Is The MSI Suprim X RTX 3090 A Machine Learning MONSTER?

[@vonwagner1257]

Love the genuine, unreserved excitement in your latest vids man. Keep on bein' a beauty bud.

[@FabiVoltair]

fuck yeah!

[@GuillaumeVerdonA]

So stoked to see these ML benchmarks from a tech youtuber. A lot of ML grad students such as myself often build hybrid gaming/ML research machines so it's nice to see both sides of the coin

[@squirrel6687]

You too? We are the minority. Dual 3090 here.

[@ops_tesseract]

@@squirrel6687 me too

[@ruanrossato9541]

I would really like to watch a video showing some entry, medium and high level machine learning setups. You are one of the few tech channels that talk about this kind of stuff and I really appreciate it

[@rick-kv1gl]

i am super depressed all the time. but 1 thing that helps is coming to this channel. thank you please keep em up Wendall.

[@vzxvzvcxasd7109]

How? it's not like this channel is funny, or tells you news that you can laugh at! just the nihilistic dooms day new since tek syndicate times! (it's scarcasm if you're american)

[@pililogan5769]

I am exactly in the same situation. My life sucks right now but Wendell being chipper, and always full of digestible but juicy tech, Ryan is always a hoot! His outlook on life really is the only sensible default. And then you got Kreestah who's pleasant vibe calms my outrage and anger to a manageable general disgust with the humans in charge, and the ones that put them in charge.

[@rick-kv1gl]

@@pililogan5769 exactly. it relieves me to see true people who love what they do and live a life they like :) give me hope

[@TaterRaider]

@Immortal Issues? Subscriptions, buddy. No half measures.

[@unskeptable]

How t prevent disk degradation?

[@MichalKottman]

Auto generated captions: "we're using the bun two".

[@henfibr]

"Machine Learning"

[@TheTylerHill]

i

[@Ryan-xq3kl]

@@TheTylerHill bun two gang

[@stevefan8283]

its actually pretty close in sense of pronunciation, "ooh/bun/too" -> "/bun/two". perhaps his accent isn't that recognizable by the autoencoder.

[@djoffski5712]

Twice as good as one. Tasted.

[@TechnoTim]

Thank you for this. I just picked up a 3090 and was looing for some benchmarking. I agree it's as easy as ever and as complicated as ever.

[@omgMBP]

This channel is on a real roll. Keep it up!!! I'll keep sharing!

[@dcphillips1991]

Molecular modelling is amazing, even with all the computational power we have though, we still have to cut corners as the calculation are amazingly complex.

[@Zarrx]

love your stove/backsplash/vent o:

[@davidn920]

Do I spy some "The Original" Glenmorangie lurking in the background? Solid choice :D

[@debdude123]

That oven joke confirmed he's a dad

[@carl8790]

Shutout to the editor for the meme edits

[@Level1Techs]

Wasn't me, but he is the meme lord between us ~ Editor Amber

[@SmokeytheBeer]

(rabid supporter here) Yay more machine learning benchmarks wahoo!

[@Dan-Simms]

Tell Ryan that I really enjoyed the kitchen meme.

[@Urhsg1373]

THIS IS WHAT I'VE BEEN WAITING FOR!!!!

[@emazzikanbodo3014]

The video 😋 or actually getting such a card ♠️ in your hands

[@sockaccount8116]

@@emazzikanbodo3014 The girl in his arms, of course

[@Mallchad]

"And here's one I prepared earlier".

[@yiannos3009]

Regarding the slide at 5:41 -- 3090 FP64 is actually around half a TFlop (556 GFlops). The Volta architecture has the highest FP64 performance of any nvidia consumer card. The V100 is at about 7 TFlops of FP64.

[@zombielinkinpark]

A100 in 54V system is configured to run 400W TDP instead of 250W. Most of the benchmarks were ran using those systems, A100 is not as efficient as you may think.

[@Tb0n3]

1:19 "No kitty this is my pot pie!"

