The WiFi on the motherboard we used for testing is WiFi 7, forgot to mention that in the video!

FYI: Not sure if you noticed, but those Video benchmarks that "don't show a doubling" seem to be limited to 16 and 32 threads respectively i.e. they're not anywhere close to loading a 64 core 128 thread system. A lot of video tasks don't scale too well with core count anyway - that's why there's software like av1an to split up encodes into parallel tasks.

Looking forward to the IOMMU tests and how the CPU lanes are laid out on the MB. The 2900 series threadrippers were already excellent in that regard.

I hope this pushes the 3K threadrippers to the 2ndary market cheap

Thank you Wendell! :) Really interested in your comments re: memory starvation.... if I remember back a few years ago your testing showed earlier Threadripper high core count processors had this issue when not populating the maximum lanes. Given the mighty strong and fast cores of ZEN4 - it will be really intriguing to see if 8 lanes vs 4 lanes is a significant factor for memory intensive tasks that exceed the large amount of cache.

Wendel is consistently delivering higher and higher quality videos with every upload! Great work! And yes, I disagree with anyone who says these new Threadripper 7000's can't game, they are a entire new paradigm for PC gaming! No console may ever match, not because of the fastest FPS on Threadripper, you don't buy Threadripper (HEDT & Threadripper PRO), Apple's M3 Max or Ultra, or Intel Xeon for that purpose, you purchase Threadripper and Threadripper PRO (and yes Apple Silicon and/or Intel Xeon w9's) for easily above playable FPS while doing literally a million and one other tasks in parallel!!!

Microsoft is trying their hardest to make Linux the desktop OS of choice by completely fucking everything up.

super cool, thanks Wendell

Finally some Linux news on these monsters!

Thanks for the video. It would be really interesting to see the threadripper plus amd workstation video cards working together in a system doing some design and blender stuff.

Thanks!

Are you going to do a separate video on how awesome the registered ECC memory is?

|*Literally*| some of the best tech content on YT. Thanks!

Good old virtualizing. I want to run Proxmox on a TR 7000 for virtualizing my servers and my desktop. So, Linux + Windows desktop (including a GPU each) and some light network stuff like pihole, vpn, Home Assistant. Basically everything I need to run my home. Also maybe a NAS. It should work, but I am not so sure about the power usage, like, if it is even less than with individual systems.

I'm glad you didn't forget to test Baldur's Gate 3 =)

No, I'm still rocking 1950X for my workstation :)

Wendell's looking good 👌

Wow, on my bucket list to own one of these CPUs.

The Orange L1 banner!

I would love to see a comparisom between 4 channel and 8 channel Threadripper / Threadripper Pro in COMSOL acoustic simulations. Idealy including paralelled batch simulations ( running multiple simulations at once to really push the CPU and memory bandwidth)

I would love to see "large" Docker (compose) projects build times. i have one project that has 13 services that each have their own docker file(s). My 7950X makes this a thing of about 15 minutes, but that's with partial cache enabled.

I'm a bit disappointed in 4-channel instead of 6 (EPYC Siena, also Zen4c, gets 6) . But getting RDIMMs on TR is a step upwards from 3000 series non-Pro.

Intel Quartus FPGA compiler works on Linux. The only problem I've found is a few years ago when the Memory IP generator messed up on Linux, while the windows version was able to generate the memory unit. For a large test job maybe one of the CPU samples would work for a heavy workload. The only problem is that the "lite" version is limited to single threaded. Note that this is less intensive than Open CL for FPGA, but single threaded my project takes over 30 minutes to compile. The process is memory limited, but my project is small enough to fit in the "3D cache" of the newer parts. So I am thinking that would be a better upgrade route.

Would love to see how fast Android AOSP could be compiled on these CPUs. It really scales well with more cores, although disk bandwidth needs to keep up

Would love to see a kubernetes/Openshift simulation/stress test

The 24 core 7960X would be such a good entry point into HEDT at $200 more than the top mainstream 7950X, plus the expected platform/motherboard costs. it costs $900 more than the part immediately below it in the stack. It's not far off triple the price. This just makes me miss CPUs like 5820K. That was a beast entry into HEDT back in the day. Decent performance and affordable enough to be bought by amateur enthusiasts.

