Checking out a Data Science Workstation
ฝัง
- เผยแพร่เมื่อ 4 ต.ค. 2024
- Checking out Lenovo's top of the line P920 Data Science Workstation.
Channel membership: / @sentdex
Discord: / discord
Support the content: pythonprogramm...
Twitter: / sentdex
Instagram: / sentdex
Facebook: / pythonprogramming.net
Twitch: / sentdex
His GPU has more GB than my Hard Drive
Feels bad man :/
@@piipiipii Yeap, I have to remove my 40GBs "Homework" folder
@@tarsala1995 in HD.
I had to replay that because i thought i was hearing things...
just use the cloud
Intel Xeon "Gold": exist
AMD Epyc: I'm about to destroy this man's whole career
Linus sentdex tips.
@Beyblade420 Let Linus drop it.
Everyone flexing their gaming PC's until my guy pulls this out.
everyones gangsta until the multi-cpu RTX 6000 worstations comes in.
Hello person from 2035! Yes we were simple people back in 2019. We liked work stations, gigabytes and polar ice caps
I can't wait for sometime in the future when all this massive workstation is ancient.
And it's calculation capacities can be put in your middle class Ultraphone.
All this for a decent FPS in Crysis
Yeah could probably hold a solid 30fps at 720p serious.
...or - even better - in Minecraft. :D
Most interesting video staring at a case for 15 minutes I've seen in a while. It's so rare seeing workstation PC being discussed with almost every other channel being obsessed with gaming and more consumer grade hardware
I wouldn't be able to resist installing Minecraft on it before shipping it back to Lenovo.
Sentdex. Absolutely amazing. #LetHimKeepItLenovo
It's already gone and out of my life :( ...maybe one day there will be another. Cool channel btw, seen some of your vids
@@sentdex hey please DM on Twitter, I have a similar contact I can connect you with! @jabrils_
That GPU is worth more than everything I own
Yeah man
Dam this thing is Beast MODE!!! I've been building rigs for 15 years and you will always learn something new! Tech evolves so quickly it's hard to keep up!!! Kudos to Lenovo and this build!!! Talk about giving HP and Dell a run for their money!!
Well one of those GPU's is more expensive than everything combined on my desk... :0
Holy cow, no wonder he trains models in 3 seconds
Exactly
Your channel is amazing, and it helps me stay motivated in computer science
What is your power bill after 1 month of 24/7 training? That's additional cost.
I calced it to cost about $120/mo
@@sentdex Definitely expected more than that, wow.
@@sentdex so you pay about ~ 10 cents a kilowatt then?
@@sentdex Thats... Much less than i thought, wow.
I train my models on gpuclub com and don't worry about maintaining these monsters. No investment is the best investment...
Alright, thank you Harrison.
Next time my mom asks me to do a course in Data Science instead of gaming,
I'll show her this video and ask her to sponsor me this. XD
lol
They should allow you keep this beast. You deserve it.
That warranty alone should be around 30% of the total price.
We have one of these at work, chugging along 24/7 doing CFD computations. We equaled the purchase price within two weeks.
Server grade parts like this costs a bit, but taking in to account how much they’re used and than any downtime rapidly costs more than the machine in lost revenue. They’re easily worth it.
They’re noicy, but we’ve got it standing in it’s own room.
"Server grade parts": were you able to make them into a server? Are there any good guides on how to convert desktops like these into a server for ML web apps?
Should've taken an AMD Epyc instead of the Xeons.
wow amazing bro , I cant wait for the new projects you will run it on this workstation.
01:58 - Dell and Alienware have been doing this for years on _some_ of their workstations and gaming PCs.
I have a P920. I'm saving up to replace my two Gold 6136's with a couple of Platinums. I'm running only one GPU (GTX 1080 ti fe). I am VERY happy with this monster.
