I will ALWAYS remember the moment the CPU pulled 1400 watts. The panic! 😂 I loled hard. Thank you for this momerable moment❤ ua-cam.com/video/vU179_czCnU/v-deo.htmlh30m21s
Anybody working in a Startup/Software/Developer environment and used to the older Intel XEON workstations know this is extremely FAST and Efficient performance with triple the I/O available. And its the lowest end version. For Software Startups or small SFX/Graphic Design/Web Design companies this is a huge upgrade in a workstation.
When Pat Gelsinger mentioned that he now considers AMD to be in the rear-view mirror I think he confused the rear-view mirror with the front windshield.
This is a niche use case but I'm so glad you guys did some benchmarking using various molecular dynamics engines. I am so desperate to find actual benchmarks for MD packages, this is an absolute godsend.
It's always nice to see these tests, but for an avg. home user that do spent some time doing some of these tasks, no way a Threadripper system is worth the cost. It has to be pretty serious work for you before Its worth spending 3-4x or whatever the total cost runs up at compared to the top Intel or AMD dsesktop systems.
@@mrdali67 it depends on how much value you place in your computer in general. For me I'd go for threadripper if the compromise in general gaming performance weren't so drastic, purely because running particle simulations in Houdini is so slow with my poor 12400F and there are still games like Cosmoteer that eat threads for breakfast.
@@mrdali67These definitely aren't benchmarks for a typical home desktop. They're useful for something like the analysis workstations in my lab, though.
And then you can speed this upeven significantly more on these new Threadrippers by setting some NUMA parameters, c.f. Phoronix. It's crazy how powerful they are for compilation workloads
7:08 It’s so cool to see something like this because usually you’d have to dig in some specialized forum to get some comparison data for different CPUs. I think it needs a little contextualization tough CFD software are limited by memory bandwidth, the more the better. So higher number of memory channels and faster ram are the most important factors. That’s why there is a significant jump in performance between the 14900K and the Threadrippers in the OpenFoam benchmark. It’s also clear that the 7960x has enough cores to saturate the memory bandwidth, so higher number of cores provide a very little benefit. I’m not familiar with how the Rodinia benchmark works but apparently it’s used to compare different type of hardware (including GPUs), so it’s not a surprise to se some mixed results
APO or whatever could be the cause, which would be a shame to see. but gimping old products has been proven to work in the past to sell more new products *cough gtx780 *cough
@@limemasonAt the time we all thought Meteor Lake would be a desktop architecture too and not just laptop. In hindsight Intel should've never released a 14th gen to desktop at all.
Spec benchmarking was *extremely* helpful. Hope that info stays available going forward b/c it really helps in terms of deciding whether the expected performance increase is worth paying for. I hope you'll consider doing it on lower end hardware. If you don't need the threadripper IO and the extra performance doesn't have enough ROI, it's good to have a comparison between the 7950x and the 14900k. (And ideally a part or two lower down so that we can get a good sense of where the ROI stops making sense.) P.S. Happy to talk to you about the details of the benchmark stuff if you think it would help you improve your coverage. (At least the details within my wheelhouse.) P.P.S. Look up the pronunciation for "Poisson distribution". It's French and pronounced like Pois-son.
As impressive as this is, I'm glad to see that it doesn't stomp my 3 year old TR 3970X *too* hard (unlike what the 4090 did to my 3090's in Redshift rendering). My usual upgrade cadence for work machines is 5 years/4x performance boost (whichever comes first) at comparable budget levels, so I'm still good for a while longer.
@@chiefjudge8456 What are you talking about?? 3970X is the 6th fastest CPU in the Blender rendering test. And still faster than 7950X. Still relevant today I would say.
Steve, I really do appreciate you putting in a few game benchmarks. My system is about 70% video & audio editing / encoding / bluray authoring (Dolby Media Encoder, MainConcept's Encoders TopazLabs Video AI), & C# development (VS 2022), about 30% some gaming nothing amazing (BG3, WoW, etc.). So seeing it's not going to just wreck my experience is nice. I have been really looking at either the 24 or 32 core
I run a 3960X workstation so this is very helpful. BTW Poisson was a French mathematician - hearing it pronounced poison (as in toxic) while eating breakfast really perked my ears up!
To follow up on the how to pronounce part, Poisson, is not poison. Its stohastics metric, named after French mathematician Siméon Denis Poisson (/ˈpwɑːsɒn/; French pronunciation: [pwasɔ̃]). I belive it was in stochastics class in 3rd semeser CS university. Anyways. P as in .. poison O as in order. I as in ardent ( damn french) SS as in snake thats very hissy ON as ON something.
Makes sense that FFT would scale very well with core count, since it's a matrix factorisation and multiplication problem at its core, and that scales almost infinitely with core count. In fact, it's almost always worthwhile to do the computation on the GPU, especially with larger data sets.
