programming

[true hemlock]

i got it alongside an adapter too

[fast pagoda]

lmfao that's teh same adapter

is there only one of those being produced ive only ever seen that same pcb for the adapter everywhere

ive not looked for another variant

128gb usb 3.0 for 7 bux a drive isnt bad

not good because wtf would i need it for but

[true hemlock]

how much i paid back then vs how much it cost rn

[patent shard]

32gb 6000mhz CL32 stick was $75
now it's.... I don't even want to know

[true hemlock]

that was in aud btw

[patent shard]

I remember the 96gb ram for $146
that was crazy

it was like 5200mhz CL42
low tier stuff.. but for that price, whew

[true hemlock]

ram was dirt cheap back then

[amber fractal]

ram is finally being treated as important, for better or for worse

[true hemlock]

project stargate is still fucking nonsense

we'll have such huge companies dropshipping ram chips/wafers soon lmao

[amber fractal]

no one is achieving AGI with an LLM

[olive sable]

nah

all we need is to make an anime girl in a glass cylinder, surely that will be agi

[true hemlock]

im talking about the fact that there is no realistic way of them using >1EB of ram let alone that much per month

[amber fractal]

KV cache for training (real)

as the curse goes, the more systems you add. The more limited you are by sheer networking

[true hemlock]

they could've gotten like

maybe 80PB of HBM and it'll still be overkill in capacity

but no had to max out dram with like 1EB per month somehow

[amber fractal]

need that 1q sized model

surely scaling will work in it's favor

-# q is quadrillion, Q is quintillion; don't know who decided that, but seems to be accepted

[true hemlock]

frontier HPC supercomputer is at 4.8PB of system ram rn

about the same capacity in terms of HBM because they used 4x MI250x per CPU

[fast pagoda]

what's el capitan at for system ram

oh

5.44pb

that one is epyc 42cs and mi300s

same lineage tho

[amber fractal]

Actually, this seems feasible as it is just over double what people are already doing. ovbs they are going for 100q for stargate

drain the entire power grid while your at it

[true hemlock]

the rank 1 on TOP500 rn uses like 43,808 APUs (24 cores Genoa + MI300A) and shares 128GB HBM per APU

[fast pagoda]

that's el capitan

i forgor cray became part of HPE

[amber fractal]

don't know enough about training to figure out how much ram that'd even require

[fast pagoda]

quack you must hate JUPITER

[true hemlock]

3rd place "Aurora" only uses 10PB of ram

[fast pagoda]

aurora runs SUSE tho

which makes it based

and also has Xe gpus lol

[true hemlock]

GH200 lol

[fast pagoda]

it's your nemesis

[true hemlock]

nah

[fast pagoda]

so many arm cores

[true hemlock]

my nemesis is B300

[amber fractal]

e waste

[true hemlock]

i can see working on JUPITER be an ass experience since you'd have to work with the ARM systems

but its still workable

[fast pagoda]

microsoft NDv5 azure VM come in at 5 in the TOP500 that's funny

[true hemlock]

they still run simulations and HPC research on it

a "supercomputer" consisting of B300 would not run HPC

and since i consider ANN dead, it is indeed an e waste lol

[fast pagoda]

https://top500.org/system/180236/ imagine this will become more common

as they build out this ridiculous bullshit

[true hemlock]

HPC6 is surprisingly efficient

[fast pagoda]

eagle is only 1900800 accelerator cores and 172,800 cpu

2,073,600 total

power bill: :^)

[true hemlock]

what counts as cores really

cuda cores?

warp schedulers?

sm?

[fast pagoda]

h100 cores

14,400x132

[true hemlock]

like... fr...?

[fast pagoda]

172,800
(3,600 × 48-core Intel Xeon Platinum 8480C @2.0 GHz) cpu + 1,900,800
(14,400 × 132 Nvidia Hopper H100) GPU

[true hemlock]

only um

132??

or

wait

oh that's sm

[fast pagoda]

14,400 h100s

[true hemlock]

fair

[fast pagoda]

they plopped that thing down in 2023

when can i see the linpack

on some more dc instances

theoretically i can purchase azure credits to use eagle

[amber fractal]

Admitting this is chatgpt clearly. This would be 1q if chatgpt didn't screw up the math.

[true hemlock]

okay

[fast pagoda]

**To Put That in Perspective **

🥀

[true hemlock]

so apparently TOP500 count FLOPS with fp32?

[fast pagoda]

yeah it's LINPACK

no sparse LINPACK

there's talk of using the hpc challenge

but it hasnt happened

since like

20 years ago

more comprehensive tho

[true hemlock]

nvm

linpack is fp64

i forgot the CPU also computes

and i forgot theoretical performance chart is at normal boost clock

[olive sable]

im getting financebro business mindset slop on my yt shorts, and its about sam altman

[amber fractal]

[true hemlock]

sam triangle when

[fast pagoda]

i dont think much was lower precision than that for hpc for al ong time

there's fp32 linpack versions tho

[olive sable]

[true hemlock]

wanna test running it

@willow path

[amber fractal]

after game graphics are entirely solved

[fast pagoda]

https://tenor.com/view/yoda-there-is-another-star-wars-jedi-gif-5140031

[fast pagoda]

https://www.netlib.org/benchmark/hpl/

oh there's a gh

https://github.com/icl-utk-edu/hpl/

[true hemlock]

if imma be honest

[fast pagoda]

HUGe numbers on this repo

[true hemlock]

her 3990x would've crushed any of y'all system for fp64

[fast pagoda]

but would it defeat the might atom

(yes)

but

[true hemlock]

[amber fractal]

Nah, due to the p40s being out of commission until I can get them working all I have is a 3060 which I'd not even bother to bench with

[fast pagoda]

 interface,  or the  BLAS  C  interface,  or the VSIPL library
 depending on which computational kernels are available on his
 system. Only one of these options should be selected.  If you
 choose the BLAS Fortran 77 interface, it is necessary to fill
 out the machine-specific C to Fortran 77 interface section of
 the  Make.<arch>  file.  To  do this,  please  refer  to  the 
 Make.<arch> examples contained in the setup directory.```

fortran 77 let's goooo

[true hemlock]

actually

imma ngl

she's crushing ALL of us

😭

i just remembered on top of the 13 TFLOPS her 3990x has

she also has 2 TITAN V

[amber fractal]

I ain't winning any race because laptop

[true hemlock]

each with like god knows how much fp64

[fast pagoda]

like 7-8 yea

[true hemlock]

they're unlocked from P2 state

its way more than 7 to 8

[fast pagoda]

Pixel Rate
139.7 GPixel/s
Texture Rate
465.6 GTexel/s
FP16 (half)
29.80 TFLOPS (2:1)
FP32 (float)
14.90 TFLOPS
FP64 (double)
7.450 TFLOPS (1:2)

at stock anyways but who knows with unlock

[true hemlock]

