the more you buy the more you save
#off-topic-tech
verbal raft
languid gulch
so when do we get to SLI 2x 5070s together
twin dew
Never?
languid gulch
π
willow pike
sli ded
twin dew
Implicit multi-GPU died with DX12.
With DX12, devs could code explicit multi-gpu, but no-one has bothered.
willow pike
explicit multi-gpu never happened in any real way
twin dew
Meant SLI and Crossfire with implicit in this case.
Which cannot be used with anything after DX11.
willow pike
even if it could, good luck getting anything running on it that uses the previous framebuffer
like dlss
languid gulch
i guess my sarcasm wasn't very clear
languid gulch
wow, palisades lost their high school
verbal raft
languid gulch
with a variable number of incorrect fingers added in
languid gulch
how about the fact that when you see a deep discount from a manufacturer, they're still pulling a profit on it, just not as much of one, which means they could have started off at that lower price & still been in positive territory
expected by investors should be a middle finger from the consumer until it's at a sane price
so, greed, & not "it's expensive to make"
there's markup for cost, then there's markup for profit. 2 very different things
edgy hazel
It's expensive to slap all these AI chips on gaming GPUs
languid gulch
i mean, they're very welcome to make something that costs them less to manufacture
just fire all the staff working on that & let AI take care of it if they're so confident that'll work
sounds like they're just slow to invest. not the customer's problem
sorry, i have extremely little sympathy for any issues the most valuable company on earth might have
soft bloom
Bruh Yt short gave me some vid of him talking in some university, and it was such a ... Crapload
Too lazy to find from history
twin dew
Broadcom isn't Broadcom anymore.
Broadcom was bought out carved out and the buying entity just renamed itself to Broadcom again.
Current Broadcom is old Avago, which got spun out of HP long time ago.
Well, with owners like this, what do you expect?
https://en.wikipedia.org/wiki/Broadcom#Ownership
The whole VMWare fiasco is just the highest profile thing that has been happening with the "New" Broadcom
Coasting along with minimal input.
Basically current Broadcom is another dying large company, being killed by financial people with underinvestment and the VMWare change type idiocy.
Yes.
With the outing of Pat, it seems that Intel will not be coming back.
Pat wasn't perfect, but he actually seemed to try to claw back.
And was ousted as that was too expensive in short term for the financial people, when the investments didn't immediately give returns.
Plan now seems to try to split the company.
Which will kill the foundry part.
And when the foundry is dead, the CPU side will die too.
And there is very high chance that GPU side will get killed very soon.
As large part of the Pat outing was the low margins on the stuff produced on TSMC, to keep competitive in server space etc.
Not going to happen.
The same rotten board has been in control from before Intel starting to dip.
10+ years ago.
As in same people in the major board positions.
No, the actual reasons have leaked and low margins in certain segments he was pushing and overall dropping yearly profits (because of the investment etc.) were the major reasons.
And that the new investments didn't yet generate new revenue etc.
Which is to be expected when the investment is into new fabs etc.
But not to the financial people, 2 years ROI was too long, even when fabs take 5 years to build.
Those long term financial only people on the boards controlling positions want quick growing profits now, even if it will kill the company.
Money from spinning foundries out right now is the latest target it seems.
Even when Intel doesn't need that money to just survive, like AMD needed when Global Foundries was created.
Just that the profits and margins are going too low for the financial people in control.
But several of the previous high margin segments production had to be externalized to TSMC to make products that would be competitive.
But that dropped the margins on those a lot.
But you cannot get those margins back until the fabs are competitive with TSMC.
And there is nothing that would actually say that 18A isn't on time.
There was bad hit job with intentionally distorted stuff about yields in it that everyone copies without looking.
Just that Intel has to have 50%+1 share controlling interest on the specific fab(s) that will get CHIPS act money.
They can sell all the others, and they can sell that almost 50% share in the specific fabs.
pure karma
so when do we get to SLI 2x 5070s together
whenever you do it yourself realistically since i doubt were going to get any kind of GPU pairing again anytime soon
twin dew
Blue Origin New Glenn flight first flight planned for 2025-01-10.
And then SpaceX Starship IFT-7 for 2025-01-13.
Ah, the official coverate for that New Glenn flight is paywalled to Amazon Prime.
Ok, that might be wrong and might be also coming to YouTube.
soft bloom
when will people learn to use log10 axis?
twin dew
when will people learn to use log10 axis?
Intentional in this case?
The whole point is how much more stuff SpaceX is launching than anyone else.
In weight
soft bloom
linear histograms don't do well in such usecase
area plot would prove the point better
what i mean
if you want to show domination: take area plot
if you want to be informative - use log10
night girder
Are those astronauts still stuck?
So who cares how much metric volumes went up to space.
If we can't get back two astronauts on time from a space station in orbit of our planet π€£
twin dew
Are those astronauts still stuck?
Not really stuck at any point?
They could have came down on that Boeing capsule at any time.
Just that at first that extra time was used to test the failure stuff, as the thrusters were on part that was burned on re-entry.
And then NASA finally decided to not take a risk with all the unknowns and not to have them come down on Boeing capsule in the end (which would in the end have worked well enough for that).
And two astronauts were dropped from next Dragon flight, which went up with just 2 of planned 4.
And the Boeing test flight crew will be coming down on that capsule.
And are now part of that normal crew.
night girder
Oh so now you are saying, they just chilling there is space.
all accordintg to plan? rofl π€£
night girder
call it what you want. Everyone says they are stuck.
Oh right, no urls. Fuck it π€£ Just google for news articles yourself and find out.
twin dew
Stuck implies they couldn't have returned.
They could have, but that was decided to be to dangerous because of the unknowns in Boeings fuckups.
Where 1% chance of death would have been way too much.
No-one outside Boeing and NASA knows what the danger was calculated to be at in the end.
And yes, I know news outlets used that term.
night girder
Stuck implies they couldn't have returned.
They could have, but that was decided...
"They could have, but that was decided to be to dangerous because of the unknowns in Boeings fuckups." that's being stuck.
If I am on an island without a boat... surrounded by deadly water with sharks... am I stuck or not?
your reasoning is; "Nah you are not stuck since you can swim and being eaten by sharks".
Or the boat that comes to rescue, can't handle the weight and will sink. But that doesn't mean you are stuck. It means you could have escaped, and probably die π€£
Anyway, the point remains. Flexing with how much we can shoot up in space.
But can't bring back two astronauts safely (atm).
it just proofs how long a road a head we have when it comes to space.
twin dew
You are on island with everything needed to live.
You have ship that can take you back to mainland, but there is 0.5% chance that that ship will sink during that journey.
You want to get back home.
Do you stay 6 months and wait for ship in better condition, or do you go with what you have?
night girder
you are still stuck.
being stuck doesn't mean you can't escape.
You can be stuck in a cave. But that doesn't mean there are no options to get out.
twin dew
Is that being stuck?
night girder
It just means you yourself can't get out.
stuck /stΕk/
noun
A thrust.
Similar: thrustadjective
Trapped and unable to move. Unable to progress.
passenger planes have landed fine with door missing.
twin dew
Well, with lot more missing than just door:
night girder
it's just annoys me when you have standard definitions of words, and then we end up discussing if that words uses that definition or not.
Open a dictionary.
twin dew
Planes are engineered to be safe without doors, as long as the surrounding structure doesn't get too dinged.
Planes aren't designed to lose half of the fuselage for multiple meters.
That Aloha Airlines case was very nearly total loss instead of what it ended up being.
night girder
How much disk space do you think I need to update an IDE?
The IDE itself takes 3.67GB.
verbal raft
i cant imagine any rooms floor being this shiny
night girder
I need more than 30GB to update a 3.67GB IDE π€£
I got error after error. Because of diskspace.
Don't pay for IntelliJ products anyone. It's getting more and more garbage with every year.
Probably fired all the good developers and replaced them with cheap juniors.
night girder
the Jetbrain AI is put in as an add to our IDE.
