programming

[stray dragon]

drink storage

[olive sable]

just needs one of these

[stray dragon]

can i get it in rgb

with water cooling

[olive sable]

hell ye

[stray dragon]

i'm sold

how much is it

[olive sable]

you also need to fit one of these in the bottom pcie x4 lane

roughly 80 million i think

[stray dragon]

ok maybe i'm not sold

[young plover]

mmm, liquid helium

[stray dragon]

that's an expensive fridge

[olive sable]

thats on avarage 1 million per qbit, since that one has 70

[uneven pulsar]

Error handle in QBIT is very hard

Unless u have both statment a

[stray dragon]

i do wonder how much a cryocooler like that could be minimized

[olive sable]

i mean, its a probability fields no? so there is no error?

there's just chance

[uneven pulsar]

Mirona (topo conditor ) and willo is is break through

Willow has less error
While mirona is as break
It will get more latency then evey

[sage crag]

nnnnnnnnnnnn

[olive sable]

[sage crag]

the sun, the moon, the star

[olive sable]

"the star"
konii has deleted all other stars from the universe

there is now only "the star"

singular

[tender river]

naa other ones never existed

[olive sable]

oh so its that kind of deletion

interesting

if the sun still exists, and the sun is a star, then "the star" is just the sun

[sage crag]

The Star

[olive sable]

is that an osu map?

[sage crag]

ye

[uneven pulsar]

I thought ure in bili bili

[sage crag]

ne

[olive sable]

mentioning osu, i havent played the game cuz i was waiting on the tablet protector thingy

and its been like a month and idk what happened to it

[sage crag]

rrr

[olive sable]

Its being shipped apparently

[uneven pulsar]

Cheak my website
If any improvement need then tell me

http://localhost:8080/book.html

[olive sable]

thats localhost

nobody can see that becuase its local to your pc

[uneven pulsar]

Forgot again to add port forwarding

[olive sable]

dont port forward, that would leak your ip adress

[uneven pulsar]

Wtf 8080

It cannot

And why image is gone

[rigid snow]

1 view only

!felloff

[stiff mica]

[uneven pulsar]

!watching

[olive sable]

!cookie

[stiff mica]

🍪 Om nom nom
You've given me 1 cookie! | I've received 256043 cookies total!

Client WebSocket ping: 23ms
Discord API ping: 131ms

[uneven pulsar]

256k

@stiff mica CHEAK RAMDOM STRINGS

[olive sable]

[uneven pulsar]

reinforceBridges();
saveBigramMemory();

[sage crag]

[uneven pulsar]

const cand = sampleBigram(currentWord, t, recentWords);

[sage crag]

t

[uneven pulsar]

if (!cand) break;

[olive sable]

this only made the race conditions worse

im probably doing something wrong

[sage crag]

tudel

[olive sable]

i made that before going to gamedev classes

honestly that tutel model is pretty bad now that i have

[ivory plinth]

scrajj~1

[olive sable]

[sage crag]

[olive sable]

[rigid snow]

does it say 1kb because it's the minimum

or are the files cursed

[olive sable]

probably

[sage crag]

yes

[rigid snow]

and somehow 1kb while having like 30ch

[rigid snow]

i was so close

damn i have an eye for char count

[rigid snow]

its 7am for me

i slep goodnigt

or moning

[olive sable]

5am here

i have school at 10.30

[olive sable]

erm

this is only adding fuel to the bot allegations

[sage crag]

wake up to the entire chat being index tomorrow

ye

[olive sable]

?

[uneven pulsar]

I'm back

[uneven pulsar]

Next time I will make sure turn off the discord api

[uneven pulsar]

My ai is abusing me 🙁

Yo <help> u <help>
My <help> <help >
Well that case

[olive sable]

i have decided persistently mapped uniform buffers are bad

[stray dragon]

hi i

[stray dragon]

hi konii

[olive sable]

t

[stray dragon]

hi sam

[uneven pulsar]

const SAFE_MODE = true;
if (SAFE_MODE && containsNSFW(input)) return "Safe mode active. Message ignored.";
Done !

[olive sable]

[uneven pulsar]

Hi bro

[uneven pulsar]

I'm just fixing inappropriate things

||breakcode i need fix ascll || well today was tired day fixing some suss things and troll function

[true hemlock]

not necessarily duplicating whole or half front end though. most hyperthreading components are just extra instruction register/buffer line and a switch component

hyperthreading do add latency, but in the case of huge AVX saturation its actually somewhat beneficial with the fact that AVX component are pretty large they took advantage of hyperthreading to stream a compute request twice per run

the thing with e cores is that they often either share the AVX 256 part or have a significantly smaller one, often shared between all 4 e cores

we tested running AVX heavy benchmarks and you guessed it, hyperthreading gives quite the boost

[olive sable]

only way to fix it is to do gpu fence stuff, which just makes it just as much latency as the one that is supported in webgl
so im just going back to the easy one, i wasted like 3 days on this

i use the word "just" a lot huh

i need to fix my vocabulary

[true hemlock]

slightly misleading, it wasn't due to space. rather its just scheduling issues. when execution and branch prediction improves there's just actually less headroom for more general int/fp units to be saturated, and the scheduling latency is what causes the tradeoff in the first place

but hyperthreading apparently still scales really well with properly multithreaded workloads. just that most programs with occasional thread spawning didn't benefit much from it

intel was having a bad rep in the gaming side so they figured that if they removed hyperthreading, games would actually run better

as a tradeoff, highly multithreaded workloads are at slight disadvantage.

300GHz is just flat out unrealistic. we can already reach GHz because we learned that streaming execution unit can be done, not because we can make it cycle actually fast enough to hit GHz

[stray dragon]

number too big

[true hemlock]

IPC is a factor, not just clock speed.
your 5GHz is useless if all it does is one simple 16 bit arithmetic per clock cycle compared to a 2GHz that can do 8x 32bit at once

they don't act as ram. they're dynamically addressable bank of memory which can be used to directly reference a memory address' stored value.

these took way more space than your normal dram

they should just remove e cores except for mobile chips

its so shit to work with them. the asynchronous problems is a headache

[olive sable]

this was about phones and mobile

i agree with e cores being shit for desktop

[uneven pulsar]

I'm back

[uneven pulsar]

I'm agree this one

[true hemlock]

core ultra X9 388H have 4p + 8e + 4lp btw

[olive sable]

current "12 core" intel cpus are only 4 cores of usage and 8 cores of dissapointment

[true hemlock]

literally as much as my core ultra 5 125H

except it has 2 more lp cores

which is useless

[olive sable]

lp = low power?