10.7 TFLOPS fp64

[true hemlock]

hers run at 2.1GHz

P0 state

stock is 1.45GHz

[fast pagoda]

such a tragedy that it went from 1:2 to 1:32

frick you jensen

[amber fractal]

such a shame indeed

[true hemlock]

2167MHz

iirc we did hit 2.2+ GHz once

okay

no

the threadripper was the least worrying part

yet it would've already beaten every single one of us

only way i could win is that i run that shit on my epyc 9655 server

[fast pagoda]

9995WX should be about 6 tflops on its own

[true hemlock]

way more

3990x is already 13.2 at fp32

6.6TF fp64 on that shit

[fast pagoda]

not real world linpack anyways

https://www.pugetsystems.com/labs/hpc/AMD-Threadripper-3990x-64-core-Linpack-and-NAMD-Performance-Linux-1666/

1.5

3990x is 16 fp64 flops/cycle/core

if it sustains like 4ghz it should be doing 4 tf

[true hemlock]

AVX 256

so 8x fp32

[fast pagoda]

they fuse 2 256bit

[true hemlock]

yeah

[fast pagoda]

it doesnt change the thruput

zen5 has 512bit avx512 true anyways

[true hemlock]

so i suppose in total of 1 AVX256 per clock cycle

3.22GHz all core for the 13.2TF fp32

[quick condor]

I mean if I crammed my 3090tis in one of my r640s somehow

[fast pagoda]

3090ti has dogshit fp64

[true hemlock]

^

[fast pagoda]

it's 1:64

[true hemlock]

afunyun beat me to it god damn it

[fast pagoda]

it does 40tflop fp32 but 625 glop fp64

glop

sao moment

[amber fractal]

glop is correct

[true hemlock]

hascrack. you're not beating the frog's TR3990x in fp64 even with all your GPU lol

1.5TF seem to be outdated score if what i read in the article is not mistaken

[fast pagoda]

the mighty atoms all combined might come close if they could work together somehow

[true hemlock]

they probably could

[fast pagoda]

cuz those were like 1.6 or something

[true hemlock]

linpack is literally made for clusters

[fast pagoda]

and i have 5 of them

rn

[amber fractal]

checks out

[fast pagoda]

the 9700/9070xt manages 1.5 tflop even despite 1:32

same as 5090 kekw

[quick condor]

The server is running dual xeon gold 6240s, maybe if I went with platinums

[fast pagoda]

1:64 moment

[amber fractal]

maybe if bf32 existed we'd complain less (no shot)

[quick condor]

I think I can do like 1.3tflop per R640

With the golds that is

[fast pagoda]

the xeon golds should be around 1-1.5 depending on how they clock

if it downclocks to 1.9 all core like they probably do unless very well cooled doing all core sustained avx512

that'd be 1.1

single core boost will get .125 per core at 3.9ghz

[true hemlock]

i'd expect her to hit 20+ TF

but depending on how good the saturation is

14TF is probably already guaranteed

on the titan v

tr not included

[amber fractal]

learning more about avx512, man that is a lot of registers for some reason. But I'm not about to complain

[fast pagoda]

she should be close if it can get 3-4 on the cpu (realistically) + 5+ on the gpus

[quick condor]

I should really be running 8280s

[true hemlock]

2nd gen xeons lmao

wouldn't even get close to the threadripper alone.

[fast pagoda]

those would probably get ~2

[young plover]

Cascade Lake should have 0.5 FMADD/cycle throughput iirc
smadge

[fast pagoda]

28c, 32fl/cycle, 2.7ghz = 2.4 theoretical max

[amber fractal]

[true hemlock]

iirc zen 2 should have 1

[quick condor]

Those also have about 9 grand worth of ram in them right now so ehh... Besides I have actual shit running on them that I'm not taking down for a highly specific non real world benchmark so who cares anyway

[true hemlock]

32/cycle?

[fast pagoda]

they have 2x avx 512 fma per core

[true hemlock]

god fuck

eh that's still fp32 though

[young plover]

It has 4 cycle latency though

[true hemlock]

and not sure if the 2 avx512 is being used in turns

because some systems does that

[true hemlock]

latency won't matter if we stream it

[amber fractal]

welp, o7 to me who is trying to minmax an avx512 program on intel hardware

[true hemlock]

god fuck, 0.5 CPI

[young plover]

2 units with 4-cycle latency -> 0.5 throughput

[fast pagoda]

it's .5 cycle/instruction = 2 instructions/cycle

[true hemlock]

that

[fast pagoda]

2x8 doubles

x2 flops

= 32

CPI is just intel autism

[true hemlock]

i am still wondering if that's for both AVX or just per AVX module though

because iirc some system do use 2 avx interchangeably

[fast pagoda]

each FMA unit does 1/cycle throughput and it is pipelined

[young plover]

Oh wait yeah

[fast pagoda]

intel should say 32/cycle somewhere in those docs

[true hemlock]

i don't believe they could do 0.5 CPI though

[fast pagoda]

as long as you have 4 fma om f;ogjt

[true hemlock]

just wouldn't be possible physically so i assume its 2 AVX being used interchangeably

[fast pagoda]

in flight

.5 cpi is the same thing as saying 2 fma units that each have 1/cycle throughput

[amber fractal]

I'd think you'd have enough register space to just have two separate loops running at once if needed

[fast pagoda]

that icelake core shows 1 cpi because it has 1 fma unit

[true hemlock]

latency is 4 cycles. i'd imagine it to do 8 in flight with 2 stream points per cycle

[true hemlock]

hmm

[fast pagoda]

it'd be 8 accumulators ye

if perfect

[amber fractal]

32 registers in total, so you get 2 operations per

[true hemlock]

so happy rn programming does not talk about llm slop and instead we talk about what god intended

[young plover]

Separate loops running on what? 🤔 The backend is saturated in this scenario so hyper threading does nothing.

[amber fractal]

seperate loops for streaming, no point in operating on a waiting state so one should swap to whatever is available at the moment

[true hemlock]

hyper threading often helps with giving instruction commands in case the other thread isn't saturating it well so it just really depends on the code

[fast pagoda]

from michael dell himself talking about this in cascade lake

[true hemlock]

32 DP/cycle god fuck

how many avx 512

[young plover]

Same throughput according to the instrinsic docs

[fast pagoda]

linpack chart

for this machine they talking about

https://www.dell.com/support/kbdoc/en-us/000176921/bios-characterization-for-hpc-with-intel-cascade-lake-processors

[true hemlock]

impressive.

