Level1 Diagnostic: Why Is 128 Cores So Janky On Windows?

[@sirius4k]

That's because the human eye can only see one core.

[@massgrave8x]

60 cores per second

[@marcogenovesi8570]

sounds like a skill issue, git gud

[@juntapiezas]

And Windows vision is based on movement... while those core are treated as static variables.

[@offspringfan89]

@@massgrave8x Cinematic 30 cores per second.

[@johnmijo]

@@offspringfan89 wouldn't that be 24 cores per second ?

[@bitcoredotorg]

Interesting situation with windows. Not a unique situation on AMD platform. We have a few Lenovo P7 workstations with a W9-3475X (36 cores = 72 HT), and windows 10 and 11 tend to break up the CPU into two groups, one with 8 "cores" and one with 64 "cores". No matter what we tried in BIOS, W10, W11, registry tweaks... The OS, and by extension the FEA software optistruct would see either 64 or 8 cores, randomly if we didn't try to pin it! Windows is NOT the platform for any form of serious computing. Pathetic, Microsoft... We are migrating all of these machines to Linux. 👍

[@Level1Techs]

Yeah, disable ht is the easiest fix. Sadly.

[@luizfcavalcanti]

Out of curiosity, which distro are you migrating to? With that kind of workload...

[@bitcoredotorg]

@@Level1Techs Is that a fix since it drops the 'threads' count below 64 in my situation, or fixes an issue with the windows scheduler? Thanks for the suggestion!

[@bitcoredotorg]

@@luizfcavalcanti Unfortunately, the software we use has limited platform support. RHEL is what we are choosing, as we use that on our HPC cluster already (IT admin experience). We don't want to jump to rocky since it's pretty new. We have no Suse footprint, so don't want to start one now, and we don't want to touch oracle products with someone else's 10 foot pole.

[@Lovedbychrist1]

@@bitcoredotorg it’s because it removes half the threads so you are under that 64 core limitation and windows schedules it as a single processor and single numa. It is a work around in hardware (disabling multithreading) that mitigates Microsoft’s scheduling issues. You are leaving performance on the table compared to Linux, but it is a work around that will provide consistent and frequently better performance until or if Microsoft fixes windows. it will also probably boost the performance on the cores as hyper-threading does require power and the Xeon W9s are power (if you have the cooling) limited like threadripper. Overall it’s probably leaving 20% of total performance on the table if you had a perfectly threaded and scheduled workload. The actuality of the figures is pretty hard to measure unless Microsoft fixes windows. Either way, Linux is more performant here.

[@steingat]

For people who are curious..... Licensing 128 core CPU system for Windows Server data center would be in excess of $50,000. There is a reason people do not have 128 core windows server systems.

[@CircsC]

That strikes me as the sort of thing you would present as evidence in an anti-trust case

[@cwspod]

U can buy windows server 2022 digital license for under 35 dollars online.

[@duckyatsea]

@@cwspodbeing able to buy it is irrelevant if it’s not a valid license. Most times, those are developer keys and you’re buying them from unlicensed retailers which can still land you in some hot water with MS

[@RowanHawkins]

​@@cwspodDeveloper license keys are not valid for production environments its part of the actual license agreement to get the developer keys. if you are buying in low volume many businesses go the embedded license route since they are selling a turnkey solution not a free use production system.

[@paulbrooks4395]

Additional background: In the very early days, multicore was the same as multi processor. Intel sold 8 processor systems for a while, and very old servers would have CPUs and RAM on daughterboards that would plug into a motherboard. In those cases, each CPU was very far away from another CPU's RAM. Thus, every CPU was a NUMA node. This broke down with true multi core CPUs that each had own-memory, leaving us in the rut we're in today. In the old days, multiple-anything was designed for what are known as "embarrassingly parallel" processes. Nowadays, compute is aggregated but processes are ad hoc and rarely parallel. In fact, the 4c processors are meant for non-parallel operations, generally virtual machines with small slices of the larger CPU. It's quite a reversal from where large systems started.