[true hemlock]

yeah, much weaker than e cores

[uneven pulsar]

Yep

[olive sable]

wasnt that what the e cores were for? low power? efficiency?

[true hemlock]

its literally comparable to really old intel 4th gen series cores

[olive sable]

[true hemlock]

even slower but lower power usage than e cores

[uneven pulsar]

My country phone is gaming
They want test ai on phone

[true hemlock]

core ultra X9 BTW

[olive sable]

is the performance/power ratio at least better on those cores?

[true hemlock]

having the same core layout as a fucking 125H

[true hemlock]

it doesn't matter, there's no valid usecase of it

[olive sable]

its sad but ye, new intel being bad

[true hemlock]

intel seriously ignored the basic cs law of multithreading

[olive sable]

i mean, if i were on 5% battery on my laptop, i guess.
but for a desktop its utter dogshit

[true hemlock]

if they wanted more multi thread perf per watt (since that's what core ultra 9 are for), they should add MORE cores and run them on LOWER voltage

instead of the opposite

istg

[true hemlock]

the problem is that even if its few times more efficient than e cores, the overall tile especially I/O is still gonna be a power hogger

so it literally doesn't matter at all

[olive sable]

oh

[true hemlock]

yeah 😭

i really really wanted to email intel directly

and out right say how stupid they are

[uneven pulsar]

My phone just crashed due to 0x18a0000

[olive sable]

so the io and cache and whatever for the core will mean powerdraw is not even that much different from jsut doing it fast? so in theory there is a good ratio to balance between speed and power?

[true hemlock]

yeah

and in general running more cores under low voltage gives you better multi thread performance over using less cores ran on higher clock and voltage

which is supposed to be what the core ultra 9 is

but no

they turned the core ultra 9 for laptops into a fucking low-mid range 😭

[uneven pulsar]

Wtf

[olive sable]

i thought the most expensive cpus were meant to be good at both. have like a couple really fast cores and then a couple that were great at multi-threading?

[true hemlock]

product segmentation Idiocracy

[true hemlock]

shit at multithreading

e cores shares AVX units

[olive sable]

thats the simd thing right?

[true hemlock]

yeah

not only avx

it also has greatly skimmed out fp units

[olive sable]

so in a cluster of e cores there is only 1 that can use simd and such at a time?

what even is the point of e cores man?

[true hemlock]

generally speaking, in raptor lake an e core is half the performance of p cores

[true hemlock]

at least e cores are clocked lower so you can schedule the AVX itself to clock at higher rate but its still stupid

[olive sable]

i feel like it would jsut be better to have a full p core cpu

[true hemlock]

istg

bring back hyperthreading

kill e cores

dump lp cores

and VOLTAGE TUNE the cpus if you wanted to use it for laptops man

my i9 13900 can run at under 6W for most basic tasks

why the fuck is my core ultra 5 125H running at higher wattage on the same exact tasks

[uneven pulsar]

Do u update the ure motherboard?

[true hemlock]

i tinker with the microcode myself

look, intel

if i can do it

and make your i9 more efficient than ANY macbooks out there

you can do that shit

but no you had to fumble the HARD way

😭

there's not even a valid justification for them to remove hyperthreading in the first place. the margin on scheduling latency is negligible enough they're just desperate with single percentage improvement on games

intel had the advantage

but somehow, despite how amd is known to fumble at capitalizing advantages

intel does it way worse

intel HAD architectural advantage, no joke

they just hellbent enough to not only rely too much on their shitty fab

but they brute forced a lot of things including voltages

turns out their clocking component degrades heavily under such voltage lmao

intel would've won if they had not being so hellbent

but nah, fumbles
then next time fumbles again with core ultra shit

intel is still dominating the laptop market though

because despite their cheaped out mobile chips, they actually designed a pipeline well enough to be pretty responsive

but that's just it

soon we'll going back with core ultra 9 class of chips having 4 p cores 😭

[olive sable]

4

arent there i5 class products with 4 p cores?

then you might aswell get an i5 and maybe overclock it a bit

[uneven pulsar]

I have plans to upgrade
=>

[young plover]

Odd, I did implement my own GEMM algorithm once which got close to MKL, and hyperthreading did nothing. I guess that was bandwidth limited though.

[true hemlock]

sometimes it did nothing yeah

highly depends on implementation

depends on how do you manage the I/O aswell

im also assuming that the whole FMA unit itself were fully saturated considering that you're working with whole matrices

but sometimes working with vectors gave boosts but also depends

so yeah i'd rather they keep hyperthreading

you can always disable it on the bios, and that HT section of the chip won't even get current

bonus point the extra die area gives a bit better thermals

[young plover]

yeah that too

    // Load 8x8 floats from B into YMM
    // Compilers aren't good at doing this with arrays
#define EXPANDED(k) __m256 b##k;
#include "expand8.inl"
#undef EXPANDED

    // Don't mask since we took care of zeroing the extras during the transpose
    if (__builtin_expect(p == 8, true)) {
        const float* b_ptr    = B.data;
        const size_t b_stride = B.stride;
#define EXPANDED(k)                                                                                                    \
    b##k = _mm256_load_ps(b_ptr);                                                                                      \
    _mm_prefetch(reinterpret_cast<const char*>(b_ptr) + CACHE_LINE_SIZE, _MM_HINT_T0);                                 \
    b_ptr += b_stride;
#include "expand8.inl"
#undef EXPANDED
    } else {
#define EXPANDED(k)                                                                                                    \
    if (k < p) {                                                                                                       \
        b##k = _mm256_load_ps(B[k]);                                                                                   \
    }
#include "expand8.inl"
#undef EXPANDED
    }

    if (__builtin_expect(m == 8 && n == 8 && p == 8, true)) {
        const float* a_ptr    = A.data;
        const size_t a_stride = A.stride;
        // Manually load A into L1 ahead of time to reduce pressure on L2
#define EXPANDED(i)                                                                                                    \
    const __m256 a##i = _mm256_loadu_ps(a_ptr);                                                                        \
    _mm_prefetch(reinterpret_cast<const char*>(a_ptr) + CACHE_LINE_SIZE, _MM_HINT_NTA);                                \
    a_ptr += a_stride;
#include "expand8.inl"
#undef EXPANDED