You can always send me a spare one 🤠
My 2008 Dell T410 used a shroud to guide air from intake to exhaust. It's nothing new. mind you the old dual X5675 processors just have passive heat sinks and rely on the case fan to pull that air thru both coolers. If I remove the shroud, no more cooling. But to say you've never seen that before, kind of suggests you've never opened a server or workstation before.. very common to use shrouding to guide airflow where necessary.
Its folding month at LTT. Wouldnt be suprised if he pumped out a few Work Units
You've never seen CPU shrouds? They're very old tech. I think you find them more in either old or compact office pre-built computers.
Wow sentdex you rock !!!
Killer machine!
I’m about to buy a Lenovo P720 in 2024😁
Dual xeon 6150 (18core x2), with 12 x 32gig RAM.
🇦🇺🤜🏼🤛🏼😎🍀☮️
My heart broken when you mentioned you had to send it back.
Ofcourse , it must be too expensive for lenovo too.
Hopy s*** Incredible pc.
Well 2Kb of RAM is enough. Apollo 11 did it so we can too..😂
I'm feeling old, i still know the old HP/Sun/IBM rules: If one person can carry it alone, it's not a workstation.
I hear 'ya brother. My media server runs 24/7. Building something reliable under those conditions is not the same as your typical kitchen table build.
My 2012 Mac mini with fusion drive ran 24/7 till now. Hdd now has 1% life left and ssd 60%
I could be crazy, but I was always under the impression it was cycled heat loads that cause damage to electronics, more so than static loads.
Can be used as radiator
Bummer there is no "Sentex Tech Tips store dot com" in intro... I'm so used to it... :(
I appreciate your video because it opens all of us up to conversation. With that said good find on the hardware platform, BUT if anyone is on the Data Science track you are correct stating that purchasing a (HPC) High-Performance Computer is the way to train data models as a novice. The hardware you require can be purchased at Nividia, or Lambda cheaper than 30K. I personally have built my own it's not that hard. For 30K you can buy an iMac Pro on the frontend for coding and have an HPC on the Backend.
Pls run it on MNIST dataset!!
Lol
Can you do a video where you just take it for a test drive with a previous network project? I want to see how much faster it does the learning process
Superb! Can you share what kind of projects you are going to run on this? Just to get an high level idea!
The thinkstation cover helps holding the cpu coolers for shipping too
Just a remark, the plastic covers on top of the cpus are very common practice on this type oh heavy duty workstation, Dell and HP have exactly the same. :)
Good video.
Great machine.. same style as dell and hp workstations. My current workstation is a dell precision with 3x GV100
Have you ever been to Pluckers on Greenville Ave? I'm really missing their fried mac and cheese. This comment is totally relevant to the video, because that machine is going to run HOT just like the wing sauce at Pluckers. I swear I haven't been smoking the reefer, just live 1000's of km away and your accent always reminds me of home.
nice landing pad for those cards
Please make a video on transfer learning 🙏🏻🙏🏻
5:23 lmao you caught me in the act I was on my second monitor looking on pc part picker
i can say airflow is a lot of trail and error. my old desktop is using the fifth iteration of the fan config and i blocked some vents and i finally got super great temps despite me hitting the old computer with blender all the time.
Very NICE!! Missed a pretty major part of the system... storage; what type, how much, etc ?
If you're a home gamer I'd say spend 5-6k and have an equivalent system. Or build two so you have a backup and still save a boatload of cash left over.
Im trying to get a job in this field. But I can only dream about putting down that much on a PC. What I got Lenovo t430 i7 16gb and an external gpu. I can only imagine what kind of data heavy algorithms the user who can afford this is doing.
Sentdex: "2X RTX 8000 GPUs - the best you can get commercially!"
Me: "I'm sorry, I didn't see you all the way over there from my 4X V100 GPUs."
If it makes you feel any better, that's one of my workstations at work; my home machine only has a GTX 970 in it (so it's pointless to break into my home!). The big workstation has a 3kW PS and the machine sounds like a swarm of angry bees when it's heavily loaded, so there's an enormous acoustic isolation/thermally transparent box around it to make the room tolerable (barely, so it's in a remote location offsite). The isolation cabinet is about the size of a coffin. Also, I don't get to keep it after my current project either (so sad), but there probably will be something 2X better at 1/2 the cost by then.