Even if you don't want it, every new project you open has the AI sidebar. With a link to their AI.
And you pay like 200 euros a year for professional IDE.
But you cannot store windows for all projects.
so every new projects, your whole window layout resets 
And then you gotta load your own window layout, but you can't set it as a default. Which we could do. But they "renewed the IDE" with a new UI.
A.k.a. they stole VScode UI.
sorry, so annoyed with Jetdeadbrains.
I just installed vscodium (open source vscode* without telemetry)
I heard true coders use vim π
use what you like, jsut don't give jetdeadbrain any money (imo, they don't deserve it).
that's the thing I currently like about vscode, it's reverse from intelliJ. VScode stripped all bloat and offers them as extensions.
IntelliJ is just bloated with plugins you need to disable one by one.
verbal raft
soft bloom
so every new projects, your whole window layout resets <:hehe:382483156542029825...
Yeah. Because duck customers. It's Java
You have it easy pal
soft bloom
I heard true coders use vim π
Neovim?
Yeah, those servers with vim as default (like for git messages) truly make you apreciate freedom
It's not number of line that makes it spin
*ide
I absolutely loved when it was just soo fed up with my Jupyter files, it wouldn't even render window resize
There's something about idea of mixing text and binary in same file that makes it cry
I mean, ofcourse, when you try to edit 200MB file, it wont be happy
On log files, PyCharm freaks out when they are bigger than 2MB
And average plotly interactive plot adds like 3-5MB to it
No, only increased the heap size or whatever to like 4G
night girder
Yeah. Because duck customers. It's Java
java isn't my issue. the IDE's are.
edgy hazel
run AI
night girder
sure
willow pike
twin dew
Fourth of a frame at that
willow pike
there does seem to be more backlash to this shit this time
i watched the 40 series reveal live and immediately called out frame generation as what TVs do
harms input latency
i guess 3 out of 4 frames is too much for people
night girder
Yeah, but some of those people that are giving it shit
also will pay 2-3K for the next gen Nvidia GPU.
So it's half-half. I would rather have peoeple voicing their concerns and show it by making better decisions with their wallet (don't buy the product).
willow pike
if I see the 8GB 5060 anywhere on the steam chart I'm declaring war on gamers
twin dew
And it will almost certainly be the most sold GPU of this year...
night girder
You don't need a better GPU. DLSS4 will save you.
twin dew
But that is 5000-series only?
night girder
What if instead of making better GPU's? We stop producing better hardware. And just fake everything.
night girder
If we can tweak performance with code.
why buy hardware π¦
Did you buy a new mouse?
Question, if DLSS is software it's still hardware locked to GPU generation?
stray badger
Question, if DLSS is software it's still hardware locked to GPU generation?
Nvidia wants money
glossy glacier
and it has some HW acc
night girder
Can Geforce GTX 1080 run DLSS4 for example?
glossy glacier
probably technically yes, just slow af because it does not have the HW-acc
night girder
probably technically yes, just slow af because it does not have the HW-acc
makes sense. And what does it mean AI is generating frames?
Why do they say it's "AI" ... where does the intelligence come in?
glossy glacier
Β―_(γ)_/Β―
night girder
What does it learn?
Like is the AI making a model of visuals and them making frames based on that model?
glossy glacier
What does it learn?
pattern recognition
night girder
pattern recognition
mhh π€
glossy glacier
all current LLMs are just autocomplete on steroids
DLSS4 is autocomplete for game frames
night girder
chatGPT is google but smarter π
glossy glacier
no, dumber
night girder
well ...
What I wanted to say.
It's smarter in the sense that it connects the information together for you, instead of you needed to read the sources and put information together.
So it's also autocompleting that step π
Normally you look up a bunch sources, and than summarize it. But that's what chatGPT tries to do.
glossy glacier
yes and no. it "learned" what words are mostly used together. then it spits out these words and its mostly on point because the same information has mostly the same words used
night girder
yes and no. it "learned" what words are mostly used together. then it spits out ...
Not what I am trying to say.
Regardless of the technical details. A lot of time, chatGPT just makes a sumarrization of the information of the internet.
As of my last knowledge update in October 2023, NVIDIA had not officially released a technology called "DLSS 4." However, I can provide information on the benefits of DLSS (Deep Learning Super Sampling) in general, particularly focusing on the advancements seen in previous versions like DLSS 2.0 and DLSS 3.0. If DLSS 4 has been released since then, it would likely build upon these benefits. Here are some general advantages of DLSS:
Improved Performance: DLSS allows games to run at lower native resolutions while still delivering high-quality visuals. This can significantly boost frame rates, especially in graphically demanding games. Enhanced Visual Quality: By using AI and machine learning, DLSS can upscale lower-resolution images to higher resolutions, often resulting in visuals that are comparable to or better than native resolution rendering. Reduced GPU Load: DLSS can help reduce the workload on the GPU, allowing for better performance in games without sacrificing visual fidelity. Support for Ray Tracing: DLSS is particularly beneficial in games that utilize ray tracing, as it helps maintain performance levels while still providing high-quality lighting and reflections. Dynamic Resolution Scaling: DLSS can adapt to the performance of the system in real-time, dynamically adjusting the resolution to maintain smooth gameplay. Wider Compatibility: As DLSS technology evolves, it tends to support a broader range of games and applications, making it accessible to more gamers.If DLSS 4 has specific features or improvements, I recommend checking NVIDIA's official announcements or documentation for the most accurate and detailed information.
night girder
yes and no. it "learned" what words are mostly used together. then it spits out ...
But what you said is correct. Just not what I was trying to say π
I kept a folder with AI fails π
Why is GPT-4o mini stuck in 2023?
So far for it being intelligent π€£
btw, I just took that screenshot.
glossy glacier
it was trained on data up to 2023 then
twin dew
LLMs "information" stops when training data collection ends and training begins.
night girder
But this is the perfect proof how dumb AI is no?
You ask it a question in 2025, if it was a little bit intelligent, it would do the math.
Regardless of its training.
twin dew
And like said previously, they don't really know anything about anything, except about word associations.
glossy glacier
no, thats just missing information
night girder
no, thats just missing information
No, that's just dumb.
Intelligence means it would know it's 2025 and just do the math.
Also meaning it fully understood the question.
night girder
And like said previously, they don't really know anything about anything, except...
So it's dumb.
glossy glacier
its frozen in time. it cannot learn anything new
twin dew
Large Language Model
night girder
no, thats just missing information
btw, missing information = being "dumb"
glossy glacier
btw, missing information = being "dumb"
?????
twin dew
Because ChatGPT and other LLMs are about emulating natural text.
Not actual information.
night girder
Well, I am not going to argue with anyone. If you want to believe AI is intelligent I am not stopping you.
glossy glacier
btw, missing information = being "dumb"
so if i lock you in a room without access to the outside, you get dumber every day because you don't learn whats happening right now?
night girder
Yes, because if I come out after 10 years. And without doing anything, participate in a show who is the smartest person around the world. I would fail every question about whatever happened in those 10 years.
glossy glacier
but you could still do math just like before
night girder
It would not totally fail true, but your question was do you become dumber. And my answer is; in a way you do. Because you don't have access to keep up to date.
I am not saying you are regressing in intelligence.
I am saying, your data in your brain is not being updated with newest information, thus you can't give as an accurate or perfect answer as people who stayed up to date during those 10 years.
If I ask someone who comes out of a cave after 50 years. Who Billie Ellis is. I doubt they know.
Not saying they are dumb.
But they are dumber relatively to others. And if they read up, they can catch up.
If I ask someone who left a cave after 50 years; ok how long is 1919 ago? They would probably first look up the current date, or ask. And then do the math.
Unless they kept a calender in the cave. Which I doubt. No day/night light so in a cave you get totally confused and lose track of time.
twin dew
but you could still do math just like before
And AFAIK no current LLM know how to do math.
night girder
Anyway, what did I say wrong?
glossy glacier
And AFAIK no current LLM know how to do math.
