[opaque sigil]

there's plenty of things apart from LLMs that benefit from the memory speed

though often it's more like you moving the bottleneck from compute to memory on purpose ig

[nocturne olive]

Well it's stripped of all compute

Well useful compute anyway

FP32 and FP64 suck compared to other cards in the enterprise class

[opaque sigil]

fp32 is fine, fp64 sucks though yea

[fast pagoda]

what precision are your datatypes? int?

[opaque sigil]

all u64

[fast pagoda]

ok that'll do it

as insane the fp16 is, it's even more insane in how shit it is at fp64

[amber fractal]

fp64 sucks everywhere, actually a challenge to find cards with good fp64

[nocturne olive]

Usually enterprise cards have good FP64, the B300 is an exception

The Titan V also has some good FP64 rates apparently

[opaque sigil]

that was their main selling point for the longest time

makes no sense financially for nvidia anymore

[fast pagoda]

only good in the sense that theyre neutered somewhat less than your average consumer suffers through

generally

outside of a few specific cards and then like the ultra top end 50k+ cards

[amber fractal]

Unless you are Intel for some fricking reason

[fast pagoda]

yeah intel said fuck all that ratio shit what is that

1:1:1 16->32->64

[nocturne olive]

If I wasn't reliant on CUDA I would buy Intel cards

[amber fractal]

50TFLOPs my beloved

[fast pagoda]

[opaque sigil]

i have to choose being a memory or compute bottleneck hmm, surely scientists are more likely to work with workstation cards right (this is specifically for biologists)

[fast pagoda]

not with this unit

[nocturne olive]

That thing is basically 1.6 times a 3090

[minor crag]

Holy vram

[nocturne olive]

Just looked at the VRAM

[fast pagoda]

yeah you should have seen iggly crash out

wwith these

[opaque sigil]

isn't that card EOL

[amber fractal]

Not that I'm aware of

[nocturne olive]

No CUDA

[fast pagoda]

it's not being made anymore

but it's still in use

[stark needle]

intel isnt that terrible fortunately

[nocturne olive]

But I rely on CUDA

[opaque sigil]

time to learn sycl

[nocturne olive]

[opaque sigil]

it's not that bad

[fast pagoda]

and i mean

b300 is like

[stark needle]

amd is the one whos dogshit

[fast pagoda]

uniquely ULTRA shit at 64

like

so fucking bad

so bad

[stark needle]

tbf most ai researchers dont use GB300

[amber fractal]

Some men in black wiped your memories because you brought it up like it was new after I already mentioned it

[nocturne olive]

Does CUDA PyTorch stuff work on Intel

[amber fractal]

Intel on these cards has CUDA at home

[fast pagoda]

i only bring that up because this was about @opaque sigil's code running way worse on a 300 than aa 5070

[stark needle]

in theory if u change .to("cuda") to .to("xpu") yes

[fast pagoda]

that's a theory with rocm too

[stark needle]

lmfao

[fast pagoda]

https://tenor.com/view/you-wouldnt-get-it-joker-smoking-gif-15952801

[stark needle]

but intel already has been running pytorch on cpus

only problem is triton kernel support

which is absent

[nocturne olive]

Going trough and changing everywhere that appears would probably be annoying
Also does OnnxRuntime do Intel?

[fast pagoda]

openvino is crazy

honestly

[stark needle]

onnxruntime does intel ye

[nocturne olive]

[stark needle]

everyone does onnxruntime lmao

[fast pagoda]

on rocm, there's HIPIFY which just translatres cuda kernels to hip

[nocturne olive]

And how pricy are those insane GPUs?