[@InderjeetSingh-im3eh]

Was not expecting Tarintino to make a cameo. 😂

[@tubaterry]

The Tarantino jump scare really made my day

[@ericthedesigner]

I felt blessed the day I was able to have 128gb ram on am4. I'm dying to get a threadripper pro for the kind of design work I do! Hail to bringing back the High Performance Compute System!!! Thank you AMD!

[@noth606]

Depending on what kind of design work you do, they might need your help soon or already. if you know how cores are laid out a bit, think about what kind of mess the routing of the interconnects are on a CPU module like this. Have you seen what happens when you have a complex circuit board trace auto-routed? If so, think of it but a million times worse, we will need to see some interesting innovation in how chiplets are designed, but even far more in how the substrate/interconnect is done, 128 cores is already mind bending but think what it'll be when more cores are involved, all needing MMU access, IO access, shared cache access etc, we'll need some non-flat arrangements very soon I think.

[@kenwilliams3279]

Yeh I'm running 128gb with a 5950x. I don't use that much RAM often but it's useful when I need it.

[@noth606]

@@kenwilliams3279 No VM's? I have an old Dell T5810, about to chuck an E5-2683-V3 into it and add 128Gb RAM for a total of 160Gb. Why? VM's, with that RAM and 14core/28thread I can reasonably virtualize the core systems of the whole company I used to work for to play with, and test various cluster message passing/load balancing systems etc in one reasonably sized tower box. Also adding a quad 4x M.2 PCIe carrier and some sort of shingled 20Tb or similar sized disk for backups. Your 5950x setup would be perfect for the same sort of thing, just a lot faster.

[@ericthedesigner]

@@noth606I'm voting for stacked chiplets with waterblocks in between! So the CPU will just be a metal block with 2 threaded water ports! How Awesome would that be!

[@donaldpetersen2382]

@2:44 Lol please editor, well done

[@FalconX88]

Microsoft is not interested in Windows (or Office) being used for actual work. Like there's this one warning if you change file extensions, something professionals do often, and you simply cannot get rid of this warning. Or like how does Excel not use full file paths for filenames? You cannot open two files with "the same name".

[@timramich]

That's because they're still putting important files right on the root of C: like a bunch of weirdos.

[@marcogenovesi8570]

@@manitoba-op4jxand phones are for the kids

[@Wheelman2004]

@manitoba-op4jx Ah yes, I remember being 13 and blindly hating on Macs because it was the cool thing to do.

[@TheFPSPower]

Linux users still daydreaming as always... Windows powers most of the world's "actual work" like enterprise applications, app development, all CAD/CAM software, corporate networks and email while linux runs mostly web hosting and virtualization, is that your "actual work"?

[@Masaliantiikeri]

@@TheFPSPower It's sad that most professional applications are only for Windows. "Actual work" users would have choice.

[@MegaJugganot]

I remember reading a paper back in college...i forget the details but the TLDR was above 70 or so cores scheduling algorithms kind of break down (with regards to performance) and you might as well use random assignment...

[@marcogenovesi8570]

probably a skill issue

[@feschber]

looks like windows took the breaking down route rather than random assignment ^^

[@MazeFrame]

Microsoft dropped the hint with "how to install Linux" last year, they do not care.

[@stefanl5183]

They do care! They care enough to hold back features like this, so they can coerce you into upgrading to their future version of windows that sucks and everybody hates but fixes the "problem". It's a feature not a bug, and it's a feature for them to leverage against you.

[@metallurgico]

They do not _core_

[@timmy7201]

_"Why Is 128 Cores So Janky On Windows?"_ It's Windows, the more cores the more janky-ness...

[@arnox4554]

"It's Windows 10/11" There. That's all that needs to be said.

[@metallurgico]

It runs great on my Pentium 4.