[quick condor]

To be fair, the AI slop I'm working with actually involves a lot of non Ai specific code... Specially as it pertains to free threaded python and proper resource utilization

[true hemlock]

no shutup

adults are talking here

aight so apparently skylake has 2x 256 bit that acts as a single avx 512

[fast pagoda]

me trying to get hpl working while not paying attention

[true hemlock]

did it actually do 0.5 CPI wtf

because if it actually does 0.5 CPI with only technically one AVX512 i'd be fucking impressed

[young plover]

I mean that has to be one of the most optimized operations on the FPU so I guess it's not that surprising.

div has 4 CPI

[true hemlock]

dawg how is skylake losing hard to consumer zen 2 though

[fast pagoda]

is it skylake SP (xeon) or skylake X being referenced (x had 1 cpi)

[true hemlock]

oh

that makes way more sense i forgot xeon also has their own skylake

[fast pagoda]

skylake sp had 2 native avx 512

yea

[true hemlock]

gaddamn

[fast pagoda]

they fucking name their architectures like a gigachip in megaman battle network

[quick condor]

Because despite the fact that amd is killing it, the market share for Intel is still 72% where it matters

[amber fractal]

I don't know anything on this so I get to learn

[fast pagoda]

aawagga

[true hemlock]

at least you wanted to learn unlike the person above who just missed a point lmao

[fast pagoda]

oh damn

https://wccftech.com/amd-breaks-gpu-frequency-overclocking-world-record-by-pushing-radeon-rx-9060-xt-to-4-769-ghz/

[true hemlock]

long live epyc

[young plover]

14-bit approximate reciprocal is 1 CPI
No one needs FP32 precision these days anyways

[quick condor]

Yes still 72% with nothing relevant released. Keep in mind 28% isn't all amd

[amber fractal]

[true hemlock]

nothing relevant
some modern xeons actually being bangers

[quick condor]

Also for the most part amd system power draw deletes an entire cabinet in 1ru

[fast pagoda]

27% unit share but 41% revenue share 25 q2

[true hemlock]

where did you get this one from lmao

[amber fractal]

yeah sad that those are used nearly as much as H100 spam

[quick condor]

Actual datacenter experience?

[true hemlock]

i don't buy it

[fast pagoda]

intel is notoriously bad on power efficiency atm isnt it

like epyc are higher perf/watt

[quick condor]

I can dig up my equinix badge if I can find it

[true hemlock]

both can run at their designated TDP. both can also deletes the whole enterprise PSU brick with 2KW+ draw

hell in my experience of managing some sapphire rapids chips they draw more than my 96 core epyc 9655 often

the die density is just larger. physics don't lie.

just that also in my experience, intel's idle power draw/management are way better.

if that's what you're trying to talk about.

because depending on specific datacenter needs, that could be relevant.

[true hemlock]

this is true rn

[true hemlock]

but they're not that bad.

[quick condor]

Dual 1600w (1.6kw) that can run in non standby config, 3.2kw if running single phase is 88% (over 80%)

[true hemlock]

what system?

[fast pagoda]

we are approaching an historic moment

[amber fractal]

like you wouldn't use the same cpu for high prefomance compute 24/7 and running vms for web hosting on the same hardware. AMD makes sense in one and Intel in the other.

[fast pagoda]

surely 18A will turn it around

i'd like intel to turn it around because i own INTC

[young plover]

I had a lot of Intel shares and dumped them when the board fired

Was a wise choice

[fast pagoda]

skylake is nearly all of intel's marketshare

in server

it's like 50% of it

if you do not count 10 year old skylake and 7 year old cascade lake

AMD = 47.74%
Intel = 52.26%

[true hemlock]

you use the e waste ARM chips for VPS or vms or web hosts

[amber fractal]

That is also a fair statement

[quick condor]

😂 That's like most of what is real world deployed

[true hemlock]

also epyc 9965 192 core sweet spot tdp is at 600W ish

[fast pagoda]

well yes but the point is that when they retire

which they are

they will have like 0 share

[true hemlock]

and that's like almost 4x highest xeon core count

[fast pagoda]

and by 0 i mean 50%

but yeah

it's a lot different of a picture in more recent generations

that's all im saying

[amber fractal]

I stand that intel did manage to predict their needs well and try to mitigate it with products. But they were simply too early for the problem.

[fast pagoda]

intel's share is riding a geriatric coattail of enterprise sluggishness atm

[young plover]

🤤 Could compile my C++ project in under a minute.

[quick condor]

Oh 100%

[true hemlock]

i was really gonna hit on the 9965

till i got offered a new 9655 for like $4k ish

or $3k ish i forgot

[fast pagoda]

if you toss out the first 3 threadripper gens and cascade/skylake so only recent, amd has actually passed intel in workstation share

[true hemlock]

was bundled with the board, a H100 and A100

and the 12x64GB DDR5 ECC

[fast pagoda]

TR9K = 4.58%
TR7K = 7.05%
TR5K = 7.13%
TR3K = 8.73%
TR2K = 2.84%
TR1K = 5.82%
Thread ripper all up = 36.14%

Sapphire 35xx = 2.21%
Sapphire 34xx = 7.75%
Sapphire 25xx = 1.41%
Sapphire 24xx = 5.49%
Ice Lake 33xx = 1.80%
Cascades 32xx = 5.35%
Cascades 22xx = 17.59%
Skylake 31/29xx = 22.26%
Xeon W all up = 63.86%

Workstation component share throwing out TR 3 - 1 and Cascades and Skylake

AMD = 50.12%
Intel = 49.88%

[true hemlock]

i forgot the total but i can afford it with monthly payment last year

[fast pagoda]

anyways just fun stats from this wccf article and its comments mainly this one

[true hemlock]

the H100 and A100 are sxm

they somehow just have those sxm to pcie adapter

[fast pagoda]

i would like 18a aka 1.8 nm if using older style to be good but who knows

[true hemlock]

i still find it fun that epyc I/O die is still gigantic

[young plover]

wccftech comments
Haven't looked at that toxic dump in a couple years

[fast pagoda]

it's so bad if you scroll down lmfao

this reminds me of liek 15 years ago:

[young plover]

:dentge:

[fast pagoda]

good ol team sport flamewar

[true hemlock]

considering that at least half my uni uses macbook im surprised this is what the graph is rn

[young plover]

1.8 NM DRAM and NAND
What could go wrong

[fast pagoda]

apple people arent real

[true hemlock]

though desktop do really took a major chunk because offices uses a lot of them compared to laptops

[amber fractal]

I can't say what I think they should be making, but a few chatters know exactly what I'm thinking

[fast pagoda]

i bet openclaw minis are somehow larger in volume than total apple desktop sales for like years

(joking but)

[true hemlock]

i thought those are 9 y.o that watch pc building slop shorts everday

[fast pagoda]

worse

49 year olds

[true hemlock]

30+ y.o be talking about how their M4 draws only 10W when x86 chips has to draw hundreds

[young plover]

How many were bought with credit cards, gambling that their botnet could somehow make the money back.

[fast pagoda]

those posts reek of genx

[fast pagoda]

i hope not many considering theyre so cheap

i mean you should buy them with a cc to get your 5% cashback but then you pay it off that month otherwise you pay interest which is a life altering failure

[young plover]

Oh that's how I use mine, I even pay off the balance before the billing date. Seems to artificially inflate my credit score lol

[true hemlock]

i love my interest free pay in 4 services.

also wtf are openclaw mini

[silent island]

dawgg, i never realized most of my switch roms are leaked copies lmaooo

[fast pagoda]

people are hoovering up mac minis because you can get an m4 or m4 pro for around 500 bux at the cheapest

[true hemlock]

expensive.