[stark needle]

good question

lmao

[fast pagoda]

ebay some of the lower ends are like

1500

pretty rare tho

and the lower end part

not the monster

[minor crag]

Why is the Intel Data Center GPU Max 1550 $8,550

[amber fractal]

Very good question I still can't get my head around

[fast pagoda]

idiots pricing yolo for a rare item

[stark needle]

intel data center gpu max with 48gb is up on ebay for like 5k lmao

[opaque sigil]

how does hipify handle inline ptx, does it just do what i assume it would? (aka give up)

[fast pagoda]

3 on ebay rn

[nocturne olive]

[fast pagoda]

a lot of exploding

[opaque sigil]

makes sense, not like they have a real choice ig

[fast pagoda]

it doesnt explode really

what happens is

it just gets left there

and then doesnt work

or works wrong

[amber fractal]

I mean the monster of the linup isn't even the one we showed, that appears after you move to SXM

[fast pagoda]

from the hipify resources: During the porting process, it is important to check for inline PTX assembly code, CUDA instrinsics, hardcoded dependencies, unsupported functions, any code that limits the size of the register file on NVIDIA hardware, and any other constructs that the hipify tools cannot convert. Since the hipify tools do not run the application, it is necessary to manually change any hardcoded constructs, an example being the warp size set to 32. For this reason, it is recommended to avoid hardcoding the warp size, and instead rely on the WarpSize device definition, #define WARPSIZE size, or props.warpSize to get the right value from the runtime. THe hipify tools do not convert build scripts either. Setting the appropriate flags and paths to build the newly converted HIP code has to be done manually.

[fast pagoda]

the 4x one is 4x the one i posted lel

[amber fractal]

Why it is called Max Next is beyond me

[fast pagoda]

https://rocm.blogs.amd.com/software-tools-optimization/hipify/README.html

[stray dragon]

intel naming schemes are always stupid

[fast pagoda]

if interested

[amber fractal]

Intel's brand decisions are also stupid

[fast pagoda]

[amber fractal]

I've seen too many failed products I'm fricking tilted

[fast pagoda]

ii purchased the ssd from a dgx spark today

that's about as close as ill get to getting one of those probably lule

[stray dragon]

laptop cpus? fucking put whatever for the model name and market it as "intel i5"
desktop cpus? surprisingly normal except for the core ultra series where they toss away the i5 i7 i9 thing they've been doing forever
gpus? break the established naming scheme with the other brands so you have to look up what gen the intel GPU is from
i'm not even surprised they call their datacenter cards "max next" without any number at all

[stark needle]

maybe we are finetuning deepseek with this one

[fast pagoda]

bro they wont even tell you WHICH fucking i5 it is

[stray dragon]

don't even get me started

[opaque sigil]

arc gpus have a pretty sensible naming scheme tbf

[fast pagoda]

every time i go and look just to see how much of a pain in the ass it is

[stray dragon]

don't you dare

[fickle rain]

How many software libraries are named Triton?

[stray dragon]

it makes sense when you look it up but then i have to go look it up again when it comes up

[stark needle]

1 triton openai

[fast pagoda]

core ultra 7!!!!
ok which one
which gen
which model
any info on core count, clock?

NOOOPE

[stray dragon]

i think the gen is marked by the letter and they do A B C with those? i don't enterely remember

[fast pagoda]

look inside

2p cores

[opaque sigil]

do you? higher letter = better, higher number = better

[fast pagoda]

150 250 350 etc

[opaque sigil]

letter is the gen

[fast pagoda]

which they dont tell you

[stray dragon]

for GPUs?

[fast pagoda]

no, cpu

[stray dragon]

yeah for the core ultra cpus

[fast pagoda]

they will say literally

[opaque sigil]

celestial when

[fast pagoda]

"Core Ultra 5!"

nothing else

[stray dragon]

i'm surprised they didn't go with a 1500 2500 3500 pattern given how many laptop cpus they have in each gen

[fast pagoda]

which could mean so many different fucking things

[stray dragon]

"i7 8th gen" and there's TWENTY NINE of them

[austere glacier]

Still not over buying a 32GB MI50 for what I thought was a deal before realizing ROCm dropped gfx906 support

[fickle rain]

There’s also binary analysis library named that

[fast pagoda]

this at least you can generally follow

but not listing the model number or any core info on the product is imo should be a form of false advertising or something

[nocturne olive]