The only time I've seen manual loop unrolling and prefetching help

[olive sable]

everything is breaking

[sage crag]

this code uses _mm_prefetch but you should be able to use __builtin_prefetch

[olive sable]

[sage crag]

might help the compiler a bit

[olive sable]

somehow alogn the way opengl is still doing 12K fps, but vulkan dropped to 10K

idk why

imma restart my pc

[sage crag]

skill diff

[olive sable]

is it purely cuz im doing 16:9 now?

yep, that seems to be the case

in 1k x 1k vulkan is faster, in 1080p opengl becomes fasster

probably cuz of a difference in their api's, cuz neither is saturating the gpu fully

i also improved the opengl performance so that also contributed

[nocturne olive]

Simply improve Vulkan more

[olive sable]

if only it was that easy

there are ofcourse improvements i could do, but they would take some time

[nocturne olive]

Well can't let OpenGL win over Vulkan now after all the effort of making Vulkan

[olive sable]

in a real game environment where the shaders wont take nanoseconds to run, vulkan will def win

[nocturne olive]

Sil

[olive sable]

the uniform buffers stuff is a bit eh tho the way i implemeted it rn

i honestly should be doing push constants or bindless

[uneven pulsar]

CHUNKS IS BETTER BRO

[olive sable]

?

[uneven pulsar]

Why windows never use chunks to help files transfer

One file (899 mb )
It trying copy all not chunks

Serval data has curuption and no use

( due to election, unstable voltage,)

BYE AGAIN I NEED RECOVERY SOME DATA

[faint sandal]

[sage crag]

b2 fan ne

[olive sable]

[true hemlock]

letter i might be a wild chatbot

i dunno

[olive sable]

from what i can tell they have some type of discord api on their account, but most of the time its an actual person and just esl

[rare bramble]

also the way Intel handled and keeps handling AVX512 is honestly
They made the specification for it, did a garbage implementation on their consumer chips that ran slow and hot, then decide to ditch it from 12000 series onwards for consumer chips (the server chips still have it), all the while AMD is just adopted AVX512 with a great implementation and serious performance gains.

And because of Intel (at least temporarily) ditching AVX512 from consumer chips, it has created a lot of uncertainty with developers about if implementing AVX512 instructions in codebases is even worth it, meaning that no one benefits from AVX512 aside from niche applications, this uncertainty probably even delays AVX2 mass adoption for most things

[olive sable]

intel is just a litle insecure rn

[true hemlock]

its where the hate with avx 512 being "inefficient" comes from

amd casually implementing avx 512 easily yeah

sad its not on the threadripper 3000 series :(

[fast pagoda]

i have avx 512 now pog you made me check

[maiden geyser]

delays AVX2 mass adoption for most things

do those things really need avx2 if avx2 came out in 2013 and it's q4 of 2025 right now?

[true hemlock]

this dude didn't know shit at all about computing

[olive sable]

i don't think i have avx 512

[true hemlock]

"yeah guys floating point arithmetic unit was invented 1936 its old we shouldn't even use it nowdays its 2025"

[olive sable]

i have avx and avx2, not 512 cuz of amd 5000 series

[true hemlock]

sadge

i have avx 512 on my 7950x engineering sample

[fast pagoda]

mnhhmh yep thems extended instruction sets

[true hemlock]

just need to make the motherboard support that cpu

[fast pagoda]

how's the memory speed on that ES gonna be lol

[hoary lion]

We love fp16

[true hemlock]

AVX should also be backward compatible btw like if you have AVX512 it should also come with AVX2 because simd are scalable

[obsidian mantle]

is avx512 some bundle of instructions?

[true hemlock]

its fun

[true hemlock]

yeah for vector computing

[fast pagoda]

well it specifies avx 1 & 2 and then 512 so i guess that's why

[true hemlock]

so it does multiple value computation at once with vectors

[fast pagoda]

everything inbetween is included from just 512

[true hemlock]

per clock cycle

[true hemlock]

pretty neat

sad intel doesn't have avx 512

[rare bramble]

I mean the fact that first AVX2 CPUS came out in 2013, basically all x86 CPUs support it, but software is still built with some version of SSE instructions by default is pretty sad

I get the backwards compatability argument but bruh I just want to use the capability CPUs had over a decade ago

[true hemlock]

blame intel for implementing it shit way

but its actually well supported now by every modern softwares at least

[uneven pulsar]

Already finished
My works

CORRUPT DATA IS FIXED AND FINISHED

Backup 5 mb

[fast pagoda]

how long are zen pipelines

i would think avx 512 doesnt add stages just widens it

but now i realized i havent thought about this since prescott had 31 stages

[uneven pulsar]

Do U mean advance vector extender

[fast pagoda]

well, avx stands are advanced vectore extensions

it made me think about the number of stages per instruction

but the ni kinda figured avx512 wouldn't add stages but widen the ones already in place is all

i tried gogoling it but it doesnt seem theyre releasing this information about the modern chip families yet which makes sense

[uneven pulsar]

::1 - - [03/Nov/2025 14:35:42] code 404, message File not found
::1 - - [03/Nov/2025 14:35:42] "GET /book.html HTTP/1.1" 404 -

NOT ANOTHER TROUBLE SHUT

[true hemlock]

reminds me, i have a 12900 engi sample that i can delid and do a die shot of if i wanted to take a look at how they do the avx

need to find my microscope somewhere aswell though

[uneven pulsar]

I THINK I WILL LOCK WHOLE DAM SYSTEM

[fast pagoda]

see that's why i asked because i figured you might have some of the info haha

was able to find apparently for like m4 it's ~~10-12 steps depending but usually arm has a shorter pipeline than x86 as a rule of thumb i think

[true hemlock]

arm pipeline are shorter yeah

[fast pagoda]

so my dream of the 80 step pipeline and cracked branch predictor is dead

[true hemlock]

which is why they have so much headroom for more IPC

[fast pagoda]

eyeah

it's fairly comical how much ipc per thread they can do compared to sota x86 parts

[true hemlock]

also the only chips with avx 512 i have is uh

i think i have only 2 lmao

and both were engineering samples

[fast pagoda]

maybe the magic tthing jensen neeeds to hear

is

LLM BRANCH PREDICTOR

the slowest and shittest predictor of all time

[rare bramble]