[fast pagoda]

not really for what it offers when you look at it because it's not about 1 unit

cluster them

[true hemlock]

still expensive

but i suppose for apple bros, cheap

[fast pagoda]

it's not that expensive to be able to run kimi k2

(1T)

[quick condor]

You mean offbrand claude

[fast pagoda]

it's like the one time an apple pc has ever been a good deal for anything

other than smelling one's own farts

at least in modern times

[quick condor]

Apple's hardware is actually (for the most part) not awfully priced (ignoring the top end graphics design stuff).

And I hate to admit it

[fast pagoda]

it's awfully priced if you care about actual specs/dollar

it is not low quality equipment but it's overpriced for what they are giving you in sillycon

minis being 599 is the best deal they have by a lot

[quick condor]

If you spec out a Dell laptop with similar specs it's less than $300 cheaper

[fast pagoda]

that's a large amount of money for many people first of all, secondly i don't use dell as a benchmark for value and neither should really anyone, third as soon as you start adding small amounts of ram and storage suddenly any apple machine is 2x in price immediately

[quick condor]

Dell is by far the standard for large scale deployment.

[true hemlock]

e waste

[fast pagoda]

talking enterprise deployments is irrelevant to the average consumer and has different dynamics completely but even still 300/unit is a 6 million dollar difference when you multiply that by 20k machines

apple is charging $1080 to go from 48gb ram to 128gb on a macbook pro, which sadly doesn't sound insane right now because of the world going insane, but it was the same price before everything went insane, or very similar

[quick condor]

300 a unit on 2000 units

[fast pagoda]

oh that's a small deployment

but still a large difference

[quick condor]

It's not when you factor in what you are paying in salary to an employee that you are giving a $2000 laptop

[fast pagoda]

600k on a 3-5m contract is a large difference

just hardware cost

and the employee wont care either way unless theyre a creative who probably should have a mac because of ecosystem shit and to stop them moaning

[quick condor]

Same with upper level IT, you give them what makes them happy enough to not complain about the pay

[fast pagoda]

upper level IT requesting a mac

where are you so i can stay as far away as possible

[true hemlock]

specs?

[quick condor]

It depends on the person but I know plenty who prefer it. Doesn't really matter what your running once your in a shell anyway

[fast pagoda]

which is an argument against an apple machine not for

[true hemlock]

doesn't really matter
$2000

[fast pagoda]

i would use a macbook pro with linux or something on it

im not saying it's a useless piece of rock

[true hemlock]

is this how it is in the enterprise space managed by a moron who has infinite budget

[fast pagoda]

kinda sometimes

[quick condor]

No it's a finite budget spent on stupid things

[true hemlock]

even my personal briefcase pc that is faster than any gaming laptops out there is still less than $2K

[amber fractal]

convince is worth a lot, and people are willing to pay quite a bit to be lazy.

[true hemlock]

it has 128GB ddr4 too

[quick condor]

I'd think something is wrong if it was allocated in a normal way

[fast pagoda]

i cant think of anything convenient about using an apple machine in an AD managed environment

or in general but especially then

[amber fractal]

I ain't have the income for convenience, good to learn how the tech works as well

[true hemlock]

eh work laptops are inherently overpriced anyways

[amber fractal]

I'm pretty much done with buying laptops, I finally have a room to myself so I can actually invest in desktops properly

or odds are servers

don't really need much more gaming prefomance

[fast pagoda]

the standard shitter laptop people get deployed at my office is basically a base model dell 16 plus

they probably run around 500 bux/unit

for that apple will sell you most of a mac mini or optionally they will take a shit on your doorstep

it is not free of charge for that service unfortunately

[amber fractal]

that requires a service tip tho

same braincell

[fast pagoda]

lolol

gotta pay the toll to see such a display

common sense really

[quick condor]

I mean if people will pay you to do it then why would you do it for free? That's not apple's fault now is it?

[fast pagoda]

well, it IS their fault that they make you pay extra for post-purchase doorstep steamer support

[orchid mesa]

hi gugs.... if u had to guess, how does neuro listen to ppl?

like... her ability to listen to vedal/collabs is via discord right? so does that mean she's (atleast part of her) is a discord bot?

like if thats true, does vedal use pycord or something like that?

[obsidian mantle]

She is agi she can speak english directly

Her speech to text connects to discord and transforms speech into text for her to use

[orchid mesa]

yea... but like what does he use

[lean stream]

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

[obsidian mantle]

@round kraken

[obsidian mantle]

Isnt marked as bot or maybe its because im on phone

[lean stream]

One message removed from a suspended account.

One message removed from a suspended account.

[obsidian mantle]

Well

Shes not a bot so makes sense

[orchid mesa]

is that the bot he uses?

i thought it was this one

[lean stream]

One message removed from a suspended account.

[obsidian mantle]

Its her account

[lean stream]

One message removed from a suspended account.

[orchid mesa]

@old pendant

i thought hes using htis?

[obsidian mantle]

Wait

Yes

[orchid mesa]

try "in" instead of ==

[lean stream]

One message removed from a suspended account.

[orchid mesa]

anyways:
i found this (pycord's audio examples)
https://github.com/Pycord-Development/pycord/blob/master/examples/audio_recording.py
the thing about this one tho is that you have to manually start/stop to listen audio

[lean stream]

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

One message removed from a suspended account.

[orchid mesa]

whats ur theory on how neuro listens to ppl?

[mighty thorn]

Twins role is crazy

[uneven pulsar]

Hi nerds

[orchid mesa]

i checked few, discord.py (the one im using, doesnt support voice recieve), pycord/discord.js/discord.py-ext does, ig my question is how does she listen to the speaker?

like
user speaks -> bot recieves message -> user stops speaking -> (threshhold for silence) -> bot saves message -> then that goes to sst??

but neuro sometimes interrupts vedal? is that because of the delay from the previous voice read?

[obsidian mantle]

Nobody knows

Some people think its actually a human called tony

[mighty thorn]

Probably either clever use of discord.py-ext OR, more likely, the real Neuro is a self bot

[orchid mesa]

he wrote his own framework?

[trim valve]

why would being a self bot change anything

[mighty thorn]

Much easier audio stuff
-# legally I wouldn’t know

[trim valve]

it's literally not easier

[orchid mesa]

i haven't looked up much more on discord bots anyways (planning to), but focusing on this, you have to manually trigger it to save audio recordings

[trim valve]

ok at a super low level, sure you could use all of the discord client as a source

[mighty thorn]

discord.py-self

[trim valve]

if you invest more time an infinitely better solution is just using a bot, because they can get individual audio tracks

[orchid mesa]

oh thats a thing lmao

[trim valve]

if I wasn't increasingly busy with life and other projects I'd work on a program for stupidly easy voice recv because for some reason everyone loves to do it the worst ways possible

[mighty thorn]

The only times I’ve done it, it has ended in code tragedy

[trim valve]

skill issue?