[@dark4krad]

Classic Cartman look alike

[@jpb2085]

It's great to see you covering 3090 in respect to ML. Up to now been doing ML with a 1080, but recently purchased 3090 FE (still waiting for PSU to be delivered!). As a Masters A.I. student/part time consultant, doing experiments and prototyping on home PC, I wouldn't go the route of multiple GPUs and 3090 solely dedicated to ML. Would have loved to have seen some benchmarks presented here in vid (for easy consumption) for 3090 compared with say A100/T100 to fully answer "Is the 3090 an ML monster". Will def be comparing my benchmarks with your forum-published ones in next weeks once new PC built (upgrading from 5820K/1080 to 5950X/3090).Thanks again for content that really interests us ML enthusiasts (ex gamer!).

[@deephousefridays1911]

Was it worth it? I want to upgrade to exact setup from 1080ti/3600x

[@benjamintrathen6119]

I love the Kitchen Aid mixer, we also have one, so many cool attachments 😀

[@alexs6986]

Love it. This is what I've been waiting for!

[@Level1Techs]

Glad to hear it! ~ Editor Amber

[@morganbertram9024]

That kitchen is wonderful

[@sandraviknander7898]

What kind of optimizations did you guys do to get that dramatic increase in performance? Awesome content!

[@andreapalazzi670]

Nice content, waiting for the same video but with amd's RX6000 gpus!

[@stevedevney7322]

Good Luck with that. Tensorflow and ADM's GPU's don't play nice, unless it's tensorflow.js

[@andreapalazzi670]

@@stevedevney7322 Machine learning is not limited to tensorflow, I personally use pytorch with my vega 56.

[@stevedevney7322]

@@andreapalazzi670 very true. I didn’t know the pytorch was working without cuda on a gpu.

[@Spillerrec]

​@@stevedevney7322 I believe this only works on Linux btw, as AMD does not provide the drivers (and doesn't seem to have any plans to do so) for Windows or Mac OS X.

[@stevedevney7322]

@@Spillerrec does it basically work out of the box, or did it take a lot of customization to make AMD work?

[@FarmerKlein]

I havent seen push button light switches in decades. My grandparents house was the only place I ever saw them.

[@iyke8913]

Your kitchen doesn't have RTX on like Jensen's....

[@Winnetou17]

Now I'm expecting that on next GPU reveal Jensen will have a cat too. Or two. Flexing on this exact Wendel's clip

[@jonhuff9524]

Thanks Wendell! Your channel is tops. Please keep the content coming for ML-related stuff, there's a huge void on youtube. Incidentally, any preliminary thoughts on how the 3080 ti will stack up? I'd think it'll be pretty comparable to the 3090, just 1/3 the price...

[@Dracossaint]

This is gonna be fun, That 3090 looks very flavorful, was it sprayed with lemon juice or our wallets tears :p

[@MoodyG]

at 5:44 the table lists the FP64 performance of the RTX 3090 at a whopping 17.85 TFLOPS.. a far cry from its actual ~ 1.1 TFLOPS (1:32 scaling). I wish it was that good tho :D

[@marcelliusika]

Wendell flexing his oven, again.

[@ThunderKat]

Love the kitchen setup, but where do you get the water? Fill you only unbox half the product here.

[@seanpalmer8472]

Question: Was that container of spatulas and other utensils put there for effect or does it normally reside there?

[@paulthebeardedonedowning6820]

I prefer your selection of spatulas to Jenson's much more of a rustic collection I like that lol

[@snikelfritz3955]

Liked the video, and loved the outro.

[@pepperjackshack2439]

Just now getting into AI and ML with tensor flow and home assistant. That tensor docker containers uses all the ram I give it haha, 6gb and counting.

[@FabiVoltair]

you really did the oven! :D hilarious intro!

[@grimtagnbag]

Love the end stay till the end

[@btrmlk]

More videos like this please!

[@dinushkam2444]

interesting stuff, love these advance videos... thanks keep the great videos

[@Paranoidtimes]

what cooler do you have on the tesla

[@andrestyle2009]

Forget the RTX3090, im here to see this glorious kitchen. Just look how shiny it is

[@wojciechkuske242]

4:24 Looks like solar panel. Finaly RGB Led inside case get application!

[@producersRus]

Thank you for the video, i really appreciate the research support. Can you try using pytorch in the future. We often run into a lot of problems with tensor flow and kernel module compatibility.