[@Lukas_Miglioranza]

Who knows how many times I will watch this video again! It's full of interesting stuff! I just wanted to add that the problems already start with the 16 cores and 32 threads of the big Ryzens (even if they don't seem very big here anymore!). The Ryzen 9 3900x and 3950x (3 or 4 cores on 4 CCXs) argued with the CPPC perf of Windows 10, which many times wanted the two preferred cores on two distinct CCXs, while AMD usually arranged its two best cores on the same CCX to make them work together and reduce latency. In addition to this, many times Windows did not take advantage of the wonderful and scalability that those processors offered and it could happen that Windows did not execute two softwares in two different CCXs, but only on different “random” cores. The situation changed with the 5900x and 5950x, and with the brute force of 8 cores on two CCDs that lovely refined scalability was not necessary anymore. Problem solved? Mh... Windows has a tendency (although it changes with overclock) to run on the second CCD only when the temperatures (not the frequencies, the temperatures!) of the first CCD becomes high, with the result that if you have a very efficient custom loop you end up with the second CCD in idle and the the first with all the 8 cores above 3000Mh just because the temperature is still low, even if you are in a multitasking session with some raw photographs, vector graphics, the video editor etcetera. Faced with the most straightforward and versatile multitasking, establishing cores-software affinities on your own becomes difficult, potentially dangerous, and in any case time-consuming. I always liked a lot the core affinity function, but Windows scheduling should however be better. I don't have big 7000 so I don't know, even because in Internet there is not a lot of this stuff. Fantastic! Thanks!

[@tommihommi1]

that article about the linux scheduler accidentally having part of it limited to 8 cores is incorrect, commit messages at the time show that it was intentionally limited because that function keeping scaling with more cores would reduce responsiveness of the system. It is correct that it wasn't documented, though.

[@timramich]

I'll try stick to low core counts and high clocks with a desktop as long as possible. They seem to keep lowering the clocks of the low core count stuff and shoving them down to the bottom of the barrel. It would probably be very feasible to have 4 cores at an insane base clock rate, like 5 GHz, maybe more. I prefer my desktops to be snappy. If I need threads to do a huge task, I'll do that work on a headless machine built for that purpose.

[@LainOTN]

In Windows 11\Windows Server 2022 affinity by default is set to all cpus on all processor groups. So That issue should not be present there. The limit of 64 bit is for compatibility reasons with old software (affinity mask was a 64 bit mask) Internally the kernel knows perfectly what to do with 128 cores, there is no limit here, the kernel knows about numa nodes, about affinity groups... it really doesn't care, jsut the affinity mask. to really understand why CS2 is not using all cpus you will need some performance metrics.

[@kennethmadsen6474]

Mighty impressive in a lot of aspects. But being able to film next to this 128-core system under load, without having to shout at the camera, that does it for me at least.

[@iaial0]

I mean, 360mm of AIO is a mighty cooler, but considering this can replace half a rack, that's impressive. Even if you just consider noise pollution

[@JimConner]

I don't think they allow MS employees access to multiple GPUs or Multiple monitors. Windows has been janky with multiple monitors since XP. For sure they don't let them dogfood on real hardware. Probably on VM in their testing operation.

[@HeyImGaminOverHere]

Windows do be like that sometimes.

[@EthelbertCoyote]

APO is showing me Microsoft is not doing as much with scheduler as it could. Don't get me wrong I know windows is s a mess of everybody wants to do things their ways and size but so much performance is being left on the table I am guessing for all processors.

[@lshallo106]

That cpu affinity window hat quite the scroll bar. You sure it doesn't list all cores?

[@stevekristoff4365]

umm. how about trying the SERVER version of windows which uses a kernel tuned for the number of cores/package architecture? Can't see anyone really using >=64 cores with a desktop OS. (then again I haven't used a 'desktop' for anything except 'play' VM's since server 2008 and before that mainly using server 2003 or xp 64bit).

