Everytime Wendell shares something niche with the internet (like the 3d printed fan adaptor for that server rack GPU) makes me wonder "how relevant does he truly is to the modern world?" Like the case of the d-pad library that broke the internet when the author pulled the rug in one of its dependencies.
I'm saying it on the context of "how many awesome things are possible thanks to the work he has put into creating niche utilities that are used by the backbone of some mayor global system
This might be a beginner tip, but I think its worth mentioning that on ubuntu installing nvidia drivers through command "ubuntu-drivers install DRIVER" doesn't always work. It works better to use "apt install DRIVER". Recently I had to update it for my gpu so just wanted to mention it. Maybe someone will save an hour :)
To whom it may concern: This message is regarding Wendel's engagement challenge at the end of the video. I shall refrain from addressing the explicit text of his query as I expect my answer should be self-evident. Such a brutal satire of the "joke" and, indeed, the very notion of "communication," is a game-changer; in a few words, you have turned my world upside-down. I now realize that the old adage about "sticks and stones" may be untrue; and, indeed, that the self-evident utility of listening to speech may not be worth the risk. I must retire to my panic room to think it over. See you in a few weeks and thank you, sincerely, on behalf of myself and all mankind, for bringing this linguo-deontological crisis to our attention!
I would love to have Idle efficiency mentioned, specifically long-term idle when there's nothing being computed (i.e. it's night, people are asleep), but model is loaded and GPU is stuck in P0 instead of going down, ideally to P8.
Im always interested in idle power usage too since most of the time computer waits for the user anyway. But its propably hard to compare due to dependency on specific driver version, program, os... etc. My Nvidia turing gpu isnt that well supported anymore on linux.
Level1Techs needs a cat or two. I'm too used to seeing Der Bauer's feline assistants whenever there's a desk full of high end computing equipment! A video on what/where/how to install the code-generating LLM you have here would be nice also!
This stuff is generally a bit above my head but more Linux content, any Linus content, is always great! PS: in the LLM prompt the second time you typed ‘numbers’ was spelt wrong, didn’t seem to affect the output though.
Thank you for the video!) Would be very much interested in a video covering machine learning on 2 * w7900 / 4 * w7900 dual slot on rocm, and how they compare to nvidia counterparts..
Isn't Post supposed to be all caps "POST" (as it's an acronym)? I'm sorry for being picky, but at the first glance I thought you're not going to tell how to fix the issue.
How hard can you stress it using vLLM for concurrent inference of multiple foundational models? Is memory bandwidth the limit there when vram runs out? I'd link, but YT would kill my comment. They have a performance benchmark too.
Would love to see a video on LDAP and web auth for self-hosted web services. I think there is a lot that can be done here that not many people are talking about.
Can you really only get 2.2kw from a whole breaker? I know yours only feed one outlet (which may have several individual sockets in the faceplate, each of which may supply a multi-way extension cord), but we can get 3kw from one plug, of which there may be several on a circuit in use at high power - our breakers are typically the power equivalent of what would need 64A to supply them at your voltage or 32A as they are in the UK on 230v. If you wire the device in, instead of using a removable plug, that can ALL go to a single unit of up to 7.36kw (but it will heat the room up). My concern is with the ultimate limit on how much power I can use in a setup that is quiet enough for deskside or desktop use without hearing protection, and how efficiently I can use that power to do fun things like AI (and including gaming during leisure time).
Pinocchio loved it! it would be nice to see a consumer & pro line evolve! but U don't want to see me after I hold my breath! ha..ha.. I'm fantasizing a hot swap bios, or CXL plug-in bios. duh...like a mini port...with a modified java egg beater.. wait what? then goto diner! sprig, sprout. Morrel...bread wine.... night..
machine learning this wendell, i got arch from scratch in parity with win 11 or exceeding it in every way, all my hardware works and all my games work, even starfield lmao, come at me when you want a real damn good os and ill share the package list
finally installed linux. on a usb. crossflashed an am gpu and back to benching in windows 😭 next time NVME's are cheap i think i'll make the system dual boot. all that $ sudo stuff wasnt too bad. need to get off windows if possible😅
2:25 editor places corrective text for 40GBs and removes the "8" from the voice, then 3:47 shows 48GB in the spec sheet. So many only have 8GB video memory, here 8's thrown in and out willy-nilly 🤣 E2A: 8 is just two 0s stacked ontop of each anyway, so there's literally zero difference, right! right?