"just like before", genius stays genius and no skill stays no skill π
but yeah, math being a perfect example of LLMs not being "Intelligent"
night girder
glossy glacier
just because they saw a lot of "2+2=4" does not mean they know why
night girder
And that's so freaking stupid.
glossy glacier
i stand by what i said: automcomplete on steroids
night girder
Personally, I wouldn't be so harsh to AI. But the way they sold it, and what they promised, and the marketing language etc.
glossy glacier
Exhibit A: "AI PSU" becasue it could be used in a PC that could run a LLM/AI
night girder
Doubt anyone is interested, but this is the first paper known to AI that mentions Artificial Intelligence.
verbal raft
Exhibit A: "AI PSU" becasue it *could* be used in a PC that *could* run a LLM/AI
"AI PSU" just makes me DIE INSIDE
night girder
and the funny part is 5.
verbal raft
you need AI to control transients
glossy glacier
"AI PSU" just makes me DIE INSIDE
i wish i was kidding
night girder
night girder
Razer announced AI trainer.
They don't record. It's not like google glasses I think.
Which was a major privacy concern. And we had accidents with it.
I believe there even was a bar fight because of the google glasses.
I am thinking.
One part of me wants to say; nah goverments wont allow it. But the other part is like; but we have smartphones with camera's, we have normal camera's, we have security camera's, ... what is the difference?
glossy glacier
"why should i? the other guy with (normal) glasses does not either!"
night girder
What difference does it make if someone walks into the toilet with their phone out (and maybe recording) or glasses (and maybe recording)?
Meta smart glasses worn by the New Orleans attacker have raised privacy concerns
I am in favor for it.
night girder
> Meta smart glasses worn by the New Orleans attacker have raised privacy concer...
it's already starting π
glossy glacier
> Meta smart glasses worn by the New Orleans attacker have raised privacy concer...
what dis? USA and privacy concerns in the same sentence?!
night girder
Ok, no. False alarm.
They are not concerned for our privacy.
their use as an aid to carry out a terrorist attack is a disturbing turn for the product
glossy glacier
ah, that makes more sense
glossy glacier
...another reeason not to get a credit card
night girder
made what illegal?
glossy glacier
sucks to be a US citicen i guess
glossy glacier
it concerns me you would even need to game the system
night girder
I am more confused now. So I pushed the AI to tell me why it though it was 2023.
October 2021. π€¨
So it's training data is based on 2021. Or it thinks so. Yet, it believes now it's 2023. While it's 2025.
If that is the case, it should believe it's 2021.
glossy glacier
you all are wrong, it the year 12025
get it right
night girder
Fuck me, the AI is brushing me off;
"I should not have assumed a year beyond my training cutoff of 2021"
night girder
Why not ask a AI
willow pike
night girder
What is that other sentence about? 'One reddit user says "Kill yourself"' ???
willow pike
the ai thinks it's a valid response to how many usb ports a motherboard has
redditors have already started filling some subs with bullshit for ais to index
alas yes
night girder
I think it might be funny to have a AI based on 4chan π€£
the pure chaos.
and adding forced subscription
languid gulch
so our chat the other day about PCBs being sideways to get back to the outputs might not have been that insane of a discussion π€£
tribal kraken
Just followed Facebook conversation of hobby/home engineers trying to solve how much a "screw lift" using M16x2 thread can lift when applying 1000nm of torque on it, with the help of chatGPT. Results were funny
I need ai solution to get my cat off the tree
mental oriole
Cut down the tree.
twin dew
Buy Nvidia AI server cluster, and use the racks as stairs to get to the cat?
tribal kraken
weather and hunger worked
tribal kraken
Yeah hes safe, he was not scared. Wanted to play, ran towards me, rolled once in the snow on my boots and then when I tried to lift him, ran up the tree. There are people at the house who would let him inside even if I did not get him down before leaving to work.
languid gulch
Yeah hes safe, he was not scared. Wanted to play, ran towards me, rolled once in...
cats are assholes. he knew exactly what he was doing
(cat owner btw)
they can smell cortisol, which explains why they love to go right for the person that doesn't like cats
yea but that doesn't mean they won't try π€£
night girder
I need ai solution to get my cat off the tree
Wait, I got you!
If it's a high tree, which is doesn't look like, listen to 5. Trust me.
I've had once somebody climb up in a high tree for a cat, all the way up there, the cat started to fight the person because it was so scared.
So the guy was hanging there in a tree, with a cat trying to scratch his face off, trying to get the cat in a basket to lower it down.
Was a professional. So ropes etc.
I did.
And a week later somebody drove over it.
Yup. I lost fate in cats because of that π¦
- Don't want to hold em hostage in a house. 2) Don't want people driving over my cats.
So, I should buy an island? π€
But there is still people there.
verbal raft
willow pike
that describes literally everyone
tribal kraken
Wait, I got you!
AI got that right
night girder
What old world?
willow pike
if AMD hadn't spun out globalfoundries, they would have died 5 years ago
nvidia never attempted to own their own fabs, which allowed them to succeed
edgy hazel
a what now
night girder
technically the truth.
Today, I updated IntelliJ IDE and it asked me permission for my microphone.
Sigh ... π
But it has amazing features!
This is the reason why you need a powerfull IDE. Machine Learning code in Marketplace π₯³
All this shit was enabled by default btw.
night girder
cry-engine π
winged valley
Guys should I get a 5070ti on launch day at my local micro center
I have a 3080ti rn and I've lately been trying to play ark survival ascended on a 1440p screen so it WOULD be a substantial upgrade for me
twin dew
Almost certainly not?
winged valley
I'm also going for DLSS 4
As far as I'm aware, micro center selling FE cards follows NVIDIA's posted prices on their website which is like $749 for the 5070ti or something like that
While the 5080 is faster in raw graphics performance as well as RT and DLSS, it costs $250 more and has less VRAM. It just so happens that ASA is a VRAM intensive game
I do want to go AMD some day but as it stands right now the games I care most about support only DLSS and/or have RT options. ASA also supports framegen which would be super helpful to me. So ultimately NVIDIA is the choice for me atm
twin dew
What about Nvidia?
And how would not counting the time Jensen was talking about his new leather jacket affect that?
verbal raft
What about Nvidia?
he didnt cover NV yet
verbal raft
And how would not counting the time Jensen was talking about his new leather jac...
that Should be included
twin dew
Meant that there would be two Nvidia numbers.
One taking into account the whole time, another just the part where Jensen wasn't talking about his jacket.
winged valley
That's comical
willow pike
Guys should I get a 5070ti on launch day at my local micro center
get a new hobby
willow pike
1x RDNA4 mention total
willow pike
I do want to go AMD some day but as it stands right now the games I care most ab...
you can use opticscaler to get FSR framegen with your DLSS i believe
winged valley
you can use opticscaler to get FSR framegen with your DLSS i believe
I'm fine not doing things with drivers
willow pike
ah it's not a driver fuckery, it's a game mod
winged valley
Although I will check it out when I eventually go AMD
willow pike
you can use it on nvidia
winged valley
ah it's not a driver fuckery, it's a game mod
That is worrying for a game like ASA which uses Battleye
willow pike
it pretty much lets you use anything on anything
ah yeah fair, it's not 100% guaranteed to be anticheat safe
verbal raft
That is worrying for a game like ASA which uses Battleye
ASA already has FG
its just that.....
it can be fixed by changing the config files
winged valley
ASA already has FG
I have a 3080ti, not a 40 series card
Therefore I cannot use FG without external stuff
verbal raft
I have a 3080ti, not a 40 series card
but i meant FSR FG
verbal raft
its meant to be enabled by default
verbal raft
but the devs fucked up
winged valley
Typical studio wildcard
Except isn't the dev for ASA grove street games of all people?