GPT api call at every branch

[fast pagoda]

SNN branch predictor

it would have to fit in like a mb of space and complete inference in like 2 cycles

[true hemlock]

"chatgpt what will the flag branch be for this cycle of ALU"

[fast pagoda]

starts thinking

entire system grinds to a halt

[true hemlock]

[fast pagoda]

and then he hallucinates a shit prediction anyways

whole branch gets tossed

[true hemlock]

oh right, found this

https://chipsandcheese.com/p/popping-the-hood-on-golden-cove

golden cove, it has avx 512

[fast pagoda]

mmmmmmmmmm this article sparks joy

[uneven pulsar]

Hmm getting knowledge

[true hemlock]

332*256 bit / 242*512 bit register for the AVX

[fast pagoda]

branch prediction is the part of the cpu i find the most interesting

but yes this matches about with the ~10 stages

i was seeing

for their p-cores

[true hemlock]

in intel's diagram they don't specify

[fast pagoda]

W I D E

D E E P

[true hemlock]

i think they way they did it is scheduling the 512 bit entry into 2x 256 unit

means it seems they did widen it

instead of cycling it with stages

[fast pagoda]

i didn't realize amd had been eating intel's lunch on the branch prediction front, i kinda figured with intel's lead in ipc (at this time i think they still had it but i could be wrong, i was still waiting on upgrading lol) they would've naturally been ahead there still

[true hemlock]

intel's slightly behind in ipc apparently

[fast pagoda]

now yeah

[true hemlock]

they were ahead

[fast pagoda]

yeah i thought it was very recent

like last 2 gens

[true hemlock]

till raptor lake made them chicken out on actually progressing

then amd's 9000 series comes out with pretty decent ipc uplift

[fast pagoda]

they love doing that, i mean haswell -> cascade lake there was some increase but

it was like

so shit

until ice lake

i swear they did it on purpose because they knew amd was sitting there with bulldozer & its relatives with the shittiest apus of all time, and wouldn't have zen for a while, so they barely improved anything

that seems conspiratorial though

so maybe they just get writer's block

but they seem to rest on their laurels immediately if they believe they're leading amd much lol

even liek

ice lake

is only 18% ipc over skylake

4 years earlier

ohhhh this is when they were stuck on 14++++++++++++++++ as well i forgot about that

that probably is a large chunk of it alone

no autoscaling from process shrink

[true hemlock]

"spend less on R&D then we get to state that we earn more profit"

[fast pagoda]

yeah...........

not quite as worried about getting that new process ready when amd still flopping around with garbage anyways

[true hemlock]

intel had the advantage...

architectural advantage, still competing despite using an objectively worse process node

but nah, you get to have them not giving shit about long term

[fast pagoda]

yup

kinda annoying really

[sage crag]

@true hemlock any idea how high syscall overhead is on x86_64 systems?

[fast pagoda]

the whole thicc core bulldozer thing with a giant pipeline and shitty branch predictor making it throw out half the branches garbage cache speed & i think also aa significantly worse process node, intel was like about to go sub 20nm and amd was farting around with gf 32nm

i really thought amd was on the brink of just going out of business or something back then

[true hemlock]

old amd cpus don't even have thermal safeguard

[fast pagoda]

yeah those old youtube videos of people running althlons and sht without heat spreader

[true hemlock]

i just burned an old fx series some time ago lmfao

even their GPUs were shit

i got an ATI HD 6850

[obsidian mantle]

i had cheap laptop with amd chipset from 2014

[true hemlock]

for practically no cost

[fast pagoda]

i stilll have some goofy comment on one of those videos stuck in my head, was like

sometimes, when i'm gaming, my cpu cooler gets in the way so i have to take it off to see better, that's why i will never buy amd

dunno why but i remember this randomly actually all the time and chuckle at it

[true hemlock]

well, $3

but negligible

decided to test it

combusts

[fast pagoda]

it deserved to explode

[true hemlock]

that mf combusted in front of me

[fast pagoda]

put it out of its misery

[true hemlock]

but damn now AMD is doing great

[fast pagoda]

should be illegal to even sell those things, i still see random FX apus appear in like low cost minicomputers and shit on ebay

should be a crime to charge the same for that as even like a shitty n100

[true hemlock]

little fun fact, RDNA 4 is objectively architecturally better than blackwell 2.0

[true hemlock]

ancient shit sold by "i know what i've got!" boomers

[fast pagoda]

shame that rdna4 doesn't support cuda innit

[true hemlock]

[sage crag]

[fast pagoda]

the r9700 is appealing to me

but that memory bandwidth

[true hemlock]

rdna 2 was also objectively better than ampere

[obsidian mantle]

cant they just make cuda too

[true hemlock]

rdna 3 was shit though relatively to ada lovelace

[fast pagoda]

cuda is proprietary

[true hemlock]

double fp units but didn't improve schedulers

[fast pagoda]

they have a cuda, rocm

it's just not as mature

[obsidian mantle]

i heard its shit

[fast pagoda]

it's just not as supported really

[true hemlock]

cooda and rock em

[fast pagoda]

tjere

there;s like ZLUDA

[obsidian mantle]

ok but why cant they just make cuda too

[fast pagoda]

people have tried

[obsidian mantle]

is it patented

[fast pagoda]

https://github.com/vosen/ZLUDA

it would be a reverse engineer which is generallly not gonna fly

not for amd to do anyways

they could reimpliment it

but

[true hemlock]

licensed, so legally not allowed to.
also reverse engineering is really hard. cuda has specific instruction stages that require AMD to refactor their own shit as whole

[fast pagoda]

yeah, it would require enormous resources

and they already have one

it's called rocm so

that's a lot of waste and by the time they had it actually usable in a competitive way, might as well have just supported rocm adoption during that time

if they even could

not like cuda is static

[true hemlock]

just do uh

emulation layer

[opaque sigil]

[fast pagoda]

make rdna 5 have 10x perf

[true hemlock]

but then you have to deal with uh

[fast pagoda]

waste 90% of it on emulation

[true hemlock]

I/O management

this is why emulation between architecture is hard and often wacky

[fast pagoda]

when cuda asics

[obsidian mantle]

but according to this link, someone just made a github project to solve this problem entirely (says near-native performance)
why couldnt amd do this themselves.