If I could afford big GPUs I would go all in on Intel

[fast pagoda]

there's a community that patches them

[stray dragon]

the model numbers in there are insanity and there is an 8x difference in performance between the best and the worst of the cpus in that list

"i7 8th gen" means NOTHING

[austere glacier]

Yeah thank god

[nocturne olive]

Except that it will allow you to install Windows 11

Not like you should

[stray dragon]

lmao

[fast pagoda]

"i7 8th gen" means NOTHING yeah this is the part i am saying is the worst

[stray dragon]

oh and did you see that they're fucking up the core ultra naming scheme even more

[fast pagoda]

at least you can still literally just go up the numbers if you have the numbers

which isnt always al lthe way perfect either

but

[stray dragon]

there's core ultra 3, core ultra 5, core ultra 7, core ultra 9, this is expected

[fickle rain]

Well the 8xx are 1st gen

[stark needle]

lemme guess this was quack

[stray dragon]

but there's also core ultra X9 and core ultra X7

[fast pagoda]

yeah the 800s were lynnfield i7ss on lga 1511

[stray dragon]

let me check

[fast pagoda]

no they are

[amber fractal]

Nah this was actually mine

[fast pagoda]

that was the first generation i bought my own pc

i had an i5 750

couldnt afford an 875k

couldnt afford nehalem 920 etc

[stray dragon]

huh ok

i genuinely did not know

[fast pagoda]

they keep looping

it's funny

[stray dragon]

because they put hundreds numbers in their thousands gens

[amber fractal]

Much to People's surprise I actually did most of the hardware hunting

[olive sable]

[amber fractal]

At least this time around

[sage crag]

t

t

t

t

be toast

[olive sable]

konii

i

i

i

[sage crag]

hi sam

[fast pagoda]

hey i am "supposedly" hella big supported on rdna4 (r9700 ai pro)

and

[stray dragon]

hi konii

[olive sable]

hi konii

[fast pagoda]

well

[sage crag]

hi

https://cdn.discordapp.com/attachments/1126496895246807123/1488042236993339483/6D972AD6C73A5B53D7A000B454B9E52A.gif

[olive sable]

https://cdn.discordapp.com/attachments/1072697081443131476/1488557244432777376/6D972AD6C73A5B53D7A000B454B9E52A.gif

[fast pagoda]

i had a 1.5tb folder on my hdd leftover last time i built rocm

[stray dragon]

you've got shit like the "Intel Core i7-1068NG7" from the 10th gen

[fast pagoda]

yeah they really went crazy style for a bit

[stray dragon]

absolute horror of a model name

[austere glacier]

Yeah I got a 20GB 7900 XT, an MI50, and a- HOLY

[tender river]

t

t

t

t

[stray dragon]

hi chayleaf

[fast pagoda]

there was a 2nd 500 gb of build files i didnt find till later also

[tender river]

hi konii

[stray dragon]

what

WHAT

[tender river]

sorry

[stray dragon]

i'm not konii.

[tender river]

i meant

hi t

easy mistake

to make

[sage crag]

hi clueful

[stark needle]

@true hemlock bro i found out today that there's a 368 core 1.5GB L3 cache amd cpu setup with 6.6TB/s ram bandwidth with 14.3TB 50GB/s read, 30GB/s writes to SSD

https://learn.microsoft.com/en-us/azure/virtual-machines/sizes/high-performance-compute/hbv5-series?tabs=sizebasic

[tender river]

kronii

[fast pagoda]

and the build was basically 3 daysof this

because i had to builtd it

20 times

[sage crag]

bzz bzz

boil me some water with your cpu

[fast pagoda]

i hadd pbo at 95

[sage crag]

you could repurpose that cpu as central heating for those 3 days

[fast pagoda]

and forgot to undo every run

[sage crag]

[fast pagoda]

[tender river]

[nocturne olive]

Cache-only Linux

[austere glacier]

I basically have given up attempting to local train at this point, but it might be worth for inference

[fast pagoda]

me too lmfao

well

[stark needle]

the highest bandwidth llm inference setup was a cpu all along😭

[fast pagoda]

i just have to be prepared for days of pain

[nocturne olive]

[fast pagoda]

sweet summer children

[olive sable]

[stark needle]

368 cores no smt WHAT

[opaque sigil]

4 sockets

[nocturne olive]

I just train on my 3090
Would train on Intel if I could get one of their insane big GPUs

[austere glacier]

I'm attempting to combine XSA and AttnRES to handroll a 1B model as we speak pray for both me and my wallet.