The same studio that fucked up the gta definitive collection
verbal raft
Except isn't the dev for ASA grove street games of all people?
nah thats just the iphone mobile port
winged valley
nah thats just the iphone mobile port
from steam
winged valley
Wait but its cross platform so they might've helped with that part
verbal raft
anyway..... here are the fixes for the engine.INI
[/Script/FFXFSR3Settings.FFXFSR3Settings]
r.FidelityFX.FSR3.UseNativeDX12=True
r.FidelityFX.FI.overrideSwapChainDX12=true
winged valley
Typical studio wildcard
My favorite wildcard moment is making the map with the most beautiful lighting effects (aberration) have the most fucked up visual glitches (my top two are the black bar of death and liquid element blinding you)
I've heard these bugs carried over into ASA
verbal raft
anyway..... here are the fixes for the engine.INI
[/Script/FFXFSR3Settings.FFXF...
you can also add : r.FidelityFX.FI.enabled=1 if you want FG to be always enabled
verbal raft
My favorite wildcard moment is making the map with the most beautiful lighting e...
i havent played AB yet
only recentely beat the island bosses for the first time
winged valley
ASE AB is goated and I argue it's the best map they've released, with Gen2 being a close second and Extinction being third
Rock Drakes my beloved
As for ASA AB, no clue. I've only been playing island so far
verbal raft
ASE AB is goated and I argue it's the best map they've released, with Gen2 being...
i dint buy ASE ABπ
winged valley
I'm running a public ASA server if you want to join it, it's for an existing (tiny) ASE community
verbal raft
also : ASA doesnt perform that much worse than ASE did for me
winged valley
It depends on what settings you run it on
verbal raft
(at least when ASA isnt have a sezeiure)
winged valley
If you go full potato mode it doesn't run much worse than ASE
verbal raft
If you go full potato mode it doesn't run much worse than ASE
and still looks a lot better
winged valley
Yes
verbal raft
on the low preset anyway
winged valley
No I mean FULL potato mode
With the launch options like forcing dx11 and disabling various effects
verbal raft
(at least when ASA isnt have a sezeiure)
by seizure i mean running out of VRAM
verbal raft
With the launch options like forcing dx11 and disabling various effects
if you disable lumen ASA looks way worse
IMO
winged valley
by seizure i mean running out of VRAM
I had to manually increase my pagefile because ASA doesn't really like 12gb of VRAM on 1440p medium-custom
winged valley
if you disable lumen ASA looks way worse
Good thing I didn't disable
verbal raft
I had to manually increase my pagefile because ASA doesn't really like 12gb of V...
oh
if you knew how good it is on 8Gb
winged valley
Frequent crashes for "out of video memory" I take it?
verbal raft
<:hehe:382483156542029825>
50% of the time the textures arent even there
winged valley
Jesus
verbal raft
Frequent crashes for "out of video memory" I take it?
nah
just 5fps
standard UE5
standard RX6600
winged valley
ASA is truly a next-next-next-gen game (it has ungodly bad optimization)
I don't know how
verbal raft
ASA is truly a next-next-next-gen game (it has ungodly bad optimization)
yeah
IMO there isnt a single game that comes close to it
winged valley
But they managed to optimize ASA even worse than ASE was back when it was a UE3 asset flip with worse spaghetti code than it has now
verbal raft
But they managed to optimize ASA even worse than ASE was back when it was a UE3 ...
the more things change
the more they stay the same
im just waiting for RDNA 4
now
stray badger
i thought ark was ue4?
willow pike
ASA is grove street games?????
stray badger
But they managed to optimize ASA even worse than ASE was back when it was a UE3 ...
response to this
verbal raft
response to this
well now it is UE4 back then it wasnt
verbal raft
*ASA is grove street games?????*
partially
i guess
i dont think grove street was there back when it launched
winged valley
i thought ark was ue4?
The original original ARK, the first release, was an asset flip on UE3
It DID have a lot of original work, but it also had a pretty high number of store assets from what I remember
verbal raft
The original original ARK, the first release, was an asset flip on UE3
is there any footage of that?
winged valley
is there any footage of that?
This would've been early access, so way before 1.0. Maybe? Probably
verbal raft
IMO there isnt a single game that comes close to it
i'd argue even CP 77 W PT looks worse
and
runs worse
winged valley
So looking into it, ark's first early access release was 2015, UE4 was released in 2014. Most sources I'm seeing say that it was released with UE4
However it is highly probable that development started on UE3
verbal raft
also
ASA looks really bad without DLSS
as if it was designed to use DLSS
especially because TSR is very blurry by default
so much so that i'd argue FSR looks better
even though you cant enable it in the settings
stray badger
However it is highly probable that development started on UE3
idk, UE3 and UE4 were very very different
they would have to rewrite the whole game from scratch
wanton orchid
good news, I'm now trained at doing the paper clip concret ECC RAM error correction and reporting actual test
and it works well with my RAM and report correctly
I mean, appart a weird delay in reporting
it does report and does correct
nothing more reliable than a paper clip
soft bloom
hmm, error correction
I wonder if it (lack of) could explain why function optimization was yielding delta y
soft bloom
I tried replicating the Length of Day plot using ephemeris
but I don't get why there some wobbles in results
wanton orchid
is it static calculation or are you doing iterations ?
did you check you are not dropping the leap seconds
soft bloom
did you check you are not dropping the leap seconds
most definitely, time counted in tdb
soft bloom
is it static calculation or are you doing iterations ?
well, I used bentq optimizer
thing is that it (delta y aka delta azimuth) goes below xtol for the most part, but sometimes it just decides to be couple magnitudes worse than requested
and these are often close by time, in little groups of 10 or so days
and that didn't happen with solar day
willow pike
kinda wanna put these into a gif
but they don't line up perfectly
are these from digital foundry?
swear to fuck these guys are basically an nvidia marketing arm
night girder
check out the light pole on the right.
it's going back n forth and than dissapears.
verbal raft
check out the light pole on the right.
nah that just goes off screen
night girder
no.
verbal raft
from the way i was screenshotting
night girder
the frames don't add up.
and the lightning pole goes from real to -> frame 1 left -> frame 2 right -> frame 3 it's gone.
willow pike
just mean in the snaps you put in discord
I can see plainly how awful frame gen is
but it wouldn't look right if I stacked them in a gif
night girder
no
I just did that.
Doesn't look right at all when you make a gif out of them π€
pure karma
i literally showed how bad it was holding up on cyberpunk a few days ago
in fact im tempted to do a frame by frame of that
willow pike
it needs to be interrogated
verbal raft
and the lightning pole goes from real to -> frame 1 left -> frame 2 right -> fra...
nah thats just from how i was screenshotting
night girder
night girder
That makes way more sense π
verbal raft
Doesn't look right at all when you make a gif out of them π€
neither does it look good in the actual footage playback:)
willow pike
nyoom
verbal raft
https://cdn.discordapp.com/attachments/1236717847011393546/1327005104225783900/D...
the guy that recoreded this decided to speed it up for some stupid reason smh
π€¦ββοΈ
night girder
We need to know how it looks on SF.
verbal raft
We need to know how it looks on SF.
well wait until we get RTX 50
verbal raft
well wait until we get RTX 50
and someone that cares enough to show us
willow pike
it can take multiple seconds at 120fps for lumen to fill in properly
night girder
to be fair, developers said that they added it. But we shouldn't expect a lot of support for it π
willow pike
if I use frame gen, monitor capped at 144, real frame rate 36, lumen will take 4x as long
verbal raft
if I use frame gen, monitor capped at 144, real frame rate 36, lumen will take 4...
oh yeah
willow pike
from 1-3 seconds to 4-12 seconds to fill in
night girder
in general or in SF?
verbal raft
lumen denoiser = TAA
verbal raft
SF specifically
yeah SF prolly has the most sploshy lumen i have ever seen
night girder
Like I said, Snutt told us not to expect too much of it.