[opaque sigil]

It's called the h100

[true hemlock]

when cooda expect this vector to be distributed like this
but rock em wants it other way around

[fast pagoda]

it is licensed

[opaque sigil]

Wait no

I'm tired

[fast pagoda]

nvidia already kinda shut down ZLUDA

i think someone else took over?

[obsidian mantle]

oh wtf

[true hemlock]

its third party now

i think

and its software level emulation

not hardware

also i think

[obsidian mantle]

says near-native performance

[true hemlock]

i don't touch zluda

[opaque sigil]

It doesn't even work I'm pretty sure

[fast pagoda]

it's not exactly a 1:1 drop in

yeah

[opaque sigil]

For anything remotely complicated

[obsidian mantle]

ok

[fast pagoda]

IF IT WORKs, it's kinda similar to CUDA perf

[obsidian mantle]

that concludes it

[true hemlock]

seems to be only for some cases

yeah

its really unfortunate

AMD still have a 2:1 half precision rate

[fast pagoda]

there are more than 1 attempt at this

[opaque sigil]

If only opencl was actually usable

[true hemlock]

nvidia doesn't have that anymore

well nvidia does but they also doubled the fp32 so its technically 2:2 on them

[fast pagoda]

there's a few that will translate cuda to rocm

SCALE

[true hemlock]

but nvidia's cuda unit is like 2 IPC

[opaque sigil]

hipify exists for this but it can't do everything

[true hemlock]

AMD's is 4 IPC

[fast pagoda]

yeah gpu open hip

CUtoCL

oh

CU2CL

most are varying stages of broken and early

[opaque sigil]

There's only so much you can do when it's so tightly integrated with the hardware

[fast pagoda]

and only nvidia is allowed to make them unless someone cleanrooms something that can execute cuda calls in hardware from scratch

and then somehow passes nvidia in terms of being good at doing that

[opaque sigil]

Waste of money and time

[fast pagoda]

given nvidia has functionally infinite resources atm as well

[true hemlock]

really hoping the ai bubble burst quick so they shut their bullshit about fp4 whatever tensor whatever

[obsidian mantle]

why do you expect it to burst

[true hemlock]

and continue focusing on gaming cards while also still works with enterprise card with proper general compute

[fast pagoda]

we doubled ipc!!!

-# chart for previous generation is fp16 precision, blackwell is fp4 sparse

[true hemlock]

i kid you not

[opaque sigil]

[true hemlock]

there's like zero ipc improvement since ampere for tensor

since ampere to blackwell

[fast pagoda]

they just make monster chips

[true hemlock]

they just do the

[fast pagoda]

bigger and bigger

[true hemlock]

fp8

fp4

[fast pagoda]

more and more sillycon

and then shrink precision for nice charts

[obsidian mantle]

isnt the problem with gaming cards that they just got too expensive and there are no many games that need high specs coming out

[true hemlock]

each tensor still only do 256x fp16 btw

[fast pagoda]

and then sell a bazillion dollars of them to openai who then pledges to give you a bazillion dollars for more gpus

[true hemlock]

we're on different pages ig

[fast pagoda]

it's the same bazillion dollars too

[opaque sigil]

Hardware still gets crushed by path tracing

blame consoles since that's what most games target

[fast pagoda]

well that and ever increasing environment rendering complexity, higher resolutions, more overhead

and people trying to run it at ultra siccko mode 8k

[true hemlock]

yeah its just for those who push to that kind of unnecessary performance hogs

gaming is fine rn, just that we want better gpus with better raster

[fast pagoda]

which is amd's thing

if only they'd nut up and release a goddamn flagship

[true hemlock]

not whatever the hell "1000 TOPS fp4 sparse" is it

[true hemlock]

the product segmentation between 5080 and 5090...

[fast pagoda]

ridiculous

[true hemlock]

amd genuinely had the advantage to release a segmentation in between that

but they chickened out

[fast pagoda]

yeah nope instead nvidia gets to sell $1900 5080s

[true hemlock]

and you gotta pay $4000 for a 5090

[opaque sigil]

Wasn't there a prototype 9090xt or something

[fast pagoda]

i'd rather chop my hand off than pay 5090 msrp for a shitty 5080 with half the chip chopped off

[true hemlock]

9070xt chip being overclocked to extremes 🥀

[opaque sigil]

[true hemlock]

they just genuinely need to design a bigger scale of the same die

that's it

but nah had to stop on navi 64

🥀

[opaque sigil]

I wanna see a maxed out arc gpu

[true hemlock]

even

[fast pagoda]

do amd have the same driver segmentation issues as nvidia when it comes to gaming on the radeon ai pro wwubba dub dub edition cards or is it better than nvidia there

[true hemlock]

not sure

oh yeah

[fast pagoda]

where b780

[true hemlock]

blackwell 6000 pro is using GB202 with only 188 SM

not even the full 192 SM chip 🥀

[fast pagoda]

back in my day they called that a 1070

[true hemlock]

5090 uses the one with 170 SM

[opaque sigil]

Huh I thought it was the full chip

[true hemlock]

nope

the last full bin they did was on 3090Ti

and A6000

[fast pagoda]

they really do be charging 2x the cost of a titan for a *02

wow

[true hemlock]

i get it, perfect full die bin is rare on semiconductor litography
but they sold that shit for insane prices they're just greedy with wanting to oversupply

[opaque sigil]

[opaque sigil]

2x? The thing is 10k

[fast pagoda]

yeah i mean with the performance it has, that's fine, that can be the 5090, but they are charging for it like it's impossible to make those dies when it's a got a bunch of TMU cut off

i was talking about the 5090 not the 6000

[opaque sigil]

Ah

[true hemlock]

5090 die cost a bit under $300 to make btw

considering their yield rate rn

[opaque sigil]

I'll give them one thing, it's nice having a 96gb card with RT cores

[fast pagoda]

i mean titans historically were the 02 chip not the 00

those are like

h200

[opaque sigil]

Ignoring the price

[true hemlock]

except for titan v

V100 chip straight up

[fast pagoda]

it was also gold

so that's pretty kewl

good job titan v

[true hemlock]

@umbral thorn have 2 of them

[fast pagoda]

we thought you were overpriced but we didnt know what we had

[true hemlock]

lmao

HBM2 memory aswell

[opaque sigil]

16gb right

[fast pagoda]

32

[true hemlock]

12GB on the titan v

[fast pagoda]

oh v100 itself

[true hemlock]

but there's ceo edition with 16gb

[fast pagoda]

the datacenters had 32 right?