[stark needle]

does it even use that much vram

or am i dumb

[sage crag]

room temp is 31c, why is acpi 16.8c

[nocturne olive]

22GB minimum for NeuroSynth training

[stark needle]

wtf is it a huge model or what

[stray dragon]

built-in air conditioning

[sage crag]

surely this is a random division by two

that found its way in here

[nocturne olive]

Not sure, but it really finds a sweet spot for scaling at 22GB

Once it reaches 22GB it's easy to crank batch size up to hundreds

Before that even 12 is a struggle

[stark needle]

do note critical batch size

[olive sable]

i didnt know this existed

[nocturne olive]

[stray dragon]

silly maximum temp

[tender river]

first has critical temp above highest temp

[sage crag]

dont question

the

[stray dragon]

i think that's like interior of the sun kind of temp

[sage crag]

sensors

[sage crag]

its threshold temps not the actual range of temps that have been measured

otherwise

-273.1 is

[tender river]

sensors

[austere glacier]

Oh is that the thing where if you make your batches too big it make training quality worse?

[tender river]

cultivated for much longer than me

i wouldnt understand their divine wisdom

[sage crag]

0.05 kelvin

so basically

[stray dragon]

that's around liquid helium temp

[sage crag]

the vacuum of space

[tender river]

its a good sensor

[olive sable]

[tender river]

detects anything from absolute zero to 10 times the sun's photosphere

[nocturne olive]

What is a critical batch size

[olive sable]

this thing says my gpu is capped at 303W?

[stark needle]

isnt the model itself like 700-800MBs tho

[nocturne olive]

Around 300MB

<100M params

[stark needle]

weird i keep either seeing 50MB or 800MB weights on huggingface for that thing

maybe ur using a custom config idk

[fast pagoda]

azure has rank 5 on the TOP500

with their ND_H100_v5 instances

jammed into a bllob

that you could rent

theoretically

https://top500.org/system/180236/

[nocturne olive]

NeuroSynth isn't anything super weird

[fast pagoda]

this would qualify as renting harder than Kazuya rentla GF

[stark needle]

i got access to our entire orgs azure setup

my boss seems to trust only me for the ml infra

cause everyone earlier invented one off non compliant scripts lmao

[fast pagoda]

plz https://www.netlib.org/benchmark/hpl/ run this

on your last day

[stark needle]

wtf is this digikey shipping

the worst one was best all along

[vague sonnet]

"digi" and "shipping"
back in my days digital goods didn't need physical shipment smh

[stark needle]

i bought the 200gbe cable for my dgx sparks

here in switzerland they cost 200 somehow

on digikey they cost 50 euro

lmao

[vague sonnet]

just make one yourself

[fast pagoda]

theyre like 50 bucks here

[stark needle]

here in switzerland theres not many datacenters

so its kinda exotic

[fast pagoda]

but on amazon

[stark needle]

amazon doesnt ship here mostly

[olive sable]

what

how?

they ship everywhere

[fast pagoda]

against the geneva convention

[stark needle]

on like 80% of stuff it doesnt ship into switzerland lmfao

[olive sable]

are you on the swiss amazon website?

[stark needle]

need to send to my mum's bf's parents in germany always

[fast pagoda]

doing buziness with bezos legall qualifies as suffering an injury of war across borders

[stark needle]

there's only amazon.de

[olive sable]

no

[fast pagoda]

amazon.swisscheese

[olive sable]

whaty

how????