willow pike
it's not tuned for it and that's fair enough, it's sectioned away in its own setting
night girder
They just added it. And that's it π€£
willow pike
DLSS4 will only make it worse
night girder
I think they half assed implemented it
verbal raft
DLSS4 will only make it worse
maybe the denoiser from it could help
night girder
just enough to keep things working
verbal raft
I think they half assed implemented it
they flipped a switch
willow pike
yeah I'm not complaining, game dev is hard and SF got hideously out of scope
night girder
I don't use lumen.
verbal raft
yeah I'm not complaining, game dev is hard and SF got hideously out of scope
and in my opinion : it is literally night and day
willow pike
s'good
night girder
It is.
verbal raft
willow pike
when it properly collects, it's gorgeous
this is the not SF channel but I guess we're talking tech
night girder
but they never made the game with lumen in mind.
verbal raft
when it properly collects, it's gorgeous
i honestly didnt stop much for these screenshotsπ
night girder
so the lightning will always be kinda off I guess?
willow pike
lighting is hard
night girder
Yeah, because you can clearly see, you almost get a different game π€£
willow pike
nvidia are not helping, their idea of help is airdropping a ReSTIR .DLL file that nobody can run properly and peacing out
verbal raft
and in my opinion : it is literally night and day
TBF my factory was 100% built to use lumen
night girder
But the same seems to go for other games.
verbal raft
But the same seems to go for other games.
well i dont think wukong was desinged to be played without lumen
night girder
This is kinda dope.
willow pike
PT is too slow
verbal raft
This is kinda dope.
medium is almost the same as cinematic
night girder
medium is almost the same as cinematic
Ther is a big difference in lightning top left corner.
but yeah, I would say cinematic and very high arre very close.
willow pike
wukong uses ReSTIR for path tracing, which completely collapses if you don't have it at maximum; without path tracing it uses lumen and scales to what you can handle
willow pike
if cinematic is the built in unreal preset, it's not supposed to be used in games
it's meant for cinema
literal 24fps
night girder
Yeah, just neat to see all 5 next to eachother. Nice overview.
willow pike
if you don't use that, wukong isn't too demanding
night girder
I haven't played it.
willow pike
but yeah the cinematic preset is intended for use with the volume
The Volume
huge LED screens used instead of a bluescreen
camera synced up to something running unreal engine to render the backdrop in real time
verbal raft
well i dont think wukong was desinged to be played without lumen
same with smear ascended
willow pike
soft bloom
oh wait. overhead displays as primary source of light? hmmm
languid gulch
oh yea the BTS documentary on it is astounding
soft bloom
though reminds me of my fuel factory
verbal raft
oh wait. overhead displays as primary source of light? hmmm
yeah,thought it looked cool
willow pike
oh wait. overhead displays as primary source of light? hmmm
imagine how many interns it would take to paint out the greenscreen from the mandalorian's helmet
misreply I am fat thumb
soft bloom
yeah,thought it looked cool
it does
found this while looking for my "fuel factory". and now this gives sucha neon-city/hong-kong vibes (reflection of zebra)
#screenshots message
#screenshots message
actually i thought it looked darker
yep, coated concrete reflects a lot
verbal raft
it does
found this while looking for my "fuel factory". and now this gives such...
i had to tweak lumen reflections so they wouldnt freak out from moving
RR is prolly doing some good work there for you
soft bloom
rr?
verbal raft
also never thought of a design like that #screenshots message
willow pike
lumen is not ray tracing
soft bloom
i am amd
willow pike
it does not use RR
verbal raft
lumen is not ray tracing
i have seen RR turned on with lumen
willow pike
are you sure
willow pike
that sounds highly bizarre
unless you can use RR with hardware lumen but hardware lumen is still partially SDF
soft bloom
also never thought of a design like that https://discord.com/channels/3704729390...
oh I got inspired by pioneer of displays as light source
but tuned it down to just cheap light strips on frames
does well in blueprints (though eats steel)
but you can easily make factories that are lit up and don't eat performance (regular lights are heavy)
don't know what's the reason, but I can have rows of light fixtures no sweat but single ceiling light halves fps
soft bloom
1440p then right?
hmmm, I think it was rendering in 1440p (less probably 1080p) with upscaling to 4k by FSR
verbal raft
hmmm, I think it was rendering in 1440p (less probably 1080p) with upscaling to ...
oh
then thats why
i am at 1080p
native
mostly
soft bloom
I only now noticed how pipes block part of light in reflections
verbal raft
with FG though
verbal raft
unless you can use RR with hardware lumen but hardware lumen is still partially ...
Stalker 2 only has software lumen so far
https://www.youtube.com/watch?v=yWkYUbdUIzg @willow pike
soft bloom
with FG though
FG makes light flicker, expecially at distance
couldn't show that wit hscreenshot, but basically i have radar towers in this design and they almost look like tall torches - flickering
willow pike
there's absolutely no way RR was built for this
it's meant to do ray trace denoising and DLSS denoising in one step
modding it onto software lumen is like putting the head of a pig onto a cow
verbal raft
https://www.youtube.com/watch?v=yWkYUbdUIzg <@333225861085921282>
though reflections have NEVER shimmered this much for me
verbal raft
willow pike
https://www.youtube.com/watch?v=yWkYUbdUIzg <@333225861085921282>
the reflections are more stable because they have less stuff in them
they're also much darker
the comments in the mod are a flame war
some say it don't do shit
be proud; don't give these fucks your money
top score in your favourite game and tell them your pc is shitass yet you won anyway
there's a lot of "my uncle works at Nintendo" going on in the comments of this mod so I'm bailing out
languid gulch
the last nintendo product i owned was a gamecube, and that was years after it came out & i grabbed a used one π€£
i think the only games i played on it was simpsons hit & run & paper mario
wanton orchid
as soon as mobiles and pc got more powerful than their custom platforms, it shifted to shit
I think Wiu was when it shifted
wanton orchid
waiting for 6.12.9 to land is frustrating
wanted to install it before going to sleep
soft bloom
Huh? Huh!
willow pike
i don't think japan has fair use like we do, so such companies often act like it doesn't exist
still, they decided to use parents to fight palworld, a much easier battle than copyright, so they know how to play the game
because the DMCA is intended for big companies to talk to other big companies, it's not fit for purpose with small creators
youtube bends to the big companies because it's cheaper, big companies exert maximum DMCA influence because they can
it's only one example of how the law works to protect big money
soft bloom
youtube bends to the big companies because it's cheaper, big companies exert max...
YT
cheaper
Since when do they care about costs?
I mean as company
They wouldn't exist if their goal was profit
pure karma
not only that but imputs are still only going to regsiter if pressed during the real frames so its going to keep getting worst
willow pike
languid gulch
even on an SSD my kerbal space program modded game takes a solid 10m
it might be a bit on the heavy side
twin dew
lumen is not ray tracing
Lumen is software/hardware raytracing solution in Unreal Engine 5.
It either does that heavy lifting with Distance Cloud Shaders in software mode.
Or with GPU RT-HW in hardware mode.
So in both cases GPU is doing most of the work, but in software mode it doesn't use the newer RT-HW stuff but just normal shader cores.
But even then it is raytracing.
soft bloom
why would Disk util report 10 more GB on LVM volume, while looking at properies reports 10 GB less?
twin dew
GB vs. GiB
1000 vs 1024 per power of 3
Well, powers of 10 vs powers of 2
10^3 (kB) vs 2^10 (kiB), where traditionally kB was that second one too, until HDD manufacturers changed it.
soft bloom
eh, no
because the warning popup appear when Disk still reports 11 free GB
twin dew
You had the most limited single-language only OEM edition, so still not without complete reinstall.
If even then.
In any other license version, in Settings.
So for Home, there is:
OEM Home, Single-Language only
OEM Home
Retail Home
Where in those two more expensive SKUs you can freely change the OS language.
Yes, you get "free" Windows license while you are playing guinea pig for MS and working as free QA.