Hi Wendell I've been following you since the tek Syndicate days, I would like to know how to connect this to a UI and how to create tokens that can be bought from what would be the users the end users using credit card or whatever like what the systems need to connect to a platform that is also connected to a client and that client is connected to the bank and you know like how would that work even the basic things that would need would be of interest to me, from the Hardware architecture to the software and all of the code that would need to be running, I know it's a lot so even just pointing me in the right direction for that would suffice but I believe a lot of people would be interested just to know like how does it work should I go and do it and so on, curiosity is the main driver for me to ask about this I'm a day Zero fan of the channel and by that I mean from the first hour of existence I've been watching.
What you are talking about is not AI, it is neural networks based on rather primitive diffusion algorithms. Humanity has not invented AI yet, it is just a marketing thing.
So you blow a shit ton of money on overpriced hardware and on all the power it consumes to get the code equivalent of 10000 monkeys with typewriters... sonders like a total winner move.
"What do you think I'm made of Threadripper Pro CPUs? - Only a little bit" - this got me, The self-aware humor Wendell is one of my favorite Wendells.
GPU rendered Wendell is superior imho.
"Asus, still in search of incredible" lol yes
Hi, I'm new around here. Didn't even know about nvtop and you have just made my life permanently better
Everytime Wendell shares something niche with the internet (like the 3d printed fan adaptor for that server rack GPU) makes me wonder "how relevant does he truly is to the modern world?" Like the case of the d-pad library that broke the internet when the author pulled the rug in one of its dependencies.
I'm saying it on the context of "how many awesome things are possible thanks to the work he has put into creating niche utilities that are used by the backbone of some mayor global system
This might be a beginner tip, but I think its worth mentioning that on ubuntu installing nvidia drivers through command "ubuntu-drivers install DRIVER" doesn't always work. It works better to use "apt install DRIVER". Recently I had to update it for my gpu so just wanted to mention it. Maybe someone will save an hour :)
To whom it may concern:
This message is regarding Wendel's engagement challenge at the end of the video. I shall refrain from addressing the explicit text of his query as I expect my answer should be self-evident.
Such a brutal satire of the "joke" and, indeed, the very notion of "communication," is a game-changer; in a few words, you have turned my world upside-down. I now realize that the old adage about "sticks and stones" may be untrue; and, indeed, that the self-evident utility of listening to speech may not be worth the risk.
I must retire to my panic room to think it over. See you in a few weeks and thank you, sincerely, on behalf of myself and all mankind, for bringing this linguo-deontological crisis to our attention!
Are you a bot? If not, are you OK?
neat videocard top is really neat
I would love to have Idle efficiency mentioned, specifically long-term idle when there's nothing being computed (i.e. it's night, people are asleep), but model is loaded and GPU is stuck in P0 instead of going down, ideally to P8.
Im always interested in idle power usage too since most of the time computer waits for the user anyway. But its propably hard to compare due to dependency on specific driver version, program, os... etc. My Nvidia turing gpu isnt that well supported anymore on linux.
My name is Wendell and I'm addicted to cores.....lol
I need 9000 cores
The 20A limit is probably from the 1895/1920/1988 power conversion designs vs the 2011 designs at 99+% efficience.
Level1Techs needs a cat or two. I'm too used to seeing Der Bauer's feline assistants whenever there's a desk full of high end computing equipment! A video on what/where/how to install the code-generating LLM you have here would be nice also!
14:02 nice oopsie there, can't wait for the review 😉
More Linux videos! Let's go!
This stuff is generally a bit above my head but more Linux content, any Linus content, is always great! PS: in the LLM prompt the second time you typed ‘numbers’ was spelt wrong, didn’t seem to affect the output though.
Thank you for the video!) Would be very much interested in a video covering machine learning on 2 * w7900 / 4 * w7900 dual slot on rocm, and how they compare to nvidia counterparts..
A quick vid or forum post on how to get from Lama to working locally AI on home box?
Does extreme demand= extreme supply constraints?
Super nice! It sucks that it doesn't post with the three GPUs, though.
Love the golden locks!
Isn't Post supposed to be all caps "POST" (as it's an acronym)? I'm sorry for being picky, but at the first glance I thought you're not going to tell how to fix the issue.
Power On Self Test
Sometimes acronyms become proper nouns/verbs. See laser as an example.
Power on self-test
@Level1Linux, when are you going to buy and benchmark a Tenstorrent Wormhole/QuietBox? It would be an awesome video!
