Threadripper 7995WX 96 Core: How's This System So Small!? Falcon Northwest RAK
Вставка
- Опубліковано 10 гру 2023
- www.falcon-nw.com/desktops/rak
Forum link: forum.level1techs.com/t/7995w...
**********************************
Check us out online at the following places!
linktr.ee/level1techs
IMPORTANT Any email lacking “level1techs.com” should be ignored and immediately reported to Queries@level1techs.com.
-------------------------------------------------------------------------------------------------------------
Intro and Outro Music By: Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 3.0 License
creativecommons.org/licenses/b... - Наука та технологія
Thank you so much for all the kind words Wendell! What AMD has accomplished with the Threadripper 7000 Series is just incredible. We're so focused on the early bring-up and benchmarking of these systems that we sometimes forget the bigger picture of what our clients will be able to do with these amazing new tools. So what you said here really floored us: "This is probably the kind of workstation on which general artificial intelligence is going to be invented. There’s going to be a sole proprietor billion dollar company, and this is what the founder of that company is probably going to use. Something like this- they’re not renting something in the cloud." You're always right (yes, even about that CPU 8-pin plug into the A6000 being OK!), so I really hope you're right about this too! Thanks again for the fantastic tour of our RAK system!
Such an insane little box, immaculately built! Super cool to see Wendell get his hands on something so awesome.
True that, and good work, but I was just hoping plain old intelligence in people were to come back in fashion and become more prevalent!
I’m so glad Wendell got a chance to check this incredible machine out! I’ve been looking forward to the day where I’d get to try a Falcon Northwest machine and call it my own :) keep up the great work!
Love your systems FNW, most of your catalogue might be a step up from my 12100 bclk overclocked machine I think
Hey guys, good to see you in the news -and- this space again! But hey! Stop monopolizing our ODM's time, dangit! You KNOW how hard it is to find a GOOD one! I got my own .. not exactly competing stuff I'm trying to launch! Maybe a peace treaty over colors and GPU counts? I just don't do the volume you do!
I wish a rack company would sell a case like this, it’s just the right amount of everything, radiator mounting, HDD hot swap etc
I think sliger has a couple of options that are close to this.
The case is a slightly modified Silverstone RM42-502, and you can find the 5.25" to 6 bay 2.5" hotswap bay online as well, so everything but the front panel piece.
The sheer joy Wendell has for bleeding edge tech is contagious. Excited to watch this one.
It's exactly why I love watching him!
@@undertone2472it resonates with all of us :)
Wendell, you sir are a scholar and a gentleman. Your enthusiasm is contagious and the information your relay to us folk is much appreciated. Thank you sir.
That chassis would be perfect to use with one of those desks that has an integrated 19" rack instead of file drawers.
Learning a lot from the server product highlights you keep sharing. I've been watching the channel for a while and now my professional life has become more and more geared towards server stuff, you're a treat for the uninitiated in the server space Wendell!
This workstation is truly epic. I’m struggling to fully utilize my 2950 X for actual tasks, not just synthetic workloads to keep it busy and here you are with 96 cores of pure powah. It tool almost 3 years but ive finally populated all of the IO on my Asrock Taichi. Hlad to see Falcon Northwest is still producing over the top systems, Ive wanted one for around 15 years now. Never had the cash to give them a try haha. Love your videos man, you’re what inspired me to build that 2950x system and to run UNRAID on it. Also running XCP-ng on my R720 with PfSense, for an older version of the forbidden router :)
Come on. You bought it to get chicks admit it.
I have been wanting to build a rack-mounted server box, but a lot of the cases I've looked at leave some things to be desired. This thing is absolutely awesome. I would love to get my hands on one (but I cannot afford anything close to this).
I've said it before. I love your passion for this. I hope you get to keep it.
Wendel multiple times in the video: "I'm not going to be surprised we we see artificial general intelligence invented in the next 5 years on a box exactly like this." Just what is Wendel up to with this hardware? He says it's just a guess, but I've heard his evil genius laugh a couple too many times to trust him.