Not for main system.
If you just select the tier that acts as last line mass testing, then it usually is ok.
But the earlier tiers break all the time in various ways.
soft bloom
eh, no
because the warning popup appear when Disk still reports 11 free GB
they decrese linearly
rn Disks reports 16free, and folder properties repots 5GB
twin dew
I have no idea what numbers this is about.
soft bloom
how does that imapct reporting?
twin dew
What reporting?
You never properly explained WHAT programs/screens numbers you are comparing
twin dew
vs the folder's (Nautilus) properties (it says `free space`)
Nautilus removes the "root only" reserved blocks
Which defaults to 5% if not specified when Ext4 partition is made.
And doesn't count those as free space for user.
As normal user isn't allowed to use them.
So in this case about 12.85 GB
If that default 5% of 258GB partition size.
soft bloom
oh, that seems about right
dam, i need to clean it
speaking of cleaning and IDE's
I spotted a 3.6GB file 'java error reporting in PyCharm'
assuming that's memory dump
was sitting there for ouple of months
steam somehow takes 20GB...
what's proper drive
soft bloom
On Linux i don't care about that.
twin dew
There is just single directory tree from / up.
And any directory can actually be another partition, or just directory on same partition as previous one.
mount is the command to attach a block device/file, that contains a file system, as directory.
And /mnt/ is just common directory for random temporary mounts.
soft bloom
There is just single directory tree from / up.
And any directory can actually be...
which is super handy
twin dew
Can be also done on Windows on NTFS partitions.
Adding more partitions as folders.
But almost no-one uses that.
soft bloom
btw i think Steam on linux somehow still knows if fodler is on different physcial partition
is there a better way to provide options to various objects inside the program other than a nested config file?
I don't know why am unsatisfied
maybe just because it's already pretty nested fuckery
limber flicker
Does anyone know why my pc wonβt even turn on
The only things Iβve touched is the gpu and ram I took out the ram and put some new 1s in as well as took out and put the gpu in to give me room to actually be able to put the ram in
pure karma
could be a million reassons with that little info
limber flicker
could be a million reassons with that little info
Thatβs the only things that have changed since it last turned on
pure karma
if theres no power at all then maybe turn back on the power supply
limber flicker
Done that
twin dew
The only things Iβve touched is the gpu and ram I took out the ram and put some ...
And does the motherboard have diagostic leds?
And have you cleared CMOS if you had XMP enabled?
limber flicker
And does the motherboard have diagostic leds?
I donβt know if it does and I donβt know how to clear cmos
twin dew
What motherboard?
limber flicker
I donβt think it has the leds
Uhh pretty sure asus I donβt know the exact details
twin dew
Need the exact model.
limber flicker
Need the exact model.
^
twin dew
Yeah, that doesn't have any diagnostic features outside of PC beeper header.
limber flicker
What about the cmos any idea how I clear it?
twin dew
limber flicker
The 1 or 2
twin dew
10:
twin dew
The 1 or 2
Read?
Page 20 is start of part 1.6: Headers and Jumpers.
And what is the first subpart?
night girder
What is 11. Intel ME jumper?
limber flicker
Aight Iβll give it a go now
night girder
It says it enabled/disable Intel ME function.
twin dew
Also allows to not have some of the ME stuff be disabled when in that update mode.
night girder
never heard of that.
limber flicker
Im so lost looking at this thing
twin dew
You remove power cord from PSU.
Then you short those specific two pins for moment.
Then replug power cord and try to start.
And if that doesn't help, you either don't have something plugged in correctly, or you broke something.
But without connecting specific type of speaker into another header, to get audio error codes, cannot say what.
pure karma
π€£
pure karma
well if theres no power at all id guess power supply either went kaboom or the mothebrored went kaput but that wouldent happen even if you somehow managed to short half the dang motherbored with the gpu so its unlikely
limber flicker
So all I do is touch those two pins at the same time with a screwdriver for like a second?
Also is a magnetic tip screwdriver all good for it?
twin dew
With that power disconnected.
Anything metallic which you can use to contact both of the pins at same time.
limber flicker
Of the motherboard?
twin dew
Q170 chipset, so almost 10 year old large OEM made machine.
limber flicker
I believe I made them touch itβs in such a awkward place to see what Iβm doing
Itβs deep in there
I found the two pins I know which ones they are
Yep pc not on
I think itβs going kaput
twin dew
If that didn't help, check if RAM and GPU are properly seated.
twin dew
Pictures from 45 degrees to the side for the RAM, and from 45 degrees from up for the PCIe connector?
twin dew
That RAM isn't seated.
dire igloo
The RAM stick furthest from cpu
twin dew
The other top was ok.
Downside was impossible to say because of that lack of moveable locks.
dire igloo
Nope ox wonβt start
Send pics of all power connections
limber flicker
Like all the cables?
dire igloo
Like all the cables?
Yup
twin dew
I would expect it to be that the other ends aren't in place either.
Or would try without that GPU installed and with monitor in IGP connections.
If that isn't F CPU
limber flicker
The cpu is i7 6700
twin dew
That needs money
limber flicker
It is bit dusty
dire igloo
Should've specified: side view of the power connections
twin dew
Probably weren't touched except for that GPU one.
limber flicker
The only cable that might not be in right is the gpu thatβs the only 1 I have touched since its stopped working
dire igloo
Figuring out what this "something" is is what we're trying to do
twin dew
Which was the point of seeing if the computer starts with IGP, when GPU isn't installed.
And if it doesn't, then trying with the previous RAM sticks again.
dire igloo
The cpu is i7 6700
Anyways, do what Baldur said.
Remove GPU, put monitor into mainboard video output.
limber flicker
Will do now
twin dew
In a way already have in latest APUs, which have mix of full and compact cores.
Where the compact ones don't clock as high as the full ones.
But are otherwise identical.
limber flicker
No display ports on the motherboard from what I can see
twin dew
Should have two DVI ports.
limber flicker
Yea it does I got no cables for that tho is the issue
twin dew
Then take out both RAM sticks, install the old ones, make sure all four corners actually go fully down.
You need to place the ends without moving clips first.
And then press those down first too.
limber flicker
The clips lock themselves in
This is the first time Iβve ever touched the components of my computer just to swap the ram and now it doesnβt start up
Yea
Still not starting up
Even with the old ram
Could do maybe
Kinda want to keep this one running tho
twin dew
Do the fans turn on at all?
limber flicker
Nothing makes a noise or moves
twin dew
Is the power switch near the power cord connector on?
twin dew
Try without RAM and GPU at all.
If fans don't at least momentarely turn on, then there is something wrong in either PSU not giving MB power, or MB not handling front power switch input correctly, or that switch not working.
limber flicker
Ok Iβll give that a go real quick
twin dew
Was mainly as next step tomorrow.
twin dew
So either there is short somewhere and PSU doesn't start.
MB is toast.
ATX connector is not properly connected.
Or power switch connection isn't working, from MB front panel header to the switch etc.
But because that MB doesn't seem to even have standby power indication LED anywhere on it, cannot easily check some of those.
limber flicker
Yea I think my best bet might be taking it to a local computer shop but I already know they gonna rip me off to check it
pure karma
Especially for a old system like that
limber flicker
Yea good ol fun
twin dew
Tomorrow you could disconnect PSU cables from motherboard etc.
Then do the pure PSU start test:
https://www.silverstonetek.com/upload/downloads/QA/PSU/PSU-Paper Clip-EN.pdf
limber flicker
Will do
I just realised it probably is the psu as whenever I have my pc plugged in even if my pc isnβt on the leds on my keyboard and mouse turn on
But they arenβt now
twin dew
That is that standby part.
So short preventing PSU starting because of protections.
PSU broken.
MB broken (causing that short)
ATX connector not properly in place.