When doing FFT, it may help to turn off extra hardware execution contexts if doing it on a CPU. FFT tends to fill up as wide as you can go, which is why it is so much better on a GPU. Multiple hardware contexts on the same CPU core is extra overhead and competition for resources. At one of my previous jobs where we were doing FFT heavy workloads on Xeon servers, the admins had Hyperthreading turned off specifically because the FFT heavy workloads ran faster this way.
It would be interesting but we all know Intel has no response there. What they can get away with on consumer oriented desktops won't work at this scale
Intel sells the Xeon W line as a counterpart to AMD's Threadripper. But I think AMD is ahead of those in price/performance and allout performance thanks to the crazy 96 core.
@@randomguydoes2901 That's the question, though, isn't it? Intel cannot compete with the 64 core and 96 core parts, but the 24 core part may be different. But because GN did not test it we still don't know.
I know you guys don't care for the 7950X3D as it gets left out of a lot of comparison charts, even if it was used in certain charts in a video it will get left off of others in the same video. I would love to start seeing it included in more testing especially in all of these novel tests.
I totally agree! It's annoying especially when they keep including the 7800X3D when no other X3D parts are shown. If they show the 7800X3D, they should show the 7950X3D as well.
Thanks for the Monte-Carlo tests! Would also be really nice to see temperature graphs. We got 3960x and 280W is no joke to cool, it also puts a lot of stress onto VRM. Wonder how 350W will do in that department.
@@kekoraaaa 99% of use cases for 14900k is gaming at 2-4 cores full load. 7960x will be 90%+ load for 24/7, probably in hot case with several GPU's, so VRM heat might be an actual issue. I wouldn't compare these two.
Thanks for the code compile test. I'm a software engineer and I'm glad to see a linear trend when going up in tr products, you will get better performance.
Always enjoy the informative content, and at last have subscribed to your channel. As an avid gamer, and engineer I've been looking to upgrade my home hardware and a 7960x might be just what is needed. Also, FYI @10:42 "Poisson" is the last name of a famous French Physicist & Mathematician (known for Poisson Equation and other achievements) not to be confused with toxic poison, or poison ivy etc.
Incredible, incredible work 👍 Even if other people are doing these tests, you guys are making it easy to find and using your huge platform to share this info. THANK YOU! ❤ Threadripper is amazing.
Remember Intel charged 2000 US dollar for a HEDT desktop CPU in gen 9. In gen 10 they charged 1000 for the equivalent CPU. And how many cores was that I do not remember but it was less than 10. And now we get 24 cores for 1500 us dollar. I would say that is a win for the consumer. And that Intel back in the days Price gouged the HEDT market. Just because AMD could no deliver anything to compete
Less than 10 cores on the $2000 9th gen HEDT???? How ignorant can you be? Lmfao. The 7980XE, 9980XE and 10980XE were all 18c/36t. You should probably use google to double check your information the next time you can’t remember.
@@phiphedude7684no you weren’t. The 3960x and 4960x Intel CPUs were $990 USD at launch…… Edit: And those are the top end skus, the 3930k and 4930k were 6 core CPUs of the respective generations and could be had for under $600 USD.
Love you guys, hope you keep doing the CFD tests too, as I know its a bit niche, but will be useful knowledge for me. Would be super interesting to see deltas between TR and Epycs!
I believe the name of the "Poisson" test and the Poisson equation comes from the French mathematician, so it's pronounced in a French way, kinda like "pwah-son", rhyming with "gone", and the ss is not a z-sound
I’m looking to get rid of my TR 3970X for a 7960X. I do reverse engineering, and CAD work but the programs don’t seem to take advantage of 32 cores. It seems like a new generation w/ 24 cores will suit me much better. I game a lot as well so it’s nice to see that it seemed to be an improvement there as well. Thanks for this!!!
@GamersNexus, the DXO photo editing software can use all cores for photo AI correction. Example my father is using 120 photos to correct, and we calculate in minutes and seconds how fast it terminates the correction process.
This is the Threadripper part I've been interested in due to the "entry" level price of the processor and the base platform. Still expensive but there is already an upgrade path to the 96 core part on a motherboard I'd get even if AMD were once again to forget about workstation. Hopefully they stick with this socket for Zen 5 and Zen 6 parts. As for the GN video editing workstation build, perhaps pair with both a fast nVidia card for CUDA based applications and an Intel ARC card for QuickSync? Best of both worlds? The unholy trifecta of red, green and blue teams?
This will be a fun revisit in 2033 to see how these wild workstation parts compare to normal desktop parts of that future date and whether games and other "normal" software will have become sufficiently multithreaded to utilize these currently-ludicrous core counts.