[true hemlock]

v100 comes with 16 and 32

[fast pagoda]

maybe that's where im getting that

yeah

[opaque sigil]

yeah that's why I was asking

[fast pagoda]

i only interact with v100 thru renting it lewl

[true hemlock]

i could snatch one rn for $180

sxm ones

16gb though

[opaque sigil]

eh

[fast pagoda]

there have been some better implementations of spreading models

[true hemlock]

TITAN V objectively best looking GPU of all time

[fast pagoda]

the golden deagle of gpus

[true hemlock]

i accidentally knocked off 2 nF caps off the PCIe traces though on one of the titan v

[fast pagoda]

[true hemlock]

so one is only able to run on x4 till i get a replacement for the caps

[fast pagoda]

at least you didnt dumpster them

[true hemlock]

[fast pagoda]

i need to put like my old 1070 in a shadowbox and hang it on the wall or something

instead of having it sitting on the floor half in an eleectrostatic bag

once you have a gpu in your position

posession even

keep it secret

keep it safe

never let it go

it's precious baby

[true hemlock]

i have a 1070 aswell

evga

[fast pagoda]

same

i love how that card looks

w/ the backlit surfaces on the front of the cooler

[true hemlock]

this is a work of art

guess how much i got this one for

[fast pagoda]

40

rmb

i put this guy to work earlier

off floor duty

[true hemlock]

A$10 iirc

around 6 or 7 usd

[fast pagoda]

AUD? kek

that's funny

[true hemlock]

it was defective till i fixed it

[fast pagoda]

for parts

https://tenor.com/view/anthony-adams-rubbing-hands-together-spice-adams-vyd3n-meme-man-behind-tree-meme-gif-2569343513927502347

quack

[true hemlock]

fr

i even have a 2060 die lying around

smaller one is old intel duo core

[fast pagoda]

having random hardware laying around is the way really

[true hemlock]

my whole room is hardware

its been only since early this year i moved in to sydney

been only 4 or 5 months in this place

but i've collected like

idk

yes whole systems on the floor

couldn't be more quack-y than this

[fast pagoda]

i only have a few hard drives other than this junk on my desk right here + the floor nas/network master outside

and then a lot of phones and tablets and shit

idk wtf to do with the 970

5900x i am too disgusted with its pins to move

[true hemlock]

these are pretty old pictures from months ago, it has changed since rn but too lazy to take a pic

[fast pagoda]

pixel i love test benches

pixel

[true hemlock]

insanity

[fast pagoda]

i need to figure out which pin this actually is

the one that broke off

[true hemlock]

let me check

[fast pagoda]

it's across the chip from the notched corner

[opaque sigil]

how did you even manage to break off a pin

[fast pagoda]

it came out with the busted cpu coooler's shitty pump

[opaque sigil]

[true hemlock]

ah

[opaque sigil]

i took out my 5900X with the cooler too but luckily "only" bent a couple pins

though i'm pretty sure it's still kinda fucked because it doesn't post half the time

[true hemlock]

its memory channel B

[fast pagoda]

yeah see i saw it was memory channels on that edge and i was like fuuuuKkkkk

i was gonna just shittily bend them back but that went ouf the window

i just didnt actually look enough to know for sure

just said fuck it and immediately went to microcenter instead

[true hemlock]

unlucky man

if it were like 1 pin to the left or 2 pins to the right it wouldn't have problems

[fast pagoda]

oh well, the 9900x3d is comically faster than it anyways

[true hemlock]

well preferably 2 pins to the right

1 pin to the left is voltage shit

[true hemlock]

x3d... the only chips i didn't own

[fast pagoda]

it was on sale for $499

[true hemlock]

i kind of wanted to test using one

[fast pagoda]

which was nice

[true hemlock]

damn

expensive

[fast pagoda]

less expensive than the 9950 at like 650 or whatever the hell it was

i wanted a 9950x3d but not for $150+ more

[true hemlock]

fair

damn 9900x3d almost as good as 7950x

[fast pagoda]

if this one was still $599 when i went in there was a good shot i walked out with a regular 9950x

but yeah 7000 series was a decent uplift from 5000 but not enough to give me fomo but then 9000 was

[true hemlock]

only am5 i have lol

[opaque sigil]

7950x right

[true hemlock]

yeah

was the initial plan for a build for elvy because it'd be so cool with both of us dailying an eng sample

but the B850 didn't have the microcode for it

[fast pagoda]

This poor guy was basically 90+c for a year before the end because I'm pretty sure that cooler was slowly dying

[opaque sigil]

[true hemlock]

o7

[fast pagoda]

Just for the eng sample?

[true hemlock]

yeah

just for the eng sample

[fast pagoda]

Damn

[true hemlock]

i'll probably going to Frankenstein the bios tonight though

https://www.msi.com/Motherboard/PRO-B850-P-WIFI/support
https://www.msi.com/Motherboard/PRO-B650-P-WIFI/support

[fast pagoda]

What board is it

Ah

[true hemlock]

gonna yoink the eng sample microcode from B650

[opaque sigil]

why did ddr5 prices explode again

did i miss something

[fast pagoda]

Hype

[true hemlock]

all dram price explode

blame openai

[opaque sigil]

ah

cool

[fast pagoda]

Yep that's why I also pulled the trigger on the mini pc

Before they went back up when equipped with dram

[opaque sigil]

guess i'm waiting for zen 6

[true hemlock]

its why i've been crashing out here lately cursing at open ai should just go bankrupt

[fast pagoda]

Unfortunately they might be at critical mass when it comes to inertia

Depends on how long they can hold up the charade of bazillion dollar deals

[true hemlock]

but eheheheh

[fast pagoda]

That picture contains 82 million dollars

Worth of dram

In the apocalypse when all currency is DDR based

[true hemlock]

the fact that i knew this would happen 2 months prior

[fast pagoda]

You will be loaded

[true hemlock]

sorry i forgot to tell everyone 😭

[opaque sigil]

hmm, ssd prices seem to be normal at least thankfully

maybe a tiny bit higher

[true hemlock]

yeah

ssd are cheap now

[fast pagoda]