I know this might sound like AI hype (and it very well could be), but there's a good chance of achieving AGI in the next 6 months to 2 years. I'd be surprised if it takes longer, given the ample training data available. While ASI is a tougher nut to crack, AGI _might_ offer valuable assistance. Governments dream of AGI for various purposes, like controlling free speech. Criminals handle the botting dirty work for them by accident, essentially putting governments in a tight spot.
Beautiful, thanks for this video. This is the perfect solution for my nex WS at work.🙏🙏🙏.
We are not billion company but we need alot of computing power.
really slick looking rack case too... to contain the insanity within.
Holey shit its been a while since I have seen wendle lv1 and hes lost some weight looking good keep it up
Love that case and the layout!
I started to call these processors "cluster on a chip" because it brings benefits of having many cores, but at the same time requires similar considerations you would meet when running workload on a cluster.
At the same time, i find such CPU in an awkward position because most workloads that benefit from going beyond 32 cores would also benefit from running them on a big cluster
Wendell, if you accidentally trained a SKYNET AI on the Falcon Northwest Rak, don’t connect it to the internet!
I have no actual interest in a product like this... But that doesn't matter at all when you've got Wendell delivering top-notch content!
Amazing , fun to watch. Cheers Terence Nelson NZ
This looks like just the rig I need for running a Windows Storage Spaces mirror for a pair of WD Blue 1 TB drives for storing my anime quickly! (Perhaps with enough threads left over to fire up an Lubuntu VM!) :) (What a gorgeous rig....!)
You’re simply the best! Better than all the rest!
This is what I need for my bigger plans
2:18 14:28 - GPU closer to CPU, power plug not seated correctly.
It was fixed for testing. It's only like that in the b roll ~ Amber
As a simulation engineer, it is cool to see COMSOL benchmarks!! Rocking last gen 64 core TR Pro at work, time for an upgrade 😅
The record in CB23 is 200k for now with one of these 96c CPUs. Bonkers.
the i'm filming skit got me in stitches
Buying now
"It can play DOOM . . . in the task manager"
Me : LOL wut?!
Falcon Northwest has been on my radar for decades because the were the biggest competitor to Alienware before they got acquired by Dell. Today, Falcon Northwest is the #1 of boutique builders in my mind.
Nice ❤
The Windows limitation about 64 CPUs comes from pretty old architectural decision from back then of times of the first 64bit Windows (for Intel Itanium aka IA64, maybe even sooner from some beta version of Windows for DEC Alpha AXP, later Windows for AMD64). BTW 32bit Windows had similar limitation of max 32 CPUs. Later, Windows lifted this 64 CPUs limitation by introducing new concept of CPU groups. Each CPU group containing max 64 CPUs. CPU group is not the same as NUMA node. For backwards compatibility with applications created before CPU groups were invented, the system is reporting max 64 CPUs. Newer, CPU group aware applications, need to ask the system about CPU count by using new, CPU group aware, API. Of course, in case of SMT (hyper threading), each CPU core is reported as multiple CPUs (Windows uses terminology such as processor core, numa node, processor package, processor group, processor die, processor module and similar).
this has apparently created an edge case in mark russinovitch's app wherein a numa node with more than one processor group needs special-case handling.. which is disappointing that the processor groups workaround needs a special-case workaround for what should have been an obvious next-thing-that-was-going-to-happen. Mark's program runs just fine with N-many numa nodes. Just not more than one processor group in a single node.
Cool stuff
I'm interested how these enterprise builds handle Meshroom. Running a 5900X/3080 build chugs while building SfM scenes and I'm looking at the workstation gpus, etc going forward.
Wendell is giddy - for good reason!
Looks incredible although most people are going to disagree with my one gripe. I think the chassis is too small, I mean I get it if you're setting up a render machine it would fit the bill. But to me if you're going to have a rack mounted server the chassis needs to be able to accommodate a lot of drives.
Question: Can you show a demo of a 70B LLM model running on this workstation? - Wondering how it compares to a homemade dual 4090 or 3090 workstation. Buy vs Build...
Compared to all the cloud vendors, where a developer would run their model in a cloud instance - versus a local offline workstation.