[@powerpower-rg7bk]

With Windows, I'd be curious if disabling SMT/Hyperthreading would make any difference. Per thread goes go up and so do boost clock as core isn't as loaded as if it were running two threads and you the magnitude of thread management Windows needs to do goes down. Throughput of course is lower due to SMT attempting to do additional work on a core that isn't being fully leveraged by a single thread. Also wouldn't hurt to check to see if the pairs of threads for a single core land in the same processor group. While they should, wouldn't hurt to check given the quality of Window's scheduler. I would be interesting in seeing how the Epyc 9684X and its 1,152 MB of L3 cache handles Starfield. That games scales weirdly focusing on memory bandwith and cache sizes. As stated in the video, no one will buy an Epyc 9684X to play games outside of some KZbin crazies, but I want to know where the current performance ceiling really is. Ditto for running this game on a Xeon MAX and its HBM memory. In fact it'd be interesting to see if lowering core and thread counts on these server centric chips alters the performance profile for gaming. 24 cores/24 threads on an Epyc 9684X and power limits at maximum would be interesting to see through a desktop/gaming benchmark suite.

[@marcfruchtman9473]

I really like the overview, and especially the explanation about NUMA nodes and gaming performance. Thanks for making the video. (Yes, we do run games on Workstations tho!)

[@DavesGarage]

It's not janky in my office. You need to set the scheduler to core with BCDEDIT, but a one-liner would be a short video :-) On the same 7995WX with 96 cores, Ubuntu 22.04 scores within 1% of Windows x64. I don't get the problem.

[@Level1Techs]

Core vs root makes a bigger difference in hyper-v context; the screenshot showing the 3rd processor group with mixed numa nodes, which is by definition jank. Also > The hypervisor will use the core scheduler *by default* beginning with Windows Server 2019. < from ms documentation, emphasis added.

[@GraysonZimmer]

Man, is there something different about setting affinity on TR? On consumer cpus, I can set it easy for the program we make at work but in trying to validate TR, I just get access is denied when trying to set it from task manager. Even Process Lasso gives a "Ehh, I'll try." This is driving me nuts.

[@noth606]

This makes me wonder a bit about the architecture internally, in the sense that at some point you'll have to come up with a non uniform connection strategy for core to ram and start building chips with cores in a donut around some interconnect stuff and io on the inside and memory access on the outer edge of the donut along with io and start caring really where in the physical memory things reside for ramcoreio sort of topology mapping which is kinda cool. The place where that sort of things start to matter is a cool place to be, even if it will be a head scratcher for a bunch of things down on from that in the OS kernel, core and memory allocation etc. At some point I imagine that there will be a need for some form of "API" for the lack of a better word, to interface the CPU architecture and kernel/process/processor-core/memory allocation stuff to begin to not only limit the wastage of cycles/contexts on the CPU level but to have applications aware enough to restructure the way they operate, including the support in compilers to allow this. What a time to be alive, that IS something I'd want to sink my teeth into. Alas it's a walking people only club as odd as that may seem, wheelchair users need not apply. o

[@Reaper_1994]

@23:00 why does it look like there is a scroll bar on the top right of the processor affinity page?

[@EasyMoney322]

Last time I've used 2x20c with disabled node interleaving on windows, it decided to put 32 cores in one processor group, and the other 8 in the other.

[@height5558]

Incredible amount of information in this video

[@thousturtles2863]

I know this video was mainly about windows, but I got distracted thinking about this system with two intel flex 170's a tb and change of ram and setting it up as 64 gpu accelerated VDs.

[@iaial0]

The ultimate terminal server

[@StingyGeek]

Finally, something with enough power to run WordPerfect 5.1 for dos without lag!

[@GnomeEU]

"Starting with Windows 11 and Windows Server 2022, it is no longer the case that applications are constrained by default to a single processor group. Instead, processes and their threads have processor affinities that by default span all processors in the system, across multiple groups on machines with more than 64 processors." Some programs might needs some tweaks by the programmers but most stuff should run just fine with more than 64 threads.

[@benjaminoechsli1941]

21:08 Nothing makes me check back in better than hearing how others have checked out. 😅 22:40 Precisely why I've stuck to 10. Great video, sir. Keep it up.