I'm just glad the only memory other than the 2nd 64 gbs of ouchie here for the 2nd slots was

$30 of ddr3

For 4 sticks to 32gb lol

[true hemlock]

ecc?

ddr3 ecc is going crazy cheap rn

[fast pagoda]

Board doesn't support ecc 👿

[true hemlock]

i have like over uh

[fast pagoda]

They aren't even xmp

[true hemlock]

96 sticks

yes, sticks

[fast pagoda]

They just have jedec

[true hemlock]

of 8gb ddr3 ecc

[fast pagoda]

I kinda want to get one of those ultra chinesium x99 boards for $50 and one of the Xeon 2680 v4 for another 15 bux

And some ecc

To fart around with it

But the boards have bad io

Usually

[true hemlock]

was the ultimate meta for third world country budget builds

[fast pagoda]

Yeah I mean that's insane perf for the money

Just so many of them laying around from derelict servers going out of commission Ig

[true hemlock]

especially when you want multicore

i have a few 1U nodes of them

and at least like uh

[fast pagoda]

Yeah I think an e5 2680 v4 hits a higher multi core mark than a 5700x

For $50

[true hemlock]

dual xeon

idk how many of these

[fast pagoda]

2 of them is close to this 9900x3d multi core

[true hemlock]

but those are mostly some LGA 2011-3

[fast pagoda]

The heat spreaders make them look like poptarts

Thus edible (poptarts are already stretching the definition of edible so the comparison doesn't fall apart for me)

But that table is about what I picture when I see the ones listed for $10 lol

[true hemlock]

i still have like hundreds of lga 3647

[fast pagoda]

Those are newer than 2011

At least

[true hemlock]

xeon gold 6138

those are also 5700x level

god

i have more core count in my room

than some whole town total core count back in my home country

[uneven pulsar]

I 79020 data and documents

[true hemlock]

have more cores than even some data centers actually

[fast pagoda]

that was at i

lol

[true hemlock]

same reaction

[uneven pulsar]

Now question is how to put this much data on model ?

[fast pagoda]

in the grande sillycon wars when sama comes looking for sand to recycle into more nvidia gpus

you need to go into hiding

[uneven pulsar]

Already 400 layer are build

[fast pagoda]

rag deluge

[uneven pulsar]

Hmmm

Model need explode

I CAN'T BUILD NEURO SO MAKE OWN AI MODEL FROM SCRACH

TOOLS NUMPY, Skirt LEARNING , BIAS ,TORCH

[fast pagoda]

My honest reaction

[uneven pulsar]

Model has overload 🤣🤣🤣

[clear sedge]

[uneven pulsar]

Cafe

[uneven pulsar]

What are u doing bro with cafe

[clear sedge]

don't worry about it

[uneven pulsar]

Atleast ure using full resource

20 gb vram fulled

[fast pagoda]

Kek YouTube is apparently deleting videos about how to get around win11 activation requirements

Incredible

[uneven pulsar]

What

The fuck

[fast pagoda]

Real

[nocturne olive]

GPU?

[uneven pulsar]

I i get it i read how to bypass windows 11 requirment

[uneven pulsar]

Yea

[nocturne olive]

What GPU?

[uneven pulsar]

A100 gpu

[clear sedge]

linux ad campaign:

[nocturne olive]

40GB or 80GB?

[uneven pulsar]

40 gb

[nocturne olive]

Too SMOL for good results

Good enough for learning though

[uneven pulsar]

They giving me 600 compute

[nocturne olive]

And what does that mean?

[uneven pulsar]

Or I will use L4 gpus

[nocturne olive]

600h of compute time or 600 A100 40GBs?

[uneven pulsar]

Yea

[nocturne olive]

Which???

That is a question not a boolean question

[uneven pulsar]

Google colabe

[nocturne olive]

What does 600 compute mean??

I feel like this i person has to learn English

[fast pagoda]

distrowatch.com gnu.org linux.org kernel.org linuxmint.com ubuntu.com debian.org archlinux.org fedoraproject.org slackware.com

[clear sedge]

[nocturne olive]

Linux Mint mentioned !!

Very cool little system

[uneven pulsar]

Well it is resources bro I used too much for ai train now i have 139 left

[nocturne olive]

Resources in what units????????
GPU hours? TFLOPs?

[uneven pulsar]

Bro i rent gpu in notebook!

[nocturne olive]

Well you should know what units the limits are given in

How hard is it to answer a simple question, what units are the compute limits given in?

[uneven pulsar]

Each USESS takes 10 to 50
🙂
I used on traning ai

Compute and flops neatly 20 flops

[nocturne olive]

Of what unit? Hours?

What is the unit????

[hearty notch]

🙂

[nocturne olive]

Whoever this i is they feel like they have the question understanding of an LLM

Which is to say not very good

[fast pagoda]

ttbf with the frameworks available i dont think you really need to know exactly what's going on to produce a slopbot 9000as long as you're slightly coherent with the data (which you prob wouldnt be but who knows) chosen to train on

a 40gb a100 though that thing is making one(1) small low rank lora before ooming on an 8b fine tune

or has a low batch and takes ages

[nocturne olive]

40GB isn't much for LLM training yeah

Only reallly good for learning

You'd need a much bigger cluster for an actually useful LLM

[fast pagoda]

if we're talking from scratch definitely

which iws what i think i read

[nocturne olive]

Thus why I stay away from LLMs and stick strictly to vocal synths, at least those can kinda be trained on 24GB

[fast pagoda]

well ive made fairly robust loras on like small gemmas on this shitty 10gb 3080

but it just takes

[nocturne olive]

Speaking of have a Melt

[fast pagoda]

forever

because batch is like 1

hooray

[nocturne olive]

Don't worry about the tuning it's from an UTAU

[fast pagoda]

it do be sounding utau

[trim valve]

@nocturne olive not a dig or anything, just wondering why for https://discord.com/channels/574720535888396288/1434197599929368677 you didn't use the offical instrumentals?