Ah yes, we've come full circle. You used to just get tons of IO on computers and it was just how it worked. Expansion cards for everything. You needed it for audio, networking, industrial equipment interfaces.. Etc. Then it all got built in and locked down. Now, hey, the professional market that needs that IO has the cash. It should be said those that need it are in a niche. Almost everyone here isn't in that category.
Not really. Adding in a second netwrok connection, be it wired or wireless, is not too niche, neither is adding 1 or 2 NVMe adapters, that is 2-3 slots, with a gpu taking at least 2. Most mATX boards already peace out, forget mini-itx, and even most ATX are done deal if you want rear channels like me. Then I just add an ascilloscope, which is too tall an order for most motherboards. I just want a barebones board with just 3-4 NVMe slots, just surround sound out E-key M.2 slot. Then I am golden. My choices instead boil down to massive highend "GAMUHRR" boards or just finding workarounds and adapters and whatever BS. Well HEDT seems to fit, sometimes. Just a 12 core is enough, I get bonus memory channels, and all the lanes I want. Sometimes there is even a mATX version of such boards, which allows me to keep my custom 2U case.
I seriously believe some of you underestimate how many machines it actually takes to "run all them simulations, and make the various media yall watch and such." Every thing outside or ones usecase or what they "see" everyday is a "niche" now 🙄 You don't bend over backwards for a "niche." Server/ workstation/ HPC has always been here.
@michaelscarportunless them chips have way more lanes than it's advertised to, I guarantee just "aren't noticing" the bottlenecks... or not necessarily using everything simultaneously for it to scream at you. If it was as simple "research your mobi better" we wouldn't be here today with non-pro Threadrippers again because that crowd (we 🤔) would have definitely sang at the top of the hills if we could have really made by with any odd consumer level platforms.
@@jamegumb7298 you have to admit that is niche though. Additional networking? Completely unessesary. Multiple NVMe? Professionals have network storage for large amounts. Storage servers aren't running thread ripper. "Ascilloscope" (oscilloscope?) hooked up to a computer is VERY niche. Surround sound out? It's emulated now over USB headphones or use an external DAC. But I used to have cards for all those things you mention. I get it. It's just not necessary for 99.95% of users. That's a niche. And niche markets are always expensive.
I appreciate that you test across a wide range of CPUs, but please put some visual highlight on the HEDT parts so they're easier to pick out. Most of us in the market for one of these cannot even consider the consumer parts regardless of how they rank in these benchmarks.
could you look into a re-review of the 7500f? it's been popular with budget builds in the eu as its an incredibly cheap entryway into am5, maybe having it into the charts alongside others would be useful
Great review, as this is the part I can most afford. I'm a 3D artist who will use the PCI-E lanes for GPU cards and GPU rendering. Would love to see you review the Gigabyte board to see if that is a good fit.
The thing that would really interest me is how much would running a VM with a given corecount of lets say 4 and 16 GB Ram affect the rest of the system for performance intensive activities like gaming that per se do not use or need these resources. These additional cores could be used to make a second PC that is being used as a server or something unnecessary.
With VMs, they don't actually lock the cores for their own use. If the VM is sitting mainly idle, it has barely any impact on the host machine's CPU usage. As a result, for a general purpose VM Host machine, you can easily get away with allocating to VMs 8 vCPUs per 1 physical core.
@@moist_ointment Sure, but my question is not even if the VM uses the cores and ram since I have more than enough left (8 cores and 48GB RAM. The question is if there are other "resources" that the VM uses/can use that will effect the performance of the system.
@jackchid6040 I would like to see VM benchmarks too, considering my job currently is to set up and maintain VM clusters, and we just best-guess the CPU based on assumptions. But the workload of a VM on CPU, RAM, and Disk IO (disk IO can be a major factor when running a lot of VMs) all depends on the current demands of the VM at that time. We comfortably run 40+ VMs on 16 core Xeons and are bottlenecked by only having 384GB of RAM of Disk IO long before CPU becomes an issue.
I prepared a large tank of copium prior to watching this video, as I already have a 7960X (but no motherboard as yet). Luckily, the copium went unused: my own guesses about the most appropriate Threadripper part for my workflow turned out to agree with GN's data.
I think it's the one actually that makes the most sense. It's got a few more cores, plus a lot more memory and expansion bandwidth, for a price that is attainable to those looking for a machine capable of that stuff. Ultimately I think the high core count ones won't really make much sense unless you go for pro with its even higher PCIe and memory connectivity.
RE: Squarespace ad... I'm old enough to remember when FrontPage, Dreamweaver, etc were new on the block and considered "cheating" when building a website. (I understood the FrontPage hate at the time with its massive markup issues behind the scenes bloating pages in an era of dial up, but tools like Dreamweaver comparatively weren't nearly as bad IMHO.) Now, this is just what we do. SMH
If only I had known this was truly around the corner, I may have waited. Instead I have been trying to create a home lab with VMs using XCP-NG, on a AMD 7950X, with 128 GB RAM. Well we will see...lol might have to trade up.