[@Bob_Da_Bilda]

Thanks for using the music you did in this video, I've heard it before in a videogame, and it was a surprise to hear again. 👍

[@middle_pickup]

What are the chances Microsoft forces users to buy pro for workstations for this many cores?

[@StephenMcGregor1986]

DragonFly BSD has decent SMP usage, and the new 6.9x kernel in Linux looks to be decent too

[@ericthedesigner]

I can't tell you enough how much I love your channel! Please keep up the good work! I also watch the "Show" . To other youtubers out there, stop yelling at the camera and for the love of god purchase a $500 waist mounted steady cam. If your passion is making videos how do you not have gimbals and steady cam mounts?

[@OrinThomas]

Was this Windows 11 Pro or Windows 11 Workstation edition?

[@autohmae]

No mention of Windows Server ? Is it the same ?

[@complexity5545]

What a time to be alive! 128 cores. But the Wattage...dang. Wendell is my favorite tech guy. He keeps me current.

[@morosis82]

The wattage per rack unit is pretty insane, but the wattage per work done is also insane from the efficiency aspect.

[@sbme1147]

I agree. 3 Watts, you would think it'd be a little more having them all talk to each other. I agree, a very, very cool tech guy. I also like Tom Lawrence (LawrenceSystems), you wouldn't think his company is doing backups for Netflix. Been watching him for a long time. Both down to earth people.

[@OLDUSAFMedic]

I am hoping the next big thing is the elevation of transfer speeds on PC storage devices. All my time is spent waiting for files to copy or move. With the 1 and 2.5 gig internet services and network connections, the transfer speeds we have are very outdated. NVME's are nice but get too hot and after 45 seconds, slow to a crawl. I'm not investing in another machine until this changes.

[@GhostLeeds]

Hope you understand how much you teach the layman. Like i'm a desktop tech by trade, 10 years in, but you blow my mind. Thank you.

[@boam2943]

I have always had the opinion that process schedulers should be mostly a hardware block inside the CPU because no piece of software can know better the current situation of the CPU than the CPU itself. This would mostly abstract the type and quantity of cores but could also potentially be faster. Instead of software schedulers deciding on which core/thread a certain piece of code is going to run, they should just communicate to the CPU the relationship that that piece of software has with something else that is already running, if such relationship exists. Then the hardware scheduler could decide where it is run and what resources are attributed to it. Example: User launches a game. The software part of the scheduler (SS) communicates to the hardware scheduler (HS) "this is a new app and this is the code". While running, the game forks another executable which is a local game server. SS communicates to HS that forked app is related to the other, does not directly share resources but might need to communicate with it (so it should be kept in a reachable core to minimize latency). When the user is starting to game, the game server creates a thread (to control bots). The SS communicates to HS that this new code is a thread of the other app and it shares resources (so should be kept on the same or adjacent core). Technology must break away from legacy thinking in order to evolve. We can not keep treating current and future systems as if they are all low count heterogeneous systems.

[@kimp_ossible]

Is this applicable to Hyper-V? How do hypervisor interactions work with the Windows scheduler anyway? Is Azure running on this janky scheduler too?

[@Level1Techs]

Hyper v/azure is less affected by this. Ms sql server, if you could afford the license, is broken too because of processor groups.

[@farhadmodaresi4182]

lol I was just messing around on KVM to get a screenshot of 128 cores in htop :)

[@vccsya]

Can you try with the current windows insider (canary/dev)? Current build is 26085 (Stable is 22621, visible as 2263x) Many things have changed under the hood. Id like to know if that improved the situation.