[mighty thorn]

Yes

[orchid mesa]

heres a stupidly easy one
https://github.com/Pycord-Development/pycord/blob/master/examples/audio_recording.py

[glass jetty]

https://github.com/XTLS/Xray-core/discussions/5724

hi chat

long time no discord

[trim valve]

eh, I have an aversion to pycord

[orchid mesa]

what would you use thne?

[trim valve]

unfortunately afaik for receiving voice in python I think it's the only option

[orchid mesa]

what would you write it on then?

[trim valve]

I've previously done it in rust and it's worked very well

[opaque wharf]

What's the purpose of Neuro-bot on this discord?

Other than feeding it cookie of course

[stiff mica]

🍪 Om nom nom
You've given me 4 cookies! | I've received 284656 cookies total!

[trim valve]

keep in mind this is a "the developers have been bad people in the past, and I don't trust a lot of the architectural decisions they've made"

you can 100% write something with pycord and have it work

[mighty thorn]

I did mine in discord.py

[stiff mica]

🍪 Om nom nom
You've given me 8 cookies! | I've received 284658 cookies total!

[ionic lintel]

Hey

[stiff mica]

🍪 Om nom nom
You've given me 1 cookie! | I've received 284662 cookies total!

[ionic lintel]

Yeaaahhh

[stiff mica]

🍪 Om nom nom
You've given me 1 cookie! | I've received 284663 cookies total!

[small anvil]

havent spoken in here for a while, but been making some cool stuff that i think would be cool to show
been working on a custom unity IL2CPP modloader for a little project me and my friends are working on and its soo cool, although im not certain about how the plugin infrastructure will look like. i've been taking inspo from geode and im trying to not stray too far off that path

if anyone could give me some feedback on how i could make the syntax for plugins a bit better without entirely stealing the look of geode pls lmk

[hallow dagger]

@small anvil

i am interested

https://cdn.discordapp.com/attachments/1061113840852533339/1456314197993459712/SPOILER_NeuroNewYearCute5.gif

idk much about loaders but i see il2cpp
i'm writing my own il2cpp dumper
for honkai star rail

hiiiiiiiiiii

[small anvil]

loaders are pretty simple, you just gotta be able to load coreclr into the process to be able to execute IL

[hallow dagger]

interesting...

is the one you're making universal or only working on this game so far? or only targeting one/few games specifically

[small anvil]

literally all the logic that goes on in finding an available CLR environment
ill probably switch this to downloading the binary itself soon but for now this is just a quick fix

[small anvil]

universal, hopefully

im gonna have support for both managed and unmanaged plugins

[hallow dagger]

[small anvil]

its a very daunting process

i decided to make my common library header only

and now im having to split it into source files

cause my dlls are reaching sizes they should never reach

[hallow dagger]

oh uhm
how are you injecting it

[small anvil]

proxy dll

unity likes to load the first version.dll or winhttp.dll they see

just

put it into game folder and run

basically

[hallow dagger]

omg i'm doing the same thing

replacing mhypbase.dll in the game folder
game runs my code blindly
boom we're in

[small anvil]

oh nice

i dont know much about honkai so i wouldnt know the process

i heard the people that make that game likes to hide their code religiously

[hallow dagger]

oh yeah hoyoverse is extremely aggressive

but hsr is their easiest game to work with

genshin is wayyy harder
and zzz is even more difficult

[small anvil]

[small anvil]

oh yeah i saw genshin i think

i first found out from perfare i think their name is

from their blog

[hallow dagger]

[small anvil]

wait let me try to find it

[hallow dagger]

yes plz
that's the person that wrote the famous il2cppdumper right
but that one works by parsing metadata
which is not possible here
encrypted

[small anvil]

blog is down it seems

i know im not schizo i wear thats were the blog was

archive.org my beloved here i come

[hallow dagger]

@small anvil hi hi u can read down my messages from here if you are interested which of course you are teehee.......

[small anvil]

oh thats really cool

i havent seen an actual runtime based il2cpp dumper in ages

only time i saw one mentioned was a while ago when i was reverse engineering a social vr game with a couple of friends

[hallow dagger]

well it's definitely much easier than working statically with encrypted metadata

[small anvil]

i have too much experience with that..

[hallow dagger]

luluscared asf

[small anvil]

[hallow dagger]

[small anvil]

and as far as i know this get decrypted somewhere after il2cpp_init gets invoked

[hallow dagger]

yes

but

with this game even that isn't possible

the metadata never exists fully decrypted in memory

game uses lazy decryption
only decrypts the parts it needs, when it needs them

so yeah...

[small anvil]

oh boy

[hallow dagger]

yeah hoyoverse is very seriously anti RE

[small anvil]

yeah i can tell

[hallow dagger]

of course that doesn't stop the awesome RE community
it's a cat and mouse game after all
but
it does mean that pisslows like me get a much, MUCH steeper learning curve

[small anvil]

it always is

i dont think im gonna even bother with splitting up my common library

ill probably do that later

[real sierra]

https://x.com/i/status/2026282714442399765

Pirat_Nation 🔴 (@Pirat_Nation)

Twitch now requires some new Affiliates to verify their identity through Persona before receiving their first payout.

Users report needing to submit a government-issued photo ID along with a selfie, with no alternative verification method offered, according to Twitch support.

X · Feb 24, 2026, 1:07 PM

is this going to become the norm for online services

[small anvil]

i hope not

[real sierra]

I don't know where the push for this shift is coming from

[small anvil]

big tech corporations bending for governments

uk pushed it the most

and everyone else followed

[maiden geyser]

filtered.

[hallow dagger]

[olive sable]

Goodmorning

[opaque sigil]

good evening

or late afternoon i guess

[hard raptor]

i got the GPU

[cosmic sphinx]

we can't discuss those topics here

[real sierra]

mb

[opaque wharf]

It's almost midnight

[maiden briar]

yo any data engineers in chat, specifically anybody that works with SQL

[opaque wharf]

Hey Shiro, quick question. What is the function of Neuro-bot on this server?

[opaque wharf]

Just ask your question

[rigid snow]

dougdoug guy sql

[maiden briar]

okay, say I have a DECIMAL(38, 2) in Teradata and I want to convert it to BigQuery. Should it become a NUMERIC(38, 2) or a BIGNUMERIC(38, 2)?

[opaque wharf]

If there's no plan to add more digits it shouldn't matter which one you choose

[maiden briar]

well it does for performance

[maiden briar]

[opaque wharf]

The only way to know is to benchmark it. Because it all depends on what you do. There's a benefit of numeric vs big numeric, namely the smaller storage requirement

[opaque wharf]

*For the same supported precision

[maiden briar]

reason I ask is because I made a python script that compares schema, and someone is telling me that NUMERIC can't take more than 29 which i was almost certain is not the case.