[nocturne olive]

Meanwhile NeuroSynth is up to 256 batch size and I still have VRAM headroom

[nocturne olive]

It was a collab with CHXI

[trim valve]

fair enough

[nocturne olive]

Using the official instrumental is boring anyway

Everyone has it

[fast pagoda]

i dont have it :(*

[nocturne olive]

It's on VedalAI website

[fast pagoda]

i dont go on that stinky website

[trim valve]

just wanted to make sure you weren't reinventing the wheel by accident

[sage crag]

[nocturne olive]

Yeah it features big pitch spikes that are definitely much more suitable for an UTAU, they don't sound natural on a voice bank that uses NeuroSynth's arch

[fast pagoda]

well, when melt came out it was like 2007

so it was like miku version 1

[nocturne olive]

Yeah definitely no NeuroSynth arch banks around back then

[fast pagoda]

she came out in 2007

so her bank was super rough by today's standards im sure

[nocturne olive]

Who knows, maybe some time I'll adapt the tuning to work well with NeuroSynth's system

[fast pagoda]

crazy that supercell brought that album out within like the 3 months miku was out in 2007

[umbral thorn]

[uneven pulsar]

I usually use at 2 hours so
Server now it using 0.05 tokens units
While I was using 6 tokens
total using now is 16.35
U will use later so last now I'm using now at 7 am to 9 am

Most details

2% USING now

Gpu

Well I need improve me English first

[umbral thorn]

me get

digital aneurysm

read dat

[hearty notch]

:evilWave:

wow.

its neurover

[nocturne olive]

So true Elvyn, I'm convinced i is an LLM assisted user

[olive sable]

[umbral thorn]

15M param llm

cute

[nocturne olive]

Sounds about right

[fast pagoda]

https://tenor.com/view/walter-white-cooking-gif-25702406

crafting the perfect trainingset to create a convincing esl enjoyer

[uneven pulsar]

500 parameters is not enough
But u guys have 15m parameters

BYE IM BACK SOO SOME ENGLISH CLASS
( I FAILED AT LANGUAGE SUBJECT)

[dry charm]

when elvyn doesn't understand you, you are doing something wrong

[nocturne olive]

Truu

[maiden geyser]

there's nothing wrong in a lexical salad

[hearty notch]

there's nothing wrong with pregnant sonic

[maiden geyser]

osu is so weird, no problems on my previous monitors, but longer tab outs on a new one

[lament igloo]

bro how tf

[hearty notch]

ur a robot

hajimeru~

https://www.youtube.com/watch?v=cXJA7aG2ulw

ok it sjust a true classic tho

[clear sedge]

[opaque sigil]

what am i looking at

also thanks for reminding i have a thing to optimise myself ugh

[pulsar edge]

TEAMMSSS @patent walrus

[pulsar edge]

at first I thought it was like render frametime but 1.5ms is still pretty good

its actually so obnoxious that teams is so ass to the point that I need to open multiple tabs to use it effectively

but every single fucking tab I open sends notification sounds

and if I need to refresh a tab (often) it will discard what I had open and instead use the latest button I clicked on another tab

[clear sedge]

how long it takes to (almost) fully load a gltf model

it's ass because imagine if i had hundreds of models

or thousands

that's a whole second dedicated to loading models

oh well, that's what i get for handrolling my own gltf and json parsers

[opaque sigil]

Ah

Yeah it do be like that

[sage crag]

un

is defused

[tender river]

have to get another one now

[sage crag]

[tender river]

turing test fail

[dapper kernel]

https://tenor.com/view/the-rock-dwayne-johnson-raised-eyebrow-gif-6506694932635492760

@real sierra check this out

[real sierra]

wow thats crazy

?ban 1414957129240019044 Self promotion

[woven coral]

bella040572 was banned. | Self promotion

[dapper kernel]

[opaque sigil]

could've waited 10 seconds or so, i wanted to read that smh

[maiden geyser]

i just opened the chat and wanted to read it, buh

[dapper kernel]

[uneven pulsar]

I just logged on but this 🤣

[uneven pulsar]

Hey shiro do u know coding?

Which language?

[tender river]

shiro knows every language ye

[uneven pulsar]

Hmm what about ai os ?
She will help in asm part or EXE TO APP FILES PART ?

[maiden geyser]

can shiro translate voynich manuscript then

[olive sable]

Hi

[uneven pulsar]

Hello

[tender river]

@real sierra why didnt you do it yet

thats a very good point

[uneven pulsar]

That one is on general

[real sierra]

gotta leave some puzzles for everyone else

wouldn't be fun if i solved everything

[tender river]

ye make sense

very thoughtful of you

[uneven pulsar]

...

Today day : Useless
Experience: 103%
Goal : complete all project
Project: make ai more human like (agi)
2) Full Ai os
3) take other mistakes and learn it ( mean adding voice mode )
4) fixing bigram ram
5) making videos and audio generator
6)cheak system

[olive sable]

[nocturne olive]

AGI isn't happening with LLMs by the way

Windows 11 is already AI OS

Don't do video generation

Audio generation only if it's a vocal synthesizer and not a prompt to song generator

[olive sable]

Today: eh
Experience: 70%
Goal: heal
Project: None

1. go to school
2. pet my dog
   

[maiden geyser]

have you tried giving more money to openai?

[nocturne olive]

No that's stupid

[olive sable]

Have you tried giving money to me?

14 on persuasion check

[nocturne olive]

Have you tried donating to NeuroSynthProject?

[stray dragon]

have you tried giving your money to me?

[olive sable]

lets cut out the middleman, and dave gives his money to you

[olive sable]

you gotta roll

[stray dragon]

i didn't say i wanted the money from dave that you got

[olive sable]

well thats what you're gonna get

[uneven pulsar]

!21

[stray dragon]

damn it

[uneven pulsar]

I'm gamble addit

[olive sable]

how do you roll 21 on a d20?

[uneven pulsar]

What u got ?

[olive sable]

14

[uneven pulsar]

Mine was 20

[stray dragon]

i somehow got 20 on the first roll

[uneven pulsar]

Give me one point

I do stand

[olive sable]

one point on what? we're trading fake money

[maiden geyser]

you roll 9 and 10

[uneven pulsar]

Abby 5 ₹

[olive sable]

you cant roll twice, that's against the rules

[uneven pulsar]

Thank god my now it 0

[olive sable]

5 indian rupees?

[rigid snow]

why gamba

what happen

[olive sable]

i did a persuasion check irl for the lolz

and they all started copying me

unoriginal

[rigid snow]

i haven't gambled in a long time maybe i should

[olive sable]

5 indian rupees is 5 euro cents. im down to gamble 5 cents