So how noisy is the P920 when fully burdened?
Just so we're clear... the RTX8000 is a better GPU than the V100. The RTX6000 is also slightly better. I don't know how to tell you how noisy the P920 is without a decibel reader, but I don't have one. I would not describe it as a loud machine, but I have a pretty high tolerance for noisy fans :P
I'm very curious about why this would have 2 $3,500 8 core CPUs. I get needing Express lanes and high core count and lots of ram but you're paying a very hefty tax there.
A single socket Epyc has 128 lanes, and you can literally get 24 cores for like... half the price of one of those CPUs on a single socket CPU that would give you more than 2x the PCI-e lanes, a complete fraction of the power draw, and way easier to cool. Hell, you could get 64 cores for the price that you paid for 16. Would it be clock speed? I just don't get where the money is going for those CPUs and would really like to know more. I personally cannot see a reason why you'd want 2 xeon golds unless your work loads favor those chips, but very few work loads do
@@lach888c2 That if a bit of a mind share answer. Epyc is a server CPU, it's literally designed with reliability in mind. I mean it's not my money so power to him but that's a huge chunk of money going nowhere.
@@lach888c2 all good, was gonna say lol
The Quadro RTX 8000 is only the fastest card but "fastest" is HIGHLY dependent on your workload.
If you're doing anything that requires double precision, than the Tesla V100 is still the fastest, topping out at 7450 GFLOPS (vs. the Quadro RTX 8000's 509.8 GFLOPS).
But if you're doing anything with half precision, then you would be correct that the Quadro RTX 8000 is the fastest card, so it REALLY depends on your workload.
the thud of putting that box on the floor in the intro lol
0:39 just like a tesla door =P
wow nice box! I want one! Great review!
Hello everyone. Do you know which of these CPUs will perform better in machine learning and data science tasks, no need to say I would use a Nvidia GPU like 3070 besides the cpu, but I wanna choose an appropriate cpu for these types of tasks. These are my choices:
1. 5900X: $250 (used)
2. 13600KF: $400 (used)
3. 13700KF: $500 (new)
But as you know there is another important factor, GPU. if I choose the 5900x then I could spend the extra money on a better GPU. If I wanna summarize both CPU and GPU configuration I afford are these three options:
1. 5900X + 3080Ti
2. 13600KF + 3070Ti
3. 13700KF + 3060 / 3060Ti
which one should be a better combination?
Finally you can play Dwarf Fortress as it was meant to be.
Air channels works. Systems like Puget Systems use them and so do most systems like this
but what is the hashrate :-)
I have two platinum with 2 tb of ram amazing and expensive very expensive around 33.000$ each
This computer is great but way too expensive for most people. I have build a machine with single 2080 Ti , and I am planning to add one more, is it worth to use NVLink, does it add extra performance?
You did not turn it on!? : ) Great vid as always.
We've already acquired what I'd call a "cloud mindset". If we need to do data science, we rent adequate CPU and GPU resources ad hoc. That is unless the data science one does is super-super confidential (but in that case, I wouldn't consider hardware from communist China). Or also unless one may precisely forecast the engine resource utilization for the next like 5 years and bother making its physical maintenance for that long, and as a result save a bit on AWS/gc/Azure, etc. :-) But for hardware-inclined guys, this is a great toy to play with, I agree.
Depends on the type of work you're doing. Most of the research labs I wind up working with, and even the work I do here, involves training models almost 24/7 for very long periods of time. So long as you're mostly a hobbyist or working on smaller problems, sure, cloud makes way more sense.