Hey Wendell. Speaking of bleeding edge kernels, what do you do when your kernel is incompatible with zfs? On arch it's a nightmare (6.6, 6.3, etc). Makes me miss btrfs. 😢

In 1979 I tried to get TI to include ECC in the RAM refresh circuit. Since you have to read and write every refresh cycle, do a read, correct, write cycle. This automatically scrubs the RAM and allows ECC on access. TI Management Ignored me as usual.

Still sad there isn't a 12/16 core HEDT chip for those who want more PCIE lanes but don't necessarily need a ton of cores. Might build one of these next year but will also have to keep an eye on sapphire rapids W and TR5000 prices, if they come down after this is available might go one of those routes instead.

Can you bench mark LLama2 70 B and Falcon 180 B on thread ripper and thread ripper pro. i am very curious about memory band with (4 channel vs 8 channel) and core count scaling. The GUF format and oobabooga_textgen repo could be very useful. if you set the temperature to 0 they will be deterministic. How many tokens per second would it get on a couple of combinations should be enough to determine the sweet spot . if the systems are idle i can write a couple of scripts to run and benchmark these ai inference tasks.

@2:20, what happened to those poor innocent pci-e slots??

thumbs up for the linux

Doing a 7970X build right now with the Gigabyte AERO D TRX50 Motherboard.

vr testing on Linux would be fun

For large memory workloads, would it be possible to use TR none pro with CxL memory to get to around the 1TB RAM area?

I need to migrate my 2990wx to Linux. I remember running Linux on a 486/dx2... As a server when I was an ISP in the 90's.

What would happen if you ran your benchmarks on a better architecture? A direct execution stack-based processor like the GreenArray GA-144?

For a college senior project, a group of four of us clustered together 24 Dell P4 machines and used BOIC to calculate factorials of 1-2000 for a total of 4000 work units with a few errors. The script took less than an hour. I wonder what that fancy Threadripper's time would be for the same calculation?

Non-pro HEDT TR can return all it wants, but AMD should understand that not even The Braindead will touch this after that 100% insane move by AMD in EOLing TRX40 after one(!!!) CPU release, despite their "long term support promise", in that now infamous 'take the money and run' move.

Would love a video on registered ECC.

i have to be honest with you guys I do not really care about gaming on linux, overclocking or gpu pass trough. I am pragmatically just going to use the os with the best support for the software or hardware I am going to use if I am not bottlenecked by cost or a certain form factor. What I am trying to say I rarely turn on your channel any more. I think Wendell has so much esoteric knowledge and interests that would make for far more interesting videos than what he is doing now. I get that he is quit busy with work outside KZfaq but I would really love for him to make more content like that. Clearly he is not concerned about generating as much views as possible so please do the needful sir.

Do you know of any motherboards that use the 2011 Ćuk-buck2 VRM instead of the standard 1920 buck VRM?

Like to know about the linux suspend-to-ram story. The lack of it is a dealbreaker for me

Hey, thank you for the lovely threadripper reviews. One question. DevOps/SRE engineer with 10 years of XP here, not sure why anyone in my space working in AWS, Azure and on-prem would need such a machine. Have been running my 32GB i7 12700H Dell XPS laptop for about 1.5 years now and to be honest I don't think I would need more in the next 3-5 years. Linting and dynamic precompilation, docker containers and WSL, no issues. Could you please give some scenarios of why DevOps would benefit from such a workstation. Devs, for sure, Ops, not so much.

what happend with the PCIe slots.. they look damaged (or dirty?) was it tape?

Will these make a good virtualization platform with Proxmox? How do you benchmark that?

11:42 😧 What happened?

Any chance of a video on Thinkpad with Linux?

I would love to see a test of IO for pro audio/visual. I currently run 3 CamLink Pros for 12 HDMI feeds and tried to run a DeckLink SDI card and couldn't due to the audio latency on SDI output. Try to do all this while recording and streaming in OBS :D

"It's gonna get fixed with a bios update. Don't worry". Right. I've been waiting for those sleep fixes for years now. Still counting :)

Who would go for these vs the epyc ones ? It's a bit confusing.

did you forget to turn on the mic during recording for this episode?

Asus Pro WS TRX50-Sage and the Pro WS WRX90 are the best for Threadripper 7000

Wendel, how does this compare to the 64 core 3990X ? In a few short years I'm behind the curve again!

Video transcoding needs to decode as well, and decoders are pretty unthreaded last time i checked. Maybe try with uncompressed yuv video?

What about 24core?