[olive sable]

[tender river]

dougdoug guy

[opaque wharf]

According to the docs, it can have 38 digits with 9 of them being the number behind the decimal point

I don't know if the limit of 38 digits can be said the same when using only 2 decimal

[real sierra]

uhh

[stiff mica]

🍪 Om nom nom
You've given me 19 cookies! | I've received 284670 cookies total!

[rough bloom]

https://docs.cloud.google.com/bigquery/docs/reference/standard-sql/data-types#parameterized_decimal_type
max precision of scale + 29

so 31 in this case

need BIGNUMERIC

[olive sable]

main job of neurobot is
trust me

[stiff mica]

🍪 Om nom nom
You've given me 2 cookies! | I've received 284671 cookies total!

[maiden geyser]

btw didn't know unkomputoble is the starknight

[opaque wharf]

So their 38 precision is only valid with 9 digit of decimal? As in, you can only go up to 10^29?

[opaque wharf]

You just realized that we are on 2 different server?

[maiden briar]

what the heck

oh, so NUMERIC(38, 9) is valid because there would be 29 digits to the left of the decimal point. So I should be checking for if (p - s > 29).

[sage crag]

rule 3

[maiden geyser]

rule 2 thoughbeitever

[olive sable]

2nd rma got approved

[nocturne olive]

RMA for what exactly?

[olive sable]

it says it on the picture no?

its my ram

[nocturne olive]

Oh I see

It broke again?

Are you sure the issue can't be your CPU's memory controller?

[olive sable]

nope

its the stick

3 of them work perfectly fine, the 4th one doesnt even work on 2133mhz

or well, it does work but i get a lot of errors during memtest

linux boots but is unstable

[silent cloak]

for all the linux users here

[opaque wharf]

Linus is right. ECC is de wae

[olive sable]

broken nand chip wont be solved via ecc

[opaque wharf]

But you'll know that it is the RAM that is faulty

[olive sable]

i already know via memtest

all the other sticks had no issues, jsut the one bad stick

[olive sable]

ive been saved form bankruptcy

742 euro

[silent cloak]

probably the worst time to get so much ram

[opaque sigil]

rgb ram no less

[silent cloak]

one could say probably one of the worst times in the history of ram to make such a purchase

[olive sable]

its RMA, im not buying

if my rma got denied i would have kept using this 16Gb stick for a year or 2

actually, i could jsut use 2 of the working sticks

[sick owl]

Gemini 3.1 way overthinking its review of my CPU fan curve

[olive sable]

now that i think about it

[opaque sigil]

sounds like something gemini would do

[olive sable]

with a 16gb stick, i have more vram than regular ram rn

[shell dome]

Currently developing my own friendly ai... thing. Hope that she and neuro will be friends one day lol

[olive sable]

i highly doubt your ai will interact much with neuro

[shell dome]

omg I just typed python to fast and typed puyjom instead >w<

[warped narwhal]

w<

chat what does this mean?

[leaden crest]

uwu-esque thing idk

[warped narwhal]

$10 a month for a minecraft server?

and why 3vcpus if minecraft servers are single threaded?

[olive sable]

the > and < represent squinting eyes. the W represnets a curved mouth

[leaden crest]

oh god it’s in their status too

[olive sable]

their status also has cat ears

^>w<^

[leaden crest]

catgirl “>w<“ lol

[warped narwhal]

kill it with fire

🔥

[olive sable]

[leaden crest]

whoa you think uwu behavior is fire

[olive sable]

its kinda cringe but ill allow it

[nocturne olive]

Chunk gen

Also they're likely not very good threads

Probably AMD Epyc threads

Which are really slow but you get a ton of them

[leaden crest]

I think the #programming channel alone has enough people to make neuro’s dream of an ai-only tinder come true

[warped narwhal]

to be fair epyc threads can be pretty fast if it is the right sku

[olive sable]

i dont think vedal would let anything close to that happen tho

[leaden crest]

fr w father

[shell dome]

Looks like ^>w<^ has sparked some attentio UwU

[sage crag]

https://cdn.discordapp.com/emojis/1360837088379015204.webp?size=48&name=minyanlookup&lossless=true

dentge

[olive sable]

hi konii

[sick owl]

https://huggingface.co/ServiceNow-AI/Apriel-1.6-15b-Thinker

This model looks really good, why have I not heard anything about it

Its not as though ServiceNow is a small unknown company

[cosmic sphinx]

is our free ai slop generator finally back

https://cdn.discordapp.com/attachments/1072697081443131476/1377029839038582956/xviangel-clap.gif

[sick owl]

Then again if it doesn't surpass 3 pro image (which seems likely given the lower param count) its not that exciting for me given I have free pro as a student

[real sierra]

ive discovered something horrible

literally ruined my night

nevermind

false alarm

user error

[olive sable]

explain yourself shiro

[warped narwhal]

issue between keyboard and chair?

[real sierra]

☁️☁️☁️

i was looking at this message

and had totally convinced myself the cloud emoji was smaller than the lookup emoji

but its just an illusion i guess?

[olive sable]

it is smaller i think

[stray dragon]

yeah discord likes to put spaces after your emojis

[olive sable]

lemme try something

[real sierra]

thats what i thought too

but it isnt

[stray dragon]

the clouds have them removed, but there's a space in there

see

[olive sable]

oh

[stray dragon]

without the space it's all the same size

ignore that it's red

[olive sable]

☁️☁️☁️
:Empty:

oh

[real sierra]

idk why the space is so hard to see

[olive sable]

i dont have the empty emote

[real sierra]

?

?

[warped narwhal]

☁️☁️☁️

[real sierra]

do you have either of those

[olive sable]

nope

[real sierra]

i guess you're more likely to have or

[stray dragon]

don't worry shiro all the emojis are the same size

[warped narwhal]

no but I have

[stray dragon]

unless you consider them getting downsized if there's text in the message

or many emojis in the message

or formatting effects

[olive sable]

nope

shiro we're friends, you can see our mutual servers

[warped narwhal]

you don't need to be friends to see that

[olive sable]

oh true

im mixing it up with mutual friends

[olive sable]

only 3, thats crazy

[warped narwhal]

I'm only in 3 vtubing servers, so unless you play a lot of squad then it's unlikely that we'd share any

[olive sable]

the vtube studio server is public so i can see where it came from,t he rest is just from a private server

[warped narwhal]

I've never understood those that are in hundreds of servers, like there is absolutely no way you can pay attention to even a small fraction of that

all my servers fit on one screen

[olive sable]

im only active in maybe 7 of the servers im in

i got dev servers, vtuber servers, servers form friends, college servers, game servers

[warped narwhal]

I have my own dev server, a server for a squad community I admin for, some servers for other squad communities (including the official ones), some vtuber servers, then some miscellaneous servers I use for emojies

[silent island]

Dawgggg DS2 demanding af
Idk if my LOQ can handle it ngl

[unkempt citrus]

150

[fickle rain]

https://lyra.horse/x86css/

Thank you Rebane

[hallow dagger]

wtf

[young oyster]

But can it run Linux?