Well, they might not turn on until first start.
limber flicker
Yea Iβll figure it out tommorow maybe
wanton orchid
weird
I keep having lower memory performance since I restarted testing some new bios settings and new kernel
I tried changing it back but no luck
it's stuck at lower memory performance for some reasons
soft bloom
If that default 5% of 258GB partition size.
I wonder how I didn't notice this before...
I thought other lvm partitions had correct reportings
wanton orchid
I keep having lower memory performance since I restarted testing some new bios s...
wait it reverted itself
transient stuff or missconfigured testing ? I will see
soft bloom
i think i need a book about how to read CWT
sometimes i think they are cursed
for example, wtf is this line?
wanton orchid
off sync again
wanton orchid
I keep having lower memory performance since I restarted testing some new bios s...
it was both testing and settings
for everyone reading this : if you use 2x1x2 RAM config at 1T(no-geardown), 1KB interleaving is better
twin dew
Probably depends on the exact load.
Interleaving type things depend on the memory access pattern very heavily.
Affects how RAM addresses are striped over memory banks and ranks.
wanton orchid
not as much as you think, it is mostly address masking
there is a big granularity (I mean as it is not fine grained)
it is also linked with actual RAM chip access config, that's why I specified it
it depends more of the RAM setup than the load
because load is always masked by cache then iod then imc
the only thing load challenge is mostly the bandwidth/latency tradeoff
but unless you pick a single load type you can mostly check there is no big downside
dire igloo
i think i need a book about how to read CWT
Channel Well Technologies?
wanton orchid
for instance a cacheline is 32 or 64bytes, here we are talking 256 512 1024 2048
you will rarely hit multiple close cachelines without it to be almost sequenced (in which case it is prempted anyway)
what is changes though is how much command and banks have to be accessed / can be loaded, with each page size access
128 would not make sense, and 4096 would not make sense either on dualchannel
wanton orchid
but there is a tradeoff where you increase throughput and does not loose latency
or increase latency without loosing throughput if you see it the other way
wanton orchid
but there is a tradeoff where you increase throughput and does not loose latency...
what I mean is that the corner cases get enclosed by cache and imc
you do not change much things in these cases anyway
but you can make it better overall which you will almost always hit in EVERY workloads
(shit corner cases will always be shit no matter what, and others are handled by cache prefetchers etc)
so I ended up with :
1T 2x1x2 1KB
2T 2x2x2 2KB
lower can make it better for big close burst of data, but increase latency and then throughput eitherway
tldr:
] sparse small data access will always hit a single bank, no matter the setting
] big sequenced access will always hit all the banks, no matter the setting
] the case it matter is in memory bus command and bank data fetching tradeoff, so it is mostly a RAM architecture/setup thing, more than workload type
anyway hope I successfully pointed out it is a general workload optimization / ram tuning, not a workload specific tuning
because we are not tuning core and cache behaviors here
soft bloom
https://en.wikipedia.org/wiki/Cepstrum
nice, i need that!
wanton orchid
maybe on 2x1x1 (sr) 2KB is better
because that is a full fetch anyway
Idk
I may endup confusing my self at this point
dire igloo
nice, i need that!
wait, unironically?
wanton orchid
ngl I'm almost pissed off, because checking these settings is getting fucking long
I mix it with ECC symbol size, to check if ECC symbol size can restrict burst memory pattern (which I do not want)
wanton orchid
maybe on 2x1x1 (sr) 2KB is better
because that is a full fetch anyway
my logic is : more dimm / ranks, more bank to use but higher latency anyway
so more interleaving means, better use of the big number of banks, so it compensate for increased latency
but latency is still there
twin dew
More ranks doesn't increase latency.
Just limits max frequency etc.
wanton orchid
I mean more bank for same size
or am I getting it backward and bigger memory have also more banks into them
twin dew
Same number of banks independent of size.
Except that x16 chips have less banks for same size compared to x8 or x4 chips.
wanton orchid
if you can burst few banks, and get your data right away, always going to be quicker than having to access 8 banks instead to fetch the same data, and then having data you dont need in the end of the burst
verbal raft
they are talking about ERROR LAKE
"AMD chalked it up to unprecedented demand and noted that Intel's "horrible" product, also known as Arrow Lake, has led to wildly increased demand that has pushed far beyond initial projections"
twin dew
DDR5 doubled the number of banks in chip compared to DDR4, which made it that on DDR4, having 2 ranks gave significant real world speed boost compared to just 1 rank, but on DDR5 it doesn't.
As 1 rank of DDR4 didn't have enough bank groups, 1 rank of DDR5 has (except x16 chips)
wanton orchid
anyway going off to test that last setting smh
twin dew
AFAIK that ECC Symbol Size depends on the chips used for the ECC.
How wide the ECC part of data bus is in reality.
4 or 8 bits.
Where wider allows for better multi-bit ECC correction before uncorrected error.
verbal raft
https://www.tomshardware.com/pc-components/cpus/amd-says-intels-horrible-product...
esentially AMD didnt think ERROR LAKE would be THIS SH*T
soft bloom
wait, unironically?
well, if it helps me understand frequency component...
althought not exactly what i asked for
there's a separate place in hell for those how place resiable text window at the bottom of the page and dump thousands oflines there leaving you with short window that you can spend hours on resizing
wanton orchid
Where wider allows for better multi-bit ECC correction before uncorrected error.
that I know, but I wonder is pushing bigger symbol also restrict burst
(basically it would always need all the data for checking the full bit size/length)
twin dew
Happens per parallel transfer.
wanton orchid
AFAIK that ECC Symbol Size depends on the chips used for the ECC.
How wide the E...
in bios it is called x4 x8 x16
so I'm testing either x8 or x16
I know physical bus is 8 bit wide
and ddr4 smallest burst is 8 too
so I suspect it is talking about burst equivalent
twin dew
Common per chip data bus widths.
x4 in lot of larger RDIMMs, x8 in most normal DIMMs, x16 on most laptop memory and smallest normal DIMM sizes.
Basically you need 4 x16 chips per rank, 8 x8 chips per rank or 16 x4 chips per rank.
- any ECC chip(s)
wanton orchid
I know x8 as almost same capabilities as x16
wanton orchid
Basically you need 4 x16 chips per rank, 8 x8 chips per rank or 16 x4 chips per ...
what, then why can I set that in settings if it is hardcoded rank dependent
the imc always know the number of ranks
twin dew
No, x16 chips have half the bank groups per chip.
x4 and x8 chips AFAIK have the same amount.
wanton orchid
I'm not talking about banks
I'm talking about ecc setting here
it is called x4 x8 or x16
twin dew
what, then why can I set that in settings if it is hardcoded rank dependent
And that was just about how many chips are needed for one rank of RAM.
1Rx8 is one rank of x8 chips (8 chips per rank + any ECC chip(s) as extra)
8GB DDR5 sticks are all 1Rx16 with just 4 chips on them.
And worse actual performance because of that bank group deficit.
But no, I don't know enough about that ECC Symbol Size setting.
But basically you are testing setting that not even server providers give any guidance on and just say to use Auto.
AMD documentation for K10:
ECC Symbol Size stuff on page 175 and some more generic ECC preinfo on previous page
https://www.amd.com/content/dam/amd/en/documents/archived-tech-docs/programmer-references/31116.pdf
Of course not directly related to current CPUs, but the newer stuff documentation isn't public.
broken stratus
twin dew
Nvidia had more mentions, but longer keynote.
edgy hazel
Yeah thank you Mr Data analyst
twin dew
Yeah thank you Mr Data analyst
Just as reply to bringing the same up, but without the mention amounts, just by the rate.
By lightning
So it wasn't really to you herg, just replied to link the image again.
night girder
π€£
willow pike
amd decided to cut all RDNA4 out just so they could achieve highest APM (ai per minute)
soft bloom
artificial intelligence points metric
wanton orchid
But basically you are testing setting that not even server providers give any gu...