TRX50 4 Channel memory wrx90 8 channel memory? I see wow great job amd and asus

Back in the day the top of the line Intel HEDT CPU cost about $2000 if I remember correct. But that was in the days when Intel could take whatever they wanted. One year later they had to slash their HEDT pricing in half due to AMD competition. So with that in mind I do not think the pricing is outraging

Yes, but the serious grown-up question is: do they still come with PSP backdoor/spywarez baked in?

@Level1Linux Wendell, would you pls run a video transcode job using Handbrake on Linux using these new Threadrippers? Details of the job: Source: Any 4K H.265 HDR 10-bit video file Target: 4K AV1-10bit SVT(do not use CUDA or any other acceleration, this needs to be purely a CPU Encode) colourspace:BT.2020 profile:Main Encoder tune:Off Level: 5.0 and up (MOST IMPORTANT WHILE ENCODING)Encoder Preset:Placebo(value 0 on the scale) I am thinking of buying a Threadripper Machine for these tasks, but before I buy I need to be sure about the speedup that I am getting. There are no benchmarks on the net that use these compute heavy settings to look for before buying them. Please please can u try these tests. Huge Fan by the way

Still salty that higher frequency, higher capacity DDR5 ECC RDIMMs become more common while ECC UDIMMs don’t even get the possible 48 GB per module upgrade :(

Not posted on Patreon first?

What about memory, does it behave with all slots loaded? Better or worse?

this video feels "personal" because of the charisma and camerawork (GOOD) but 2 camera agles nullify the effect a little bit.

Am I blind or can I not see a comparison between the new 64 core and the old TR Pro 64 core?

Please do a rdimm video

No 3D v-cache versions??? 😢

i can only dream about this CPU, cost to much !!!

Sound is weird. Has some background hissing.

Looks like someone sneezed all over the motherboard like 5 times.

but at those prices is it not HEDT as we all are used to and remember.. this is proper workstation stuff.

will this platform run ableton ?

Where did the Linux in 1812 joke originate from lol

Well it’s Linux so how about running Openfoam, code Saturn,SalomeCFD and Code Aster?😊 Pretty Please. I have an aging HP Z840 which back in 2017 was a beast, but my company went under, and I am semi poor, and need to upgrade. A thread ripper pro with a 100 gig ram disk seems to me like it would handle just about everything I could throw at it. I have 48 cores of Haskell the v4 chip I believe. I have been watching and listening for years. You are my tech God. Can 32 cores of threadripper pro be my answer to simulation heaven?

They should give you comsol for Linux

4 channel and restricted to only 1 DIMM per channel? YUCK! For that price I expect 2 DIMMs per channel at the very least, memory capacity is one of the key selling points for HEDT in my opinion.

Really doubtfull of these CFD results, generally CFD scale very well with CPU cores and tend to be quite optimized these days. I would like to have confirmation if you could do more tests and dig into it. You migth want to give a look at OPEN FOAM wich is a beautiful open source software for CFD and CLI friendly, a less known alternative could be Code_Saturne. But as you tell this beahviour could be a software or system, don't think so since you are in control, optimization wich could dramaticaly hammer performance. To my knowledge Comsol is more into FEM whereas Ansys is more FV (for real CFD purpose) wich lead to very different beahviour on CPU usage, it might be that comsol could scale more with memory (or cache) because directly solving this big matrices rely on RAM. I had to check papers, this should be a well known territory for computer scientists. Thanks for your work ! PostCognatum : Might be, that the previous pro version has much more compute units or are better used by the library used with the bench ?

350w? Nice of AMD to put in 64 efficiency cores.

i thought you're matt damon

Things like video compression benchmarks simply do not scale on their own to so many cores. They do not have enough parallelism to exploit it all. I am pretty sure if you look at cpu utilization, you will see it cores are not loaded super high. Others on 7980X are memory limited. 7980x definitively not great for memory hungry and parallelization simulations (pro and epyc is for that), because they will saturate memory bus with their super optimized codes, and basically get same result as 7970x, which also saturates the bus, but if you use stuff using less memory bandwidth (i.e. compilation / build kind of things, doing a lot of io), then 7980x can almost double over 7970x. CFD, I am pretty sure big part of CFD is dead on CPU. Everything runs on GPUs. A lot of CFD packages only support CPU, because their models and code are really complex, and decades old, so hard to port to GPU. But some do support GPU. Plus for decent results, you need to have double precision capable GPU, and most GPUs are really slow at this, or do not have enough memory to run big models. But if you get big HPC GPUs, then it is the best option. I am pretty sure if you are into CFD, then you will have way bigger machine (pro or dual socket epyc) or run simulations on the cluster / cloud.