[stark needle]

bro switzerland is not eu bro

[fast pagoda]

2026 bro anything's possible

[olive sable]

that shouldnt matter

what the fuck

[stark needle]

u can just buy cheap junk u could alr get in switzerland but with shipping from eu + import

[fast pagoda]

try to buyss some nice goods on ebay japan

[olive sable]

almost all the central european countries have their own amazon site

[opaque sigil]

no they don't

[olive sable]

how the fuck does Switzerland not have one?

[fast pagoda]

it's switzerland also

[opaque sigil]

amazon germany serves half of europe

[safe path]

oh tim is no longer cooking

we have a new apple ceo now

[fast pagoda]

switzerland is their own thing

[olive sable]

what

[fast pagoda]

WAHT

NOOOOOOOOOOOO

not timmy

[safe path]

https://www.cnbc.com/2026/04/20/apple-names-john-ternus-ceo-replacing-tim-cook-who-becomes-chairman.html

[fast pagoda]

WH

[opaque sigil]

is that the hardware guy

[stark needle]

what

[sage crag]

ternus

[stark needle]

who tf is this mf

[sage crag]

supervillain name

[safe path]

https://www.apple.com/newsroom/2026/04/tim-cook-to-become-apple-executive-chairman-john-ternus-to-become-apple-ceo/ apple news room

[fast pagoda]

john termius

[opaque sigil]

oh it is

[fast pagoda]

ol johnny ternary

[amber fractal]

WHO IS THIS MAN

[stark needle]

[opaque sigil]

he was vp of hardware engineering

[fast pagoda]

i guess they were kinda killing it

the only good part of apple

[tender river]

who let him

cook

[fast pagoda]

🥀

[safe path]

vibe coded on apple silicon ™

[fast pagoda]

wwith apple intelligence

[stark needle]

bro how was wikipedia alr updated

??????????????????????

[fast pagoda]

the intelligence we used to purchase gemini credits

[fast pagoda]

https://tenor.com/view/ghost-in-the-shell-keyboard-gif-7519694

[stark needle]

pov wikipedians the millisecond someone dies

[fast pagoda]

there's no faaster race in the world

[safe path]

btw did u guys see, github copilot finally introduced session limits and users are mad

https://github.com/orgs/community/discussions/192485

[fast pagoda]

except the one to immediately purchase and scalp all new tech anything

[amber fractal]

What do they do when they come back to life a day after

[fast pagoda]

https://tenor.com/view/godzilla-let-them-fight-fight-ken-watanabe-watanabe-gif-5687182

[fast pagoda]

RACE AGAIN

[stark needle]

me when companies don't give me free stuff for no reason that i kept abusing constantly

also hi vanor

[safe path]

hii

[opaque sigil]

can't have shit

[safe path]

100% people connected their subscriptions to claw or hermes

[stark needle]

openclaw isnt even good😭

[safe path]

for what copilot offers, its dirt cheap

[fast pagoda]

i mean i agree they cant give it away froeever and such but how about maybe then we dont use 95% of the economy to vacuum up alll resources and sell iit back for a loss just to get ahead in the slopwars and then complain

[stark needle]

its free with student account, just exploit ur kids student edu email

afaik

[safe path]

have kids to use free copilot subscription

[opaque sigil]

worth it

[fast pagoda]

i have kids every year for jetbrains

[safe path]

wait this solves the global population crisis

[stark needle]

yea u even get gemini pro with student

[opaque sigil]

i wish

[stark needle]

for 1 year

[fast pagoda]

i got gemini pro for a year with my pixel 10 proxxl

[olive sable]

netherlands has amazon.nl
belgium has amazon.com.be
france has amazon.fr
poland has amazon.pl
spain has amazon.es
czechia has amazon.cz
italy has amazon.it
de serves germany, denmark, austria and Switzerland i think?

slovakia, hungary, croatia and such i cant find info on

[stark needle]

rip lmao

switzerland is not eu😭

[olive sable]

that shouldn't matter

[stark needle]

eu free trade

[opaque sigil]

i don't feel like searching for the page but german amazon ships to most of the eastern countries (not every item but most of them)