I do share this "Child like wonder" feelings for this amazing tech, unfortunate that my child like wonder is so expensive. Maybe some AI jobs could offset the cost ;)
This might be one of the nicest rack workstation cases I've seen, do they sell it without the hw?
Sorry, we only sell full systems.
The camera doing a lil bit 🤏 of trolling with that focus
*LEVE LONE TECHS*
Hi Wendel, can I hang RTX 6000 off the last slot? It looks like this motherboard has only 5 slots vs traditional 6. Hard to tell if the clearance is there.
Given that its a Rackmount, can we stack multiple Falcon Northwest Threadripper 7995WX's together?!
Yes you can! All airflow is front to back in these. GPUs that don't vent directly into the top (i.e. Founders Edition, Quadro/RTX Professional) would be preferential to keep heat transfer down if there's no air-gap between the RAK units (though an air gap is always preferred).
@@Falcon_Northwest Mad respect for the direct reply. This company is one of the best in the industry to ever do it!
Just wanted to point out the power connector on the top RTX A6000 is slightly loose.
The silverstone you are reffering to is the IceGem 360mm ? Love the video!
No, it's a custom 280mm by Silverstone with a very different pump configuration than the IceGem.
Do you guys ship to Colombia? if so the warranty is the same? THX!@@Falcon_Northwest
Wendell has a way of getting me very excited and depressed at the same time.
That blender graph only scaled to the monster scene scores?
I would love to see one of these 96 cores hash bitcoin
Wish I could afford this beast for ML from home, but got used mi100's for cheaper and better half-precision and double-precision.
Run Einstein@Home or Milkyway@Home with that beast
How good is this for NAMD ? does the system uses all cores?
I love the case but bit to much power for my home. Rack mount are normally so ugly this sits nicely in the home. But a threadripper for lite youtube would be massive over kill like 2 GTX 1080 in Sli was over kill at the time for 1080p gaming.
Can you share the cooling solution? I am running a 24 core threadripper (building one) and using the Noctua cpu fan cooler. I want to go water cooler on it, but after seeing this cooler you have, and not able to find it I was hoping to find that one. I am also considering the EWKB option.. they have some nice options too for custom water cooling.
Mad respect I'd like to see a sentient AI project with this😊
Please tell me you pushed that power cable in the rest of the way on that GPU :D
I'm guessing Falcon Northwest wants this mailed back... I can't imagine this being used for an HTPC.
🤤+ it's over 100K!
AMD really needs to hire ? and the Mysterians to do a revised version of "96 Tears".
Not sure if I missed it or if it was not mentioned, what is the depth of the chassis?
I'm just hoping for the Ryzen 9 7900X to drop down to around $300.😢 Droolworthy🤤 as this is, it's way above my budget!🤑😵💫
What Silverstone AIO is this? Looks like XE360-SP5, but compatible with SP6/Trx5. I want this one in my 7975wx build.
Came here to ask the same thing!
At 2:20, the inboard GPU Power Connector is not fully seated.
FNW make it look so easy.
That half connected GPU power cable at 2:15 triggers me a little.
Why cpu cables to those add in cards/GPUs and that's no issue?
There is rumors that in 7th gen TR chip there is a fuse, that burned when you activate overclocking, that will help AMD with warrenty cases with cpus
Those aren't rumors, AMD themselves were talking about it
@@SaHaRaSquad I hope. Because fuse for vendor locking exist in ryzen pro, TR and EPYC series and its very bad :(
I didn't quite understand your comment on NUMA nodes in Windows. You said it maxes out on 64 threads, so how did you solve it for processors with 32 cores and up? And aren't modern multicore systems usually setup as SMP? You don't have multiple sockets in the system, just 96 (!) cores on a single die. So the distance to memory is equal for reach core, right? So why then NUMA limitations?
Asking as I'm about to buy a 7985WX, though that's 'only' 64 cores, it may still suffer from the limitation (ie 128 logical cores).