So I end up testing it
to understand how it behave
because all I know is bigger is better for ecc sake
but I don't want to sacrifice on it
so still investigating if x8 is better
paper mentionned x8 and x16 have similar correction capabilities (even though of course x16 is safer and more reliable, x8 still offer similar tracking even if less reliable)
I want to avoid it being defaulted to x4 simply because more performance or something
twin dew
Part of that document:
A single symbol error is any bit error combination within one symbol. Both of these ECC codes are able to detect and correct any number of incorrect bits in a single symbol, to detect any number of incorrect bits in two separate symbols, and may detect errors in more than two symbols depending on the position of corrupted symbols. (The x4 code is able to detect 100% of errors in two symbols. The x8 code is able to detect 99.99999963% of errors in two symbols.
But seems to be mostly about if broken chip can be detected/corrected (chipkill) or not.
Which depends on if the two memory channels are working in separate or combined mode and what is the data width of single RAM chip.
(2x 64 vs. 1x 128 mode)
The tables on pages 175 and 176.
twin dew
So needs to be large enough compared to the per chip size and that channel mode, but smaller is better once that chipkill is reached as it allows recovering from more simultaneous errors.
Where if the symbol is too small, you get results lower in the stack.
But if the symbol is too large, you lose ability to correct multiple errors at same time.
1: The failure of a DRAM device results in errors to one symbol, and can be corrected (chipkill).
2: The failure of a DRAM device results in errors to two symbols, and can be detected but cannot be corrected.
3: The failure of a DRAM device results in errors to multiple symbols, and can be detected with a high probability but cannot be corrected.
DRAM devices wider than the symbol size are not recommended for high reliability or high availability systems, due to the higher potential for uncorrectable, undetected, or miscorrected errors.
Which also explains why server RAM uses x4 chips when possible for the capacity.
wanton orchid
(2x 64 vs. 1x 128 mode)
my assumption was : x16 -> 1x128 ; x8 -> 2x64
twin dew
No, more complex.
Like I said, check that section in that AMD manual.
But if it works the same, then that x16 mode would be right for x8 chips with ganged memory channels.
And x8 mode would be right for x8 chips with unganged memory channels.
But basically higher the x#, better recovery chance from complete RAM chip dying no matter what the exact RAM config is.
But less chance of recovery from spread out multibit errors.
And with smaller per chip widths, smaller symbol sizes still give that Chipkill capability, so you can get both that, and better overall multi-bit ECC
Exactly the opposite, x16 for two unganged channels with x8 chips.
And x8 for two ganged channels with x8 chips.
As in unganged mode you are storing two sets of 64 bits of each 128 bit chunk that ECC is applied on, in same DIMM, one after another.
So losing one chip makes you lose 2x8bits on x8 chips.
But I'm off to bed now.
wanton orchid
it keeps getting more clear then more blurry
good night
I'll reread "ganged talk" and stay on x8 for now as it still sounds like the sweetspot for me somehow
wanton orchid
AMD documentation for K10:
ECC Symbol Size stuff on page 175 and some more gener...
damn I'm understanding it all, I didnt not expect for it to be layed down so clearly
that's why I didnt read this paper before
this is super nice from amd ngl
it also explain the "syndrome" code I was getting when I paper clipped my ram channels
though if this does not impact memory performance, this is a pretty big fluck
wanton orchid
probably
pure karma
i feel like all of us in this discord probably have number syndrome in one way or another
languid gulch
THE NUMBERS MASON
languid gulch
sharp oasis
I want to see quad xfire 6900xt on ln2 for timespy
stray badger
I need to check cal fire
languid gulch
interesting quote from linus during WAN show tonight that i think should resonate:
"i'm not going to buy a $2000 GPU to turn on DLSS" π€£
excellent point there
twin dew
damn I'm understanding it all, I didnt not expect for it to be layed down so cle...
Bulldozer: https://www.amd.com/content/dam/amd/en/documents/archived-tech-docs/programmer-references/42301_15h_Mod_00h-0Fh_BKDG.pdf
Excavator: https://www.amd.com/content/dam/amd/en/documents/archived-tech-docs/programmer-references/55072_AMD_Family_15h_Models_70h-7Fh_BKDG.pdf
Unfortunately for Zen 1 and later, AMD hasn't been releasing that document publicly anymore.
But just more limited Processor Programming Reference (PPR)
soft bloom
isn't that true in base 2?
it's true in base javascript
twin dew
damn I'm understanding it all, I didnt not expect for it to be layed down so cle...
But that is the full documentation for the specific CPU models for BIOS/UEFI and OS Kernel developers, like the name says.
But for Zen 1 and later, it isn't public anymore, but behind NDAs.
That Bulldozer document also contains what that Error Thresholding does, and guidance on ECC scub rate etc.
And that Excavator document has much less on that ECC symbol part, but has section about Data Poisoning.
Seems Bulldozer to Excavator only supported unganged mode for the two channels.
And only x8 symbol width, so that only RAM with x4 chips could get that chipkill level of ECC functionality.
languid gulch
i managed to make a dell prebuilt temporarily useful by finding out that the 9700 APU in it could be swapped out for a 1700X
wanton orchid
I still don't quite get thar killchip thing
if Symbol is x8 and chip is x8
then it's 1 symbol per chip
no multiple symbol defects
twin dew
Basically how many bits are lost if one whole RAM chip dies.
With unganged mode, on x8 chips, you lose 16 bits.
With ganged mode, on x8 chips, you lose 8 bits.
And if the ECC can still completely repair the loss or not for the computer to keep functioning, just with all memory accesses creating corrected ECC errors.
As in that unganged mode, you have two separate 64/72 bit channels, where 128/144 bits are stored in two transfer cycles.
And in ganged mode, you have one combined 128/144 bit channel.
So if that symbol size is same as how many bits are lost on complete dead chip, or larger, you can still recover all the information.
But larger symbol size means that it is more likely that some other multibit error patterns cannot be recovered from and end up as uncorrected errors, or even undetected errors.
So balancing between more robust ECC against error sources that aren't RAM chip dying.
Or being able to continue functioning if whole RAM chip dies, just with constant errors.
So if I understood it correctly:
Larger that symbol size, more consecutive bits can be lost and still recovered.
Smaller that symbol size, more nonconsecutive bits can be corrupt and still recovered.
When all those nonconsecutive bits aren't inside same symbol.
So normally for best stability without possible uncorrectable or undetectable errors, you want the smallest symbol size, that still can recover everything when one RAM chip completely dies (Chipkill)
wanton orchid
so x8 in "ganged" mode
I don't know where to setup this mode though
wanton orchid
amd pushed for maximum reliability
which is nice
128/16 is pretty chonky
twin dew
But probably not changeable.
No idea which one is in use.
Probably ganged only.
On AM5 at least.
But cannot confirm yet.
Ok, so first dual-channel was ganged only.
Then Phenom II at least you could select one or other.
And Bulldozer was unganged only.
No idea about Zen 1 onwards.
But the move to that two independent channels was because while ganged mode is best for just one thread, separate channels is better for anything else, so almost all real life software.
And remember that DDR5 split it to two 32/36/40 bit channels per DIMM.
And LPDDR4 already had 32bit channels.
Which makes it so much pain that the details aren't released anymore.
But I really would expect that leaving that one on just Auto would make the AGESA optimize to just that specific point.
Or was it using x4 based on the MCE shown symbols?
Of course in your case, it is either 2x64/72bit, or 1x 128/144 bit
More probably that 2x unganged.
Unfortunately I'm not going to reboot my own server to check right now.
And I don't have access to another AM4 machine either.
wanton orchid
my guess is ganged and unganged is too simple
and you can have things like "ganged" which splits for smaller access
it indirectly simplify design as you main buffer behave as in ganged mode
and still gives you best of both worlds on memory bus
but I don't know how that would play with ecc
I guess ecc too can be 64/72
it keeps symbol boundaries
if you ever need both data
you can still combine after performing preliminary tests
after all in x8
it's 8+1
tribal kraken
wanton orchid
or that's what x16 is for ?