How hard can you stress it using vLLM for concurrent inference of multiple foundational models? Is memory bandwidth the limit there when vram runs out? I'd link, but YT would kill my comment. They have a performance benchmark too.
Would love to see a video on LDAP and web auth for self-hosted web services. I think there is a lot that can be done here that not many people are talking about.
1:35 I had to install a 10A breaker just for a 850W psu, the inrush current would always trip the 5A electronics breaker.
k8s sounds nice in terms of what content should you make :D
0:00 _dives headfirst into the backseat_
Can you really only get 2.2kw from a whole breaker?
I know yours only feed one outlet (which may have several individual sockets in the faceplate, each of which may supply a multi-way extension cord), but we can get 3kw from one plug, of which there may be several on a circuit in use at high power - our breakers are typically the power equivalent of what would need 64A to supply them at your voltage or 32A as they are in the UK on 230v. If you wire the device in, instead of using a removable plug, that can ALL go to a single unit of up to 7.36kw (but it will heat the room up).
My concern is with the ultimate limit on how much power I can use in a setup that is quiet enough for deskside or desktop use without hearing protection, and how efficiently I can use that power to do fun things like AI (and including gaming during leisure time).
I think RTX 6000 Ada prices have come down a bit, in Europe at least
I wait for all the AI startups that are going to fail soonish. Should be a lot of nice hardware for cheap.
Is that a HAF 700 case?
I wanna see a render of a level one techs, links with friends show….but everyone is Danny devito
This video wasn't "Rack-Tastic" It was. "Off the Rails".
Papa Wendell, tell me about the shared exponents...
Pinocchio loved it! it would be nice to see a consumer & pro line evolve! but U don't want to see me after I hold my breath! ha..ha..
I'm fantasizing a hot swap bios, or CXL plug-in bios. duh...like a mini port...with a modified java egg beater.. wait what? then goto diner! sprig, sprout. Morrel...bread wine.... night..
So how much fps in minecraft??
Sooner or later someone needs to talk about those customised models that are being to do harm
I wonder how A800 cost nowdays because it has 10 TFlops of double precision.
Try the GreenArray GA-144. I44 CPUs (not cores)/chip.
machine learning this wendell, i got arch from scratch in parity with win 11 or exceeding it in every way, all my hardware works and all my games work, even starfield lmao, come at me when you want a real damn good os and ill share the package list
*A4000 are now 4000$??* dear god..
Lets be honest that is probably cheap
0:55 that spinning rust sounds like a angry frog
Racktastic indeed
Let’s talk AI and FORTRAN optimized hardware configurations? A little interesting math optimization?
finally installed linux.
on a usb. crossflashed an am gpu and back to benching in windows 😭
next time NVME's are cheap i think i'll make the system dual boot. all that $ sudo stuff wasnt too bad. need to get off windows if possible😅
2:25 editor places corrective text for 40GBs and removes the "8" from the voice, then 3:47 shows 48GB in the spec sheet.
So many only have 8GB video memory, here 8's thrown in and out willy-nilly 🤣
E2A: 8 is just two 0s stacked ontop of each anyway, so there's literally zero difference, right! right?
well we have 40gb and 48gb with a6000 and a800 so it's really more mixed than I thought lol
Oh sh, Wendell knows I'm a poser.
Hi Wendell I've been following you since the tek Syndicate days, I would like to know how to connect this to a UI and how to create tokens that can be bought from what would be the users the end users using credit card or whatever like what the systems need to connect to a platform that is also connected to a client and that client is connected to the bank and you know like how would that work even the basic things that would need would be of interest to me, from the Hardware architecture to the software and all of the code that would need to be running, I know it's a lot so even just pointing me in the right direction for that would suffice but I believe a lot of people would be interested just to know like how does it work should I go and do it and so on, curiosity is the main driver for me to ask about this
I'm a day Zero fan of the channel and by that I mean from the first hour of existence I've been watching.
no, it's "neat videoCARD top"
what an """expert""" you are
Ni...
Don't say it!
What you are talking about is not AI, it is neural networks based on rather primitive diffusion algorithms. Humanity has not invented AI yet, it is just a marketing thing.
Man, your most popular videos are from 9 to 4 years ago... I think you got too complicated for general population, lol.
So you blow a shit ton of money on overpriced hardware and on all the power it consumes to get the code equivalent of 10000 monkeys with typewriters... sonders like a total winner move.
First!