[@TheRetarp]

This CPU doesn't seem like a good deal unless you have a critical need for this density. My one year old Ryzen 7900x build which needs a cleaning just put down an R23 score of 27,000 and some change. $10,000 for a cinebench score of 110,000 using 450 watts. vs $350 for 27,000 using 155 watts 110,000 / 27,000 = 4.07 Ok so about 4 times faster. $350 x 4 = $1400 150W x 4 = 600 watts Bottom line is this cpu costs nearly 10 times more for 4 time the performance and it's only 25% more efficient than an enthusiast desktop CPU. I understand needing ECC memory, the additional memory channels, and the extra pci express lanes, but this is a hefty price premium for those features. No wonder AMD as a company is valued so high right now, they have successfully implemented Intel's market segmentation approach.

[@GalvayraPHX]

It's a chicken and egg situation - no one is using it because if they tried, they'd have to get through all the jank. What I'm worried about is the future - right now, 128 cores is clearly server/workstation only, but wait some years and it'll eventually be desktop standard. And knowing M$ they won't do anything to improve the situation. Sometimes, I get the feeling they WANT to force everyone off windows...

[@yancgc5098]

HEDT is is still only 64 cores, and AMD’s workstation is just now coming with 96 cores. Having 128 cores as desktop standard in just some years is laughable to think about, it’s gonna take longer than that. Heck, it might not even happen if the industry finally finds a way to increase CPU clock speed substantially again

[@geoffmm2]

You missed that theres a scroll bar in the process affinity popup. It shows a larger swath of cores

[@Level1Techs]

Selecting them doesn't assign right still because of the processor glrpups/Numa node edge case tho

[@matthewhorwat7540]

I guess maybe that explains the other CS2 cities skylines 2) 64 core limited in the LTT video...

[@DynotoeTube]

I was under the impression that MS suggests Windows 10/11 Pro FOR WORKSTATIONS to support and unlock higher core counts and hardware?

[@harveyweizman]

What about windows server edition? How does it address all the nodes/cores?

[@Level1Techs]

It's a different scheduler out of the box but it doesn't mean. Huge difference except in hyper v contexts

[@harveyweizman]

@@Level1Techs The reason I asked is because I read a article on Microsoft’s website about 2 years ago, which I saved. Doesn’t this solve the performance issues on a system running windows server with more than 64 logical cores? “Behavior starting with Windows 10 Build 20348 Note Starting with Windows 10 Build 20348, the behavior of this and other NUMA functions has been modified to better support systems with nodes containing more that 64 processors. Creating "fake" nodes to accommodate a 1:1 mapping between groups and nodes has resulted in confusing behaviors where unexpected numbers of NUMA nodes are reported and so, starting with Windows 10 Build 20348, the OS has changed to allow multiple groups to be associated with a node, and so now the true NUMA topology of the system can be reported. As part of these changes to the OS, a number of NUMA APIs have changed to support reporting the multiple groups which can now be associated with a single NUMA node. Updated and new APIs are labeled in the table in the NUMA API section below.”

[@cluberti]

Is there a reason this is being done on a desktop OS and not Windows Server or something like Azure Stack? I'm not 100% sure how much change you'd get, but they do behave differently. This is odd, but I'm wondering if it's odd because desktop is tuned differently.

[@mkedzier123]

What Windows are your using - I would assume Windows Server? Or was is just Windows 11?

[@Zwank36]

love to see elden ring have a hemorrhage on that CPU, it does not like my 24 core threadripper (have to turn off all but 6 cores on 1 CCD for it to not stutter)

[@JerameatC]

So i have a 2920 processor and 1950x are they gonna get better results with 11 than over 10?

[@vh9network]

I'm glad Microsoft finally got the whole 32 core / 64 thread issue sorted out. 2990WX runs great in Windows 11. Not so much back when I first got it in 2018.

[@WMRamadan]

Are these builds made for someone, or do you use them for yourself, or are they sold after you create the review?

[@AndyDavis007]

??? The closest I ever get to Windows is checking in for an appointment at a kiosk station but I doubt there's more than four cores inside that kiss-all-your-privacy-goodbye disclaimers box. But I make sure I get some sanitizer on my hands right after. Can't have my Commodore 64 getting infected cuz I touched a Windows device and didn't wipe after. What was the answer again? 😬

[@Lukas_Miglioranza]

mh... I think that in reality you too, like aaall of usss, have been positive since the days of... shall I say it? Huh? Do you want me to say it? I'll say it! MSDos!