[fast pagoda]

cant use gemin i pro for much because gemini still cannot into tools and i never use chat that isnt claude

[opaque sigil]

which is what i meant

[fast pagoda]

switzerland is just the cayman islands if you slammed them in the middle of yurop

[safe path]

😔 when will vedal release subscriptions to neurogpt

[olive sable]

the specific website makes sure you only see items that ship to your country. which is why i find it strange switserland doenst have that

[fast pagoda]

ponzi scheme gpt

[amber fractal]

One he doesn't die of a million prompt injections to backslash

[vague sonnet]

erm
actually you can ask a question when you sub

[amber fractal]

-20k moment

Also -$5 moment

[safe path]

just 4 more of those will get u another claude sub

so, 4x neuro = 1 claude?

[vague sonnet]

neuro is cheaper and answer quality is about the same

choose wisely

[safe path]

ok yeah this is insane tho

[vague sonnet]

lol

[fast pagoda]

he got margin called on his tokens

[vague sonnet]

the corporations are so smart
when their paid service costs them more than subscription earnings they can just cut access from paid users

can't wait for sponsorship messages to get injected into chatbot responses

[safe path]

u get an unskippable ad tailored to ur prompts

[olive sable]

like god intended

[fast pagoda]

actual pharma (both sides of the board) practices

give samples and free use

get people hella dependent

alright yoink

pay more or use less or both (involuntary)

except it's ai psychosis

[stark needle]

when do u make ur own llm

and host it via api

[fast pagoda]

tbf it's still googl model but i randomly tested gemini 3.1 pro vs gemma4 31b hosted locally over api and gemma4 won

not in speed tho

[safe path]

ya me and my infinite ethically sourced data that definitely exists

[fast pagoda]

fair use if u steal everything everywhere all at once

[safe path]

likely to sample ethically sourced data

[stark needle]

im gonna try running llama 400b in 2 days when i get my connector lmao

[safe path]

should just train LLMs on the library of babel

[fast pagoda]

they basically did

they trained on aa

[stark needle]

what a genius

infinite training tokens!!!!

[safe path]

!!!

[olive sable]

just

use the library of babel

no llm

[nocturne olive]

400B

[fast pagoda]

1.4PB of data that prettty much all the labs accessed

[sage crag]

generate the library of babel with llms

[olive sable]

tower of babel 🔺

[stark needle]

library of konii

[stark needle]

its prob gonna run shit on like 3t/s

afaik

[nocturne olive]

[sage crag]

tiring

[stark needle]

my bet is that better llms are gonna come out soon

[fast pagoda]

[stark needle]

i just need gemini 3 pro performance

and im set

[sage crag]

my bet is that all llms will vanish from the earth in 3 days

[fast pagoda]

gemini 3 cannot into tools

[stark needle]

multimodal

[olive sable]

they will a l start speaking different languages and get scattered over the world

[sage crag]

in 1093 years, it will be discovered that they migrated to the moon io

[fast pagoda]

when an insider speaks you listen

[sage crag]

https://tenor.com/view/zato-zato-i-love-the-world-zato-i-love-the-world-and-everything-in-it-maria-maria-kaplan-gif-2976066453239429674

[stark needle]

im waiting for qwen 3.5 omni

thats gonna be god tier

[olive sable]

you're waiting for the sequel

im waiting for the prequel

qwen 0

[stark needle]

im just gonna rewrite lilac as an agent from scratch, run on 2 dgx spark

and hook her to a phone app

with a camera

[fast pagoda]

🥀

[stark needle]

wait wtf i could afford buying a live2d model easily

[nocturne olive]

Could one of those DGX thingies even run NeuroSynth at real time speeds
Or run NeuroSynth at all

[stark needle]

live2d models are how much? like 1k?