Right now most programs on windows that can use a lot of threads work better in numa mode than they do with processor groups. Processor groups was added as a workaround for large numa nodes but lots of apps still don't implement it right
@Level1Techs wow, didn't expect an answer so soon, and on a Sunday! Tx so much 🙏❤️
Anyone able to find that cooling solution from Silverstone? AMD lists the socket as sTR5 but Silverstone only lists that heatblock under SP5 and its a 360 rad not the 280 shown in this case. I'm working on a Threadripper Pro build for work to do CT reconstruction and analysis, going to need a beefy cooling solution.
It's custom version Silverstone did for our systems.
WANT.
Omg, why TRX50 on 7995wx? This system has 192 threads but only 4 memory channels with ~150GB/s total memory bandwidth.
2:25 CPU ? for the GPU :D ?
Thats a CPU score.
Rtx a6000 uses eps12v not pcie12v, it's a workstation card thing. Ada6k uses the new 16 pin connector tho
@@Level1Techs thx 🙏🏻
👍900
96 cores in 3? 4U? Isn't exactly that much, there are servers that pack two EPYC servers into 1U node. Still, one hell of a workstation if you need to do some real number crunching.
WSL is a non-starter in a roaming environment. It's pushing us off of software like Altera Quartus as WSL users have to be admins and have to not roam as the whole WSL setup goes into the user LOCAL profile.. Quartus moved to WSL :
could I interest you in ssh? If you are willing to apply a patch to your gnome desktop, remote desktop runs with full-tilt h264 acceleration and it is amazing for remote users and this kind of use full-fat gui remote use case...
in 2032 Wendell will be playing 2016's Doom on the task manager in 4k if the cpu core race doesnt stall again between AMD or intel or (insert new ai Chipmaker)
Affinity Designer will automatically use those cores and GPU's
I'm utterly confused on the people who are rich enough to buy a Threadripper but too poor to buy EPYC.
Epyc usually comes in a server chassis - big and loud. There are many people who need a workstation sitting on their desk. How is the price comparison though? I thought Epyc of the same generation would run relatively pricier for the whole platform.
GPU on the right has a loose PCIE power connector. LOL.
Think it can run Minecraft?
Had to go configure one of these god boxes. $40k! Hey... anyone lend me $40k? Also, this is practically illegal levels of tech porn.
Try folding at home with it.
But.... How many copies of DOOM can it run at once
🤔😜.............
lucky SO*.
im triggered by the not-fully-insterted cpu-power plug in the gpu.
It being halfway in was fixed after the b roll, and as for the CPU label, it's fine for these Tesla's. ~Amber
Crysis?
dude those 2 gus will meld the CDs/DVDs after a few hours of crunching numbers, why would you put this thing in the living room ? only if u looking for a divorC
Lmao I gotta admit, 'agi will be invented on a machine like this' is far and away the most absurd thing I've heard in a youtube video about a computer.
Sound levels weak Wendall
YOUR DATACENTER IS PLEASANTLY QUIET. I'M A BIT JEALOUS.
Is he on diet?
let me guess: it's small because it's loud
you can't run TR at stock limits without being loud
someone didn't watch the video. Half the video at the desk was with the system on! And you can * barely * hear it in the mic. Madness!
@@Level1Techs if you can hear it it's too loud, if your mic picks it up it's WAY too loud
LC allows you to dissipate a lot of heat but it adds a quite significant floor to your idle noise level, even if it's robust enough to never get above that level under load
I know industry standards are "Ninja is too quiet, it should spin faster!" but that's far from what I expect of a well built machine, for me if I can hear it (including full load during summer heat waves) it's broken and that level requires either some really clever engineering and a reasonable power limits or quite strict power limits when using only aftermarket parts, with 7950X exceeding 100W of extended load using only aftermarket parts is a challenge, if you have some spare time you should grab some good fans (new A14 when?), good heatsinks, some accessories (like offset mount) and see for yourself
I'm sure in case of TR that's just gigantic you could increase the limit 20-30% when using the right cooler when compared to tiny 7950X but that's still far from stock, when you have such a high amount of cores even running at the top efficiency settings will push your power output above capabilities of most cooling systems
The L1 in Taskmanager is nice
Your cable isn't plugged in all the way. ua-cam.com/video/BhMQa9rEZFo/v-deo.html
! ƎƆIИ