i would say hell no just wait and get zen5 with 20% more ipc and better efficiency and build 2 or 3 and still save money - zen5 will have much better support and development - if you need more powah then go to server chip with 12 chan rather than zen5 cluster

AMD statement to every influencer - "Do not EVER compare TRX50 motherboards with TRX40". Want a shocking eye opener? compare every non-pro TRX50 previewed (of course, you can't buy any!) with the greatly extended feature sets of TRX40, in particular, asking yourself "Where are the Zenith II / Aourus Extremes etc??". For your answer, ask MSI - they pulled out of supporting TRX50, primarily because of how AMD treated every customer and motherboard maker of TRX40... AMD themselves said that "non-pro TRX50 is a welcome return for the desktop HEDT enthusiast". That is, exactly 4 years ago, after they publicly baited potential buyers with "long term support promises" for their then new TR system, TRX40, and then proceeded to lie to every TRX40 purchaser by breaking those promises in immediately EOLing it. No more expected CPU's. AMD literally killed their own non pro HEDT userbase! So, what about TRX50? Well, in truth, TRX50 is a shocking, feature-depleted embarrassing shell of what TRX40 was and stood for, a shell that not even The Braindead enthusiast would purchase.

So freaking glad cheaper Ryzen Threadripper CPUs finally returned - a lot of people have been clamoring to AMD for them for a while, it's good that AMD respected their wishes. I am not upgrading for a while though as Dark Silicon issues are getting somewhat bad particularly with TSMC 3N fabrication (Silicon is becoming useless the smaller the transistors get), so I am waiting for Ryzen Threadripper CPU to finally switch to III-V semiconductors transistor materials before I buy it to upgrade from Threadripper 1900x, this older Threadripper was and still is a cheaper option even brand new right now - I bought for work and play (I'd game on Linux OS as it got better CPU scheduler - I bought my Threadripper CPU more for astrophotography as 25+ MP RAW image files took 15 minutes on Phenom II 955BE processor to denoise). Still, the Ryzen Threadripper 7000 series CPUs are truly off the cuffs, 7995wx is as powerful as a decent video card, up to 96 to 128 CPU cores.

I'd be super curious about how llama.cpp's llama-bench performs (say on a llama2-7b Q4_0/Q4_K_M quants, 4K context (I do all my testing -p 3968 -n 128). Might be also nice to see how it does on a 33/34B model as the SOTA open code assistant LLMs are all that size atm.

Never knew level1 had a linux channel

No RGB? I'm out

sorry to say it but wendell you lost alot weight! since i watched last time!! well done Sir

I received an e-mail claiming that the Linux scheduler had a hard-coded limit of 8 threads. Is this true? How does having more than 8 cores help if there is a limit of 8? What is the current Linux limit?

I hope in 10 years these chips will be $10 on ali express like the old xeons

Couldn't disagree more. I have a Threadripper 3960x. It performs the same as a similarly clocked 3800xt in consumer workloads. My x79 and x58 systems before it all did the same. I was looking forward to the Threadripper 7000 launch, and was planning on buying a 16 or 24 core model at launch, but I was very disappointed. This time they have moved more towards traditional workstation. The consumer workload performance is terrible, falling behind even the Ryzen 5 7600. Just not worth it if you ask me. I'm keeping my 3960x for a bit longer and reassessing later once the 8000 series comes out. If someone would only make an AM5 motherboard that lets me use a full electrical x8 slot without dropping the primary x16 slot into x8 mode I'd be thrilled. The lanes are there! I'd even take gen 3 lanes off the 670E chipset for my x8 slot. The fact that not a single board maker is doing this, when the lanes are there and available is just so frustrating.

At the 10:00 mark you mention a xeon system, what are the specs of that system?

No RGB headers? So RGB RAM and GPU to compensate, lol

Speaking of CFD -- I would LOVE to be able to see what it can do with OpenFOAM.

There are some strange artifacts in your audio recording. Really distracting.

with those ripoff prices, HEDT is absolutely dead LMAO.

But why linux? It is less reliable than Windows these days. And now with SNAP it's twice as slow. Linux has been infiltrated and sabotaged by MS. They've won. Linux is dead.