@@alexdomio4982 Naahhh, I'll get rid of my Phenom BE from 2010 before I'll get rid of my "old equipment". Doing information security research and relearning the basics of networking, filling in the gaps in my knowledge base from incorrectly taught garbage I didn't learn over the years. Basically, working on a Masters>>PhD degree in my spare time so that one day I can truly quit my "Day job" and enjoy living. 😉
Did you trade up? I kinda in the same boat. I'm waiting for a faster cpu to slot into the socket, not that the 7950x is slow but damn the 7960x can do ddr5 6000 on 48GB dimms. 4 x 48 and cpu overclocks to 5.2GHz all core rock solid. No harm to review but it dosnt show its full potential but I, just built the pc 4 months ago so cant really justify changing.
Yeah I hope AMD new 8950x or whatever they will call it that I can just drop in and ebay the old cpu on. Maybe even X3D on all cores with 6GHz Probably too much to ask in one gen. @@renzokuken2g
Great vid as always Steve et all. I would love to know more about the Blender benchmarks. This is EXACTLY what I'm interested in, and I appreciate that you're looking at it and not just games etc. "Even for those using EEVEE with CUDA, it's still impressive to see how much can be done with CPU rendering" OK but why? Why focus on rendering for CPU benchmarks? Is the CPU really going to matter when you have a 4090 in there and you're rendering on that with Optix? 1. if we're using GPU rendering, I'd be much more interested in knowing how fast the CPU did a liquid simulation (which I think is CPU bound?) Why are the CPU benchmarks focused on The Thing That Will Be Running On The GPU? 2. I would expect most anyone with a recent NVidia GPU would be using OptiX not CUDA, but same applies Thanks for the vid! and ps it's Poisson pwa-sson
For local code compilation for my development work, I wish I had a Threadripper system! Okay, I also wish the build server were an appropriate Epyc but alas... The speed is impressive - even more so on Linux with SNC/NPS tuning as per Phoronix.
Well in the video it is shown that the 24 core performs close to the 32 core, but for a significant price decrease. I think that justifies this product a lot
I would really appreciate if alternating row coloration could be used in the benchmarking charts, on more than one occasion watching this video I had to pause and visualize which row reference was which chart bars
Need to remember that these target businesses where they may purchase a few hundred or thousand units. If there’s no benefit to CPU tasks at higher core counts (or diminishing returns)-particularly where GPUs take the majority of the load)-then saving a few grand times 500 units results in a large savings.
Would it be possible for you to include idle power going forward? The added IO makes this really attractive for homelab, but not if the platform as a whole pulls 150W sitting idle.
Just a heads up, whenever people say things like "convolution" and "filters", they are nothing but matrix multiplications! Convolution and filter both are types of getting smooth data from a relatively grainy set.
Five years back I bought a 2965X 16-core gen 2 threadripper. At first I was very pleased with it. Then the Enermax 360 gunk problems occurred. Ugh. After the E(holes) claimed they fixed it I got another one. Fixed? Not so much. My choices became going full custom loop (not interested in the maintenance) or going with an air cooler (but it was build in a PC-011 dynamic case and hardcore air coolers didn't fit). Or use another 360 aio but the cold plates didn't give full heat shield coverage. I ended up with the Dark Rock cooler with a full coverage cold plate. While it does a fairly decent job, the 2965X at full boogie overwhelmed it. Not as good as the E(hole) 360 cooler (when it was working). It also meant I had to move my builtiful custom cabled installation to another case (Fractal R6). Bummer dude. So! I noticed you didn't seem to be using a cooler with a full coverage cold plate. Are there any AIOs with such a cold plate? Can an air cooler even think about 300W+ cpu? How much performance was being left on the table with such a cooler? Does one need a case that allows dual 420 radiators for the cpu? I recently upgraded to a 7950x which comes close to doubling the 2965X scores in Cinebench and appears to render DaVinci twice as fast. Plus it has a more standard heat spreader so I could stick on a lian li galahad II trinity which allowed me to built in my pc o11 dynamic again. I would *really* need a hard core video editor rig again before I'd pay the extra $$$$ for a thread ripper again.
Honestly the only maintenance I've had to do on the open loop keeping my 3960x cool, was a top up after swapping out an air cooled GPU for a liquid cooled one and stuffing a secondary pump in for a bit of redundancy because I could. Most maintenance problems come from people putting stupid coolants in with dyes that gunk up the works, a lack of biocide that gets algae growing in there, and mixed metals that just corrode everything. Stick with copper or nickel-plated copper