[@amigaworkbench720]

Can we get proper desktop Linux test?

[@jttech44]

The most likely use case is a linux server, likely with a ton of NVME stuck in front of it running debian and slinging data over fast network interfaces. Not sure why you'd whitebox that kinda thing, but, it's certainly happening.

[@666Tomato666]

@@jttech44 Because there are people doing fluid dynamics, 3d rendering, etc. on Linux.

[@amigaworkbench720]

@@jttech44 Yes Blender test please, gaussian splatting, and I would like to know how AI models work on this beast! What about pci pass-trough on Linux? I can't see anything interesting that Windows can do in 2024.

[@bobbastian760]

Never mess with a man who buys thermal paste by the barrel...

[@personaldronerepair6141]

3rd gen thread ripper here, 3960x. I would love to know more about the Silver Stone aio. I know that I am on sp3 and this aio is built for sp5. I also know that Noctua's new sp5 tower cooler will mount to the sp3 socket as well as the sp5. Does any one reading this know if this Silver Stone aio will mount to the sp3? I know this comes across as lazy but I can't think of any other audience that may have already looked into this may have a positive result.

[@Z4KIUS]

kinda off topic but looking at all these memory banks I remembered I'd love ITX form factor that sacrifices PCIE to support more memory channels to feed that beefy APU we aren't going to ever get

[@ShinojiResearch]

I wonder if it's mostly good for VMs, Cinebench wise it's about the same as my EPYC 9654 (96 core 192 thread).

[@cwspod]

Am running 64 cores, epyc 3rd gen , 64gb ddr4 , windows server 2022 and have no issues . Paid $30 for digital license.

[@idahofur]

I was going to leave more of a comment. But, most hit the nail on the head below. I can't say they are lying about specs. Just seems that anybody tries to push it. They run into issues for misc. reasons. Reminds me of IBM talking about mainframes. They had a long list of happy users. They also had a long list of fixes, bugs, or features for hardware / software. They had to decide what to fix. In my opinion. Microsoft doesn't care unless you are a big enough customer.

[@UnknownProductions0]

what about windows for workstations? or windows server?

[@Terran.Marine.2]

I had thought about windows server, also.

[@jamieknight326]

I’d love to run some of our dev workloads on this and see if we could saturate it :)

[@MadMeddler]

The task manager CPU affinity dialog is crap, but there's a scroll bar to the right, then you can select all the individual cores.

[@Level1Techs]

that doesn't seem to set the correct affinity on this system, though, when you go look at the process after. give it a try?

[@MadMeddler]

@@Level1Techs Fair enough. Alas I'm limited to 16 cores at this end 😉

[@keyboard_g]

BSD 14 added cpu core count up to 1024. LFG.

[@chebrubin]

It booted. All that matters. Are the cores allocatable with HyperServer?

[@MitchellWilsonII]

I was commenting a recommendation for Sliger and then you mentioned them. lol

[@keyboard_g]

Windows 11 is a consumer OS. Even Microsoft doesnt want it for the cloud. Benchmark it against their own Linux distro Mariner.

[@Primetime94]

Windows 11 installs fine with an fTPM. Are you saying this CPU does not have an fTPM?

[@Appleloucious]

Wow haven't watched L1Techs in a long time! You are looking fecking fabulous! : D

[@garrettkajmowicz]

It would be neat to see how high they could get the base and boost clocks if they were to produce a Threadripper CPU with only 1 or 2 CCDs.

[@michaelmoses8745]

I would intuitively say that would be a worse part. 96-128 cores in 1-2 CCDs would create a lot of heat in a relatively small area, probably beyond the ability of an AIO to cool effectively. It would also be a little harder to manufacture as each CCD would be relatively large. Tl:dr; worse performance.