[hallow dagger]

asking the real questions here

[opaque wharf]

Movuscator exists. So it's only time until someone made C compiler to css

[young oyster]

I remember playing with a javascript x86 emulator running Linux.

Without looking over the source code, im guessing this css emulator makes heavy use of calc, guessing that set of squares at the bottom is memory with each box being a byte encoded in the shade of gray.

There is no reason the box holding data needs to be visible...AFAIK calc will work just fine, so that can be a good way for handling storage needs.

[young oyster]

So long as the css is working as an x86 emulator, you shouldnt need any special compiling...it should run x86 code

[opaque wharf]

But that would be slow since it needs to emulate x86

[young oyster]

Yeah...but also its Linux...you just need to get to command line.

No one saying it should run KDE

OK, having seen the demo...never mind

That is stupidly slow

[mighty thorn]

Skill issue

[olive sable]

what exactly is the skill that they have issue with?

[mighty thorn]

Lack of mutual servers

[olive sable]

7 is rookie numbers

[amber fractal]

8 here

[olive sable]

hi iggly

[amber fractal]

[hallow dagger]

me who has like 70 mutual servers with my bestie

[leaden crest]

70???

[olive sable]

70 is a bit excessive

[hallow dagger]

i've been fighting 200 server limit for months now

[leaden crest]

the server limit is 200?

how does one even reach that level

[hallow dagger]

well

100 normally

200 with nitro

[leaden crest]

if you lose nitro do your servers get sent to purgatory

[hallow dagger]

if you're above 100 and you lose nitro they don't go away but you can't join servers anymore unless you leave enough and go back under the 100 limit (or buy nitro again)

[leaden crest]

is like 90% of your servers just emoji servers

[hallow dagger]

yes

[olive sable]

my highest is maybe 13 mutual servers, idk

[hallow dagger]

i only really talk in one lol

close friends server

i used to be more public and stuff
some years ago
but that shit drains you

now i hate people

[leaden crest]

I have a 19 mutual server friend

lemme see if there’s someone else

nope 19’s my max

[opaque wharf]

I don't have that many friends on discord to begin with

So my highest is 2

[queen ginkgo]

I'm in sixteen servers period. Seven are private, three of the public ones have less than five hundred members.

In the words of my friend, I can be a bit of a boring arse

[hallow dagger]

hi

[quick condor]

I have my active group, irl and ones I look at from time to time/not at all

[true hemlock]

wait

even more

[amber fractal]

3 main servers, 5 friend servers, 6 announcement servers, 3 emote servers, 4 dead servers that just exist.

That's also how I organize them

[azure lynx]

i left all the servers i was in that i couldn't rememeber why i went to them. and then i had a managable amount.

(I also don't have a great memory)

[quick condor]

I think I have a problem... I'm attempting to do an all core compile of flash attention... On a 72 core system

[kind nimbus]

Jesus Christ i'm always afraid of my phone's battery when i join too many discord servers lol

[snow creek]

Chat is flutter great for developing a p2p multiplayer game

I want to make an Android game as a hobby

[quick condor]

normal flash attention build

gotta be honest, building flash attention in under 8 minutes felt pretty good

[silent island]

btw for anyone who owns a console, can we turn off motion blur in games?

[amber fractal]

That is called Temporal Anti Aliasing, and no; your stuck with it

[rigid snow]

flutter is generally meh for games, it's built for making apps - though there is stuff like flame

one thing flutter has going for it is you have good cross-platform out of the box, so it'll be a ios, windows, macos, linux game as well if you just build it for that platform

[silent island]

bru imagine looking at 30 fps with motion blur...
painful ☠️

[amber fractal]

Your also stuck with it on PC

[rigid snow]

wait what

that's like

wrong

two different things

[amber fractal]

Yeah honestly it isn't the same thing

[rigid snow]

[amber fractal]

One can get rid of motion blur, but that aliasing is motion based

That's on me for not explaining everything that was in my head at the time

[rigid snow]

taa isn't remotely as offensive as motion blur and vice versa if i actually want motion blur

motion blur can be nice sometimes

[amber fractal]

It can be, I do agree. I just wish to have an option to not have any blur

[rigid snow]

it isn't either or, it's two separate things

[amber fractal]

I'm aware of that

I just dislike TAA as a forced option

[rigid snow]

tbf i've yet to see a game where i can't turn it off

but also i never do anyway it looks fine

[amber fractal]

Makes sense

[sage crag]

bird

[rigid snow]

if i have a dlss option i'll turn that on

[sage crag]

did you know there are some butterflies in this genus with a wingspan of 20cm

which is larger than some songbirds

[opaque wharf]

What a peculiar species of bird

[sage crag]

[rigid snow]

does not make it a brid

[sage crag]

wrong

it has wings

[opaque wharf]

Bird

[sage crag]

and a mouth

therefore its a bird

[quick condor]

IMO motion blur when it's not used to hide low frame rate is fine. Problem is when it's used for something it shouldn't be (hiding low frame rate)

[proven merlin]

I like motion blur (when done properly, there are different kinds of motion blur, radial vs linear movement and such) in racing games. However only on low intensity. For everything else i dont particularly like it

I find TAA quite nice, but it needs post-sharpening, only a touch, but yea. The one issue i have with "It takes Two" is that the TAA is not sharpened afterwards so i have to do it in the driver which probably doesn't look as good as it could.

(i add like 12% sharpening in nvidia or so. 7 and up is enough, 25 and above looks odd to me. 50% looks stylized cartoonish to me

[glass jetty]

ahahaha

just check this out

that so funny

auto lock = save_buffer.lock_nowait();
while(not lock.has_value()){
  co_yield{.timeout = 40};
  lock = save_buffer.lock_nowait();
}

auto success = save_buffer.write_ifFreeSpace();
if(not success){
  co_return;
}```

[tender river]

meganeur

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

wing span 70cm

insect

[sick owl]

Apparently Google models know what they are now

That's new

[glass jetty]

cloude 4.6 thinking is peak

idk about prev, havent used much

but out of everything i used it is most useful model

AND NOT ONLY FOR SEARCH

it just wrote me lua script that ACTUALLY WORKS

[sick owl]

Gemini 3.1 Pro is my go to now but Opus 4.6 is great too

[glass jetty]

i just checked everything in script following docs and other examples, and it is correct

[sick owl]

I hate this fedora wearing reddit fuck so much

[tender river]

im looking at various large insects and i really really would not like to meet any of them
butterflies look fine by comparison but 20cm

[sick owl]

Due to price/performance mind, Opus 4.6 is a stronger coding model while 3.1 Pro is a stronger generalist but 3.1 Pro is good enough at coding for the price

[stray dragon]

amount of false positives this thing gives is awful

[rigid snow]

system prompt change

[sick owl]

I have a 100% false positive rate with it because I'm on Mullvad

[stray dragon]

i don't even use anything like that

[sick owl]

Makes sense

[stray dragon]

it just blocks me for no reason

default browser no incognito mode no vpn no weird network shenanigans besides being on uni wifi

"you've been blocked by network security" i'm about to block someone from continued existence

[rigid snow]

literal data center ip btw

no clue what they base it off

[stray dragon]

they base it off "fuck this random user today"

it tricks the shareholders

[sick owl]

https://www.bbc.com/news/articles/cjrq1vwe73po

Disconcerting

Apparently Anthropics redlines are using Claude for mass surveillance and autonomous weaponry

But the Pentagon is pushing for it

[sage crag]

orm

llm guided weapon

im reminded of

the vending machine

are they sure its a good idea

[tender river]

they dont really need perfect accuracy or decision making

is all i'll say

[jagged turtle]

what vending machine?