[@TheShawMustGoOn]

Would love to see how it performs on NLP and NLU tasks!

[@evilgibson]

those in cybersecurity have gpu farms 100% dedicated to password/hash cracking. Depending on the length of time given by the client to do a pen test, we might only have 12 hours to and getting a crack on a system might allow pivoting to other machines. for example, if we can get domain controller access, a lot of stuff comes available if the admin logged into it recently and forgot to set the the policy to not store the admin hash on the machine

[@WandererOfWorlds0]

Can you please benchmark some variant of a UNet architecture?

[@gojiraforbreakfast]

Since monsters don't exist.....yeah it is a monster since it's never in stock! Love the content dude.

[@zencontroller]

I have a titan V, just bought an evga rtx3090 gaming ultra. This video is particularly interesting for me.

[@mcmormus]

I’m a student. I’ll write a thesis in Maschine learning. This will include training. Can someone recommend a card? Priority is price per performance. Used parts are definitely an option.

[@proterotype]

I’m 3 years late to the party, but I am one of those outlier viewers who wants more Machine Learning content with these GPUs. For what it’s worth

[@kifni41]

almost hit 'X' then suddenly the surprise ending showed up LoL

[@datapro007]

Terrific video Wendell, even if I am late to the party. I use my GPU for ML, not games. Many thanks.

[@somjio5775]

Would have been helpful to many If you included Pytorch benchmarks.

[@francistembo650]

You make good content Sir.

[@Level1Techs]

We appreciate that! ~ Editor Amber

[@TheGamingIndian]

Who is here after learning of the A2000? I made a sort of mistake getting a 3090, it's not bad but no certified drivers, extreme cost not as huge a performance jump as I thought compared to my older Tesla were bummers BUT I can easily game on it even though that's like 1 hour every Saturday.

[@nanolog522]

The machine learning team at my department bought a bunch of RTX 3090s, because they cheaped out, and will now have to put into the "setup" section of their papers, that they are using gaming GPUs.

[@DrNioky]

That's what everyone in ML has been doing for years though... The most used GPU in research papers is probably the 1080Ti.

[@nanolog522]

@@DrNioky That is probably true, before we used 1080s. And in most cases, that should be fine, if you don't need a maximum of performance.

[@Kabelman]

i got a 3090 and my company does machinelearning on price prediction of secondary markets and forex. I may be able to give some info on the performance there.

[@pililogan5769]

Do you visit the forum? Please do, I know that kinda info makes the community excited!

[@Kabelman]

@@pililogan5769 Not yet, but i may later, just a Lot of work right now

[@Auttieb]

You even have the spatulas!

[@jeffcav2119]

Great video!

[@shadowblack5455]

Hows machine learning with the 6900 xt and on amd in general

[@will000748]

I tend to get better perf using the nvidia container repo versions of tf/torch. Maybe also try with those?

[@SniperMayer]

Is that a Glenmorangie 10 behind you on the counter in the beginning?

[@REDACT3D]

Hum, back to hit subscribe as for some reason it's showing I'm not subscribed.. >.< keep up the good work bud!

[@paulomartel9618]

The A100 is superior in several ways but at a cost 6x higher (at least). I would say, for the home ML enthusiast the 3090 seems a dream come true. Thank you for the very informative video! :-)

[@MEATHEADBooYA]

I wish they would just open the market. I hope Intel does this with their new graphics cards. I want SR IO and be able to calculate server level stuff on a home pc. :P

[@Argoo41]

Hi! Can you make a video on budget ML setup?

[@prabhavkaula9697]

Do you recommend cloud services like gcp or ec2 over getting a 3090?

[@ashishpatel350]

When will people learn it's better to wait for the drivers to be updated than buy something at luanch. Either case they are hard to get.

[@kevlarandchrome]

Never.

[@pililogan5769]

that makes sense to me.

[@dmingod999]

You need to use nvidia-docker to run GPU workloads --- you cant use docker-ce directly. nvidia-docker uses the same docker but has some wrappers on it to be able to use GPU on docker

[@denvera1g1]

He said he would do a 3090 review in his kitchen, he delivered better than Jansen did.