[@levygaming3133]

@@michaelmoses8745I assume he wasn’t talking about 96-128 cores but instead threadripper versions of like the 7800x and 7950x. Still 8 and 16 core parts, just in a threadripper socket with the threadripper IO die. And basically it should absolutely be able to clock just as high as the desktop parts, considering it’s got a larger power budget (320-400w instead of ~185w) and a larger surface area on the IHS and even more space between 2 CCDs. Not to mention all that extra memory bandwidth. I don’t understand why it isn’t part of their lineup already?

[@jooch_exe]

0:10 yes and no. If rackspace is your only concern then yes.

[@GorditoCrunchTime]

Is it “math libarry”, “math library”, or both

[@tim5187]

Would be interesting to see some Factorio benchmarks on this system.

[@Alex_whatever]

Next video with this machine on www.youtube.com/@Level1Linux pls!

[@biomechanism1]

Hey yo, Ryzen 7000 with 196 gb ram is that stuff working now?

[@peacemaker9807]

Going out on a limb..... But I'd assume windows server has no such issue. Product separation. And other stuff.

[@jondonnelly3]

Old McDonald had a farm, A. I. A. I. O.

[@WeiserMaster3]

Even made a song about it! kzbin.info/www/bejne/m6bHgKqbaLetY5I

[@jondonnelly3]

Numa numa song

[@paulrallyfan417]

I'm hoping the 7995WX we have awaiting unboxing will not be a white elephant.

[@ericthedesigner]

I'm voting for stacked chiplets with waterblocks in between! So the CPU will just be a metal block with 2 threaded water ports! How Awesome would that be!

[@Elinzar]

Who squirted Liquid metal into that sage motherboard? , so lucky those pcie wherent full x16 lanes

[@dannotech2062]

Hi Wendel, I am working on a software renderer that uses AVX512 and I would *** love you *** if you would be willing to run my app your system. I've been thinking of buying a Threadripper for this but it's massively expensive and I'm doing this is a hobby, so I can't justify to the wifey why I would spend new car money on a CPU. Would you be interested in trying it out? I have a couple of videos on my channel of it running on my OG Threadripper and my 18 Core Xeon if you wanna check it out.

[@UmmerFarooq-wx4yo]

A lot of peripherals being designed for not windows compared to a basic android phone.

[@KiraSlith]

I could see using it for linux and data science tasks, Xeon Phi was cool for the same reason,and that RAM pool is bonkers (Imagine the absolute unit of a motherboard required to take full advantage of this chip). But let's be honest, this chip only exists for Amazon, Google, and Microsoft Azure. It's meant for Superscaler data centers for whom $15k/cpu is chump change and has their own proprietary fail over virtualization system over machine to machine 100gbit fiber channel.

[@GorditoCrunchTime]

“Show me you checked out as an operating system company without showing me you checked out as an operating system company” 😂😂

[@CaraesNaur]

At what point was MS ever really checked in as an operating system company?

[@BansheeBunny]

This was a fun video.

[@nunyobiznez875]

Nobody uses Windows to do anything important anyway, but this was interesting none the less.

[@WilliamTaylor-h4r]

You are makingloveeating food,you are the king of the interwebs.

[@BrunodeSouzaLino]

1300W? Eh. Considering all the expenditure and device capability, why not 2000W?

[@666Tomato666]

I wonder how fast would run CS:GO on Linux... it is natively supported.

[@Lazllb]

Im convinced Windows runs better in a KVM virtual machine so it can let Linux handle the scheduling.

[@gwenmorse8059]

Have you broken any parts on your Framework laptop yet? :D

[@jamesesdaile4696]

microsoft stopped caring about their OS years ago, when you had two control panels that was it. they never finished it, they don't care.

[@unlimitedslash]

Cloud Imperium Games need a few of these to build their new servers just so we can finally play Star Citizen with stable 30 server fps.

[@michelians1148]

Even almost infinite speed is not enough to make bad code run fast.