[tender river]

vending bench

https://arxiv.org/pdf/2502.15840

[jagged turtle]

ah that

uh

it just needs Good Enough™ accuracy

[sage crag]

llm scapegoat :rrrr:

[midnight sigil]

Gigabyte G24F

This symptom(random lines appear and screen jittering) happens every time I turn the monitor on(even on its logo page, so not gpu issue).

It'll slowly recover after a certain amount of time(anywhere from 4secs to 6mins) after lighting up, the artifact disappears like something scanning upwards. Remains functional without any other visual artifacts.

Initially occurred after OCing the monitor from 165Hz to 180Hz(their promoted overclocking refresh rate)

My questions:
Would it worsen in the long term?
Is it fine if I just leave it untouched and keep on using it?

[brisk laurel]

Hmmm

[sick owl]

Drop the overclock, if it causes instability its not worth it for 15 extra Hz

You can damage monitor hardware with a dodgy OC

[midnight sigil]

the artifact still appears after dropping the OC and returning to its original working refresh rate, maybe it fucked up already

[sick owl]

↓

This is why I don't fuck with monitor overclocking, you risk hardware damage for (in this case) 0.5ms of frametime improvement

[midnight sigil]

well I guess it's fine if I just have to let it warm up for some minutes

at least not some huge visual issues that stay consistently

[azure lynx]

probably some component got stressed at some point and now is acting out.
(as in "is incapable of performing in spec any more")

i made an LCD screen have grayscale by screwing with the driver for it... but that only cost a couple of dollars vs a monitor. and it looked pretty bad.

[midnight sigil]

I am suspecting some thermal-sensitive capacitors got a bit cooked while I was OCing

only 2 years of use

[small anvil]

rip

[true hemlock]

did you by any chance also touched some voltage related shit

because that's the one probably causing hardware damages
else just normal overclocking shouldn't actually damage it and i suspected its just buffer related issues and you can try disconnecting it from any power

[midnight sigil]

nope, I used the overclocking feature built-in to the OSD

just a toggle from no to yes, and perhaps also switching my refresh rate in the nvidia control panel

[true hemlock]

idk how does gigabyte handle the overclocking

[true hemlock]

welp

assuming its gigabyte

[ornate bane]

interesting read, might shed some light on those recent moves in the hardware market, such as the hoarding of wafers by the biggies..
https://www.nytimes.com/2026/02/24/technology/taiwan-china-chips-silicon-valley-tsmc.html

[midnight sigil]

[true hemlock]

everything everything ai now ig

im just gonna silently watch as these "ai" models slowly just fucks up by some sort of any unlucky probability distribution and cause them damages

we have the us and openai hoarding 900K dram wafers per month and causing the shortage

that 900K wafer is about at least 1.4EB on reasonable yield rate. not even any supercomputer with most system ram came close to that capacity

and the fact that its strictly dram and not even hbm

its questionable as fuck
all those ram are just going to sit in warehouse and waste away

[stray dragon]

has this not been blindingly obvious for years, even before ai

[true hemlock]

god, honestly i would really like to see what happened if the chip supply were cut off from the US because that'd be funny as fuck

we're doomed either way. either they get it and waste it in the name of "AI"

or they don't get it at all

[stray dragon]

did they not just pre-order tons of the memory wafers and buy out most of the scheduled future supply for the next few years

it's not even sitting in warehouses it simply doesn't exist

[jovial remnant]

they reserved wafers on terms that when they are available in future, then they can choose what they will be made into from some predetermined options

[stray dragon]

i see, thank you

[true hemlock]

the contract said 900K wafers per month

so 40% of what's currently being produced

[stray dragon]

christ

[true hemlock]

its still very unrealistic

i do not see how they could even utilize that much

let alone that much per month

[sage crag]

t

https://tenor.com/view/kitten-hit-kitten-hitting-kitten-hitting-camera-kitten-slapping-cat-hit-gif-5248570956277630634

[stray dragon]

yeah seems like they'd be limited by infrastructure way more

power, facilities, etc

[true hemlock]

nodes.

[stray dragon]

hi konii

[sage crag]

https://cdn.discordapp.com/attachments/1421861060507602944/1425936142754123927/neuroCatModeOfOfnnfOOf.webp

[true hemlock]

no one cram more than 1536GB on even a modern 12 channel epyc

else you run at way lower bandwidth

even if they managed 2048GB per CPU

[obsidian mantle]

It always seemed crazy to me that they have not moved tsmc somewhere else just in case

[true hemlock]

its still just 65PB of ram on an extremely massive 16K dual socket node

[stray dragon]

i thought the high-bandwidth memory wafers were also used for VRAM as well as DRAM DDR5 and it was going mostly into the gpus rather than memory sticks

[true hemlock]

because they don't own tsmc

[true hemlock]

contract specified strictly dram

[stray dragon]

oooh ok that's news to me

[true hemlock]

HBM isn't even included yet on their contract

i don't fucking know why 900K wafers of dram

they could've do like 10x less capacity for like hbm with similar cost

and get even more benefit because hbm is just way better for whatever the fuck ai shit they do

you don't inference nor especially train models on dram.

great lol, all of it is getting wasted

[azure lynx]

the T stands for Taiwan.

[true hemlock]

either that or open ai would start drop shipping dram when they fold

[azure lynx]

also: moving fabs is very hard.

[true hemlock]

you literally would have to rebuild the fab and sorta start all over the R&D

[azure lynx]

i'm almost certain there'll be a video on the asianometry channel about the article in the next week or two.

[olive sable]

[azure lynx]

damnit. i wrote it but still read it as fabs() instead of "fabrication plants".

[midnight sigil]

new discovery, also happens when switching to a dedicated fullscreen with different resolutions

[glass jetty]

i used same monitor for 7 years
buttons almost not work, but monitor is fine

[azure lynx]

might be time to look for a replacement.

[sick owl]

https://huggingface.co/Qwen/Qwen3.5-27B

[glass jetty]

i gifted pc, so now i just use my small laptop

[sick owl]

This may be the GPT OSS killer

[glass jetty]

i like have to write everything i designed, but i tired now
randomly found critical memory leak on coroutines stack destruction when exception occur
i just have to put code i wrote in actual repo (in 3 different generations of our software)... a lot of time just for copy pasting and running compilation + fixing compilation issues

• i have to write logic for writing and reading from disk, to save actual output data and will use disk as swap memory

also, who need it