[@JoeNguyen9x]

I need to call you sensei

[@mayurkulkarni1990]

The performance of the model training cycle is usually bottlenecked on input stage. Where are those images for imagenet datasets are stored? are they on ramdisk? Are you doing some kind of image augmentation on them? Using tensorflow using tf.data.Dataset and optimizing it helps a lot.

[@samuelschwager]

It's kinda ironic that we tried and succeeded to get away from 8 bit integers and now they are making a comeback with ML.

[@calaphos]

Sure the table at 5:45 is correct? I always thought the consumer chips only have 1/64 fp64 performance and not 1/2 like the A100. There Tensor cores of the 3090 aren't capable of fp64 and also gimped in the fp32 accumulate part AFAIK. Adding to the point of ease of use with nvidias software stack - There is even a beta driver for WSL passthrough were CUDA calls in a WSL VM are forwarded to the (windows) host Nvidia GPU driver. Works flawlessly on consumer GPUs and is frankly extremely useful as it allows both to avoid the windows development experience as well as the Linux desktop environment one.

[@somebrains5431]

There were complaint about FP16 and FP32 3090 driver nerfs in September. Did anything change?

[@syedalamdar100]

Can you do an install video on how you got RTX 3090 work with Tensorflow and have a decent utilization?

[@iyke8913]

Your RTX3090 double precision figures assume 1:2 ratio, when it's 1:64. Should be ~550GFlops

[@uss_liberty_incident]

I have no idea what I would even do with ML, but your videos like this help to give a better surface-level understanding of what otherwise would be a rather intimidating topic. So far the only non-gaming work my 3090 has done is waifu2x upscaling and folding at home lel. Thanks for another awesome video!

[@deephousefridays1911]

Lol, loved the stove reference

[@zockerimbett]

Could you do a video about ml on amd cards and the whole rocm stuff?

[@abc20723]

Im a PhD student working on antenna design and electromagnetic surfaces. Im building a Threadripper 3970x system for CST Studio Suite 2020. CST uses GPU for 3D modelling of antenna and also for computations when we are running simulations of our antenna design. Based on my budget i can either go for a Quadro RTX 4000 or GeForce 3070, which one would you recommend?

[@hwy9nightkid]

11:30 wat.... no pytorch !!!??? you google shiill Wendel! jkjkjk

[@sebastiangabriel3392]

What CPU you recommend for Deep Learning ?

[@particles1101]

I love your stove!

[@ginomoussa4344]

Which fractal design case are you using for your supreme 3090?

[@Apollo-Computers]

i want to learn how to do this stuff

[@itsdeonlol]

Wendell baking GPUs like Jensen!

[@DimasKurniawan]

hey bro did u ever tested deep learning on pcie 1.1 x16 ?

[@jaredkomoroski]

Which stove/oven is that? Do you have a thread in the forum on your kitchen work?

[@erikrosvall4387]

Can u explain how to install Rocm and use it? :) I did something wrong...

[@AvindraGoolcharan]

4:47 does it run well with the Titan driver?

[@dissident4226]

Aww, you be nice to the kitty! It probably just wants your attention.. lol

[@satanspy]

i only bought mine for videogames..not mining or machine learning or anything like that.. just for framerates.

[@bozidarprotulipac3771]

I hope you'll review rtx a6000 :)

[@zacharystreeter1965]

What you said around minute 2 is exactly correct. It's the reason why most theoretical physicist use Macs --it's Unix.

[@Dannicus117]

Very nice kitchen, by the way

[@ibrahemjoud1851]

Hi Amber have a good day!

[@jeffshee8969]

I would prefer doing ML on Fedora instead of Ubuntu. Although it seems easy to install the Nvidia driver and CUDA on Ubuntu, for some reason it's not that stable IMO. Still, you will need to do some config on CUDA after the first install, things are not really working out of the box on Ubuntu. Also, you definitely don't want to update the driver or CUDA, things may break. On Fedora, I never run into any Nvidia related issue on my setup. Enable RPMFusion, follow the Nvidia and CUDA guide, then you are good to go. No further config is needed since everything is working out of the box. Having no problem updating the packages as well.