[olive sable]

shadow the plebian still in the 2D era

where im going we have the third dimension

[fast pagoda]

they can run most things just not necesarrily blazingly fastr

[stark needle]

dgx is for ml training mostly

[fast pagoda]

big pool of mem _+ clusterable

[olive sable]

lmao

i misread that

as

[fast pagoda]

not good throughput

[nocturne olive]

NeuroSynth speed null

[olive sable]

"big pool of men"

[fast pagoda]

he saw what he wanted to see

[stark needle]

?

[olive sable]

[stark needle]

bro

are u aware

i have android app with live2d waifu support

[nocturne olive]

triangle shape with 3 triangle shaped sides

[stark needle]

i just need to wire it to lilac

[fast pagoda]

is that the one that i have on my phone

[olive sable]

no 3D?

[fast pagoda]

from u

[stark needle]

3d controlling is annoying asf

[olive sable]

not really

[stark needle]

im just gonna try to beat operators shit inference stack

@viral oasis dont let me beat u ok thx bye

learn all ml in 5 days

[olive sable]

[fast pagoda]

every inference stack is a house of cards made of hope, a dream, and the most brittle pile of dogshit that ever somehow worked

[stark needle]

im gonna code something openai would envy

[fast pagoda]

"holy shit he wrote a comment like manually???"

"how do i get 5.4 to marionette my hands tro do that"

[stark needle]

wtf live2d model costs only 800$

thats actually cheap

[nocturne olive]

Almost 2x 3090

[stark needle]

looking at vgen rn

[fast pagoda]

there's actually free ones there's one named hiyori

[nocturne olive]

I want to be able to consider that stuff cheap too
Broke

[stark needle]

i need to have her as live2d

[olive sable]

she's already rigged in 3D no?

[stark needle]

yea but controlling is garbage as hell

unless u make preset animations

[olive sable]

you jsut gotta make presets

ye

[stark needle]

still if i make her live2d ill get more details

[olive sable]

not specifically animations, but like face expressions at least should use presets

[stark needle]

3d looks inherently kinda crap

especially since her model isnt a 2000$ 3d model

[nocturne olive]

[olive sable]

the biggest difference with 2D is probably the line art

[stark needle]

wait lemme see if that artist is taking comms

nope shes not

rip

theres a god tier artist that makes 3d look like 2d

wtf her model prices were 1000-1500$ now they are 4000$ minimum nvm

bruh

corpa

[warped narwhal]

I have used some maths and a 60 line python script to make a rocket in ksp.

the kerbals will probably not survive, but it will be fun to watch this thing fly

[stark needle]

we are vibe coding rockets with this one

bruh

[trim valve]

why does my userscript sign me out of discord

also uh

[nocturne olive]

Average

[stark needle]

@nocturne olive if u want i can run benchmark with that tool that ur using for neurosynth for training on what the biggest batch size possible is

[trim valve]

also it seemingly leaks vram

im such a good developer

[ruby star]

Someone get vedal to see this, hearing that he’s making them generate full usernames instead of just converting casual/nicknames to IDs was.. painful :P

[warped narwhal]

there's nothing like a nice launch in the morning with a solar eclipse in the background

[ruby star]

I mean I should’ve guessed from the “Minik” thing when interacting with mini, but

[nocturne olive]

[kind nimbus]

<im_start>spearfishing

[stark needle]

<|vq_15363|> my beloved

[kind nimbus]

from where is that from? lol

[stark needle]

supposedly gemini

[rigid snow]

[nocturne olive]

How does one even benchmark "biggest possible batch size"?

[fast pagoda]

vibe

[leaden crest]

NeuroCheer

took me way longer to get the extension than I shouldve

[keen hatch]

2xDGX Spark

[cobalt vessel]

slightly delayed search but here https://cdn.discordapp.com/attachments/1059569601144442911/1067819008717496320/countdown.mp4?ex=69e78ef2&is=69e63d72&hm=1cad8bc237e0d3b30bdf78d6c0d02501c2a1eedd3c41f658dc8b7581e9cf81e0&

[glad path]

WE STOLE THE SUN FROM THE SKY

[glad path]

the audio quality dying because of the amount of people is so funny lol