This power on AWS is $13K a month. If you're a hobbyist or doing smaller data, then cloud is obviously better. If you're doing real enterprise type work, which is what this machine is for, you'd be saving money *not* using AWS
@@sentdex I totally agree that AWS is going to earn some margin on that. A dedicated instance for 3 years will be a bit cheaper. However, it is fair to note that, for the enterprise setting, that AWS/gc/Azure price tag will include fees for the broadband Internet access, electricity for the hardware operation, electricity for cooling the premises (not a big deal, of course), and what is much more important for the business - the physical security of the hardware. Otherwise, the company normally needs to install cameras and hire people to ensure physical protection and physical access control. That won't be cheap for a relatively small business that is highly dependent on uninterrupted operation of their hardware and reliability of its connection to the Internet.
@@sentdex could a machine need more than 1tb on data science ? doing things like what?
what a balling workstation!
Are you able to do a video on what attributes of a deep learning machine affect neural net performance? i.e (GPU tensor cores, cuda cores, etc) Especially with regards for convnet inference? Sorry if this has already been answered. Thanks!
How did I know u got this a month ago but miss the video until now, I blame yt algorithms and my laziness for not searching up and checking
But can it run Crysis?
Would you recommend for mining ETH?
Try to run keras implementation of mask RCNN model with live video feed, if it runs well please do a video on that.
Do you have a link to the entire specification of the computer?
At what temperatures were the GPUs and CPUs running on max load?
I want a full spec one so bad, wish I had 20k lying around
Special thanks for translation to normal metrics(cm) from usanians :D
thank for the content,
192GB of RAM and 98GB VRAM, still cant get 60FPS with RLCraft and Ray Trace on.
This is the Linus Tech Tips video we deserved but never got
I'm a new subscriber and I do not know if you have ever reviewed a mac for data science. So, can you review something like a Macbook or the Mac mini for Data Science? Thank you!
Great presentation of that workstation. I just would like to know, what BIOS version it has, GUI based BIOS, or text based BIOS? Thank you in advance,
At those prices you might wouldn’t it be better thinking about a server rack? Or buy a boatload of tpu’s edge devices?
It's going to come down to the task. There will always be a way to build a machine for a niche/specific task better than this. This is clearly a general purpose data science machine which can do anything you want. It's not JUST for deep learning, or just data processing..etc.
Um I was actually waiting for him to turn it on.
Did you buy it? or they sent it to you for free? or was it just for a review?
-IBM/Lenovo Think* Series are always great.
-No one can build a machine close to price / perf / reliability to Think series.
-Doesn't matter how much VRAM RTX series has, Nvidia is still slow for running that amount of VRAM on RTX8000, might work on 8K Ray Traced rendering but.. for Machine Learning..? Use AWS.
-$30K not expensive when you consider you pay that money in a month for 2 ML engineers.
-Compared to coming NPU's NoC's, these GPUs are simple calculators.
AWS runs the V100. The RTX cards have more VRAM as well as execute more TensorTFLOPS than the V100. It's definitely a better, faster, card.
@@sentdex You don't get it. How many current real-life models require TOPS?
It's sad that those Xeons are not worth more than an ryzen 1700 in pure performance ratio, yet they put such a big margin on them :/ Other than that nice vid mate :)
you cant keep it? :(
I know, very sad :(
@@sentdex i thought you were the guy who earned a workstation like that
Plz make a video and show how fast it is by training a massive model with massive data.
Anyways that cool AF
the old xps 700's had a similar airflow thing. I wish threadripper boards had more memory slots to get higher ram support. Theoretically they should support 2TB but motherboards just aren't there. Kinda sad that new mac pro will be able to hit 1.5tb of ram while pc space tops out around in the hundreds of gigs. hopefully in this next gen of threadripper we'll get relatively high clocks and massive ram support.
I would be happy with a quarter of the ability that your computer has.
had no choice but to open it up.. LOL xD
Wish I could afford something like this.
Watching this video thinking to upgrade my pentium 3
That thing is huger than all my gadgets combined
But can it open more than 4 tabs on Google Chrome? I don`t think so.
Coil whine i think
re: airflow
Or...you COULD just run computational fluid dynamics and figure it out from a scientific/engineering perspective.
It's really NOT that difficult.
ready for 1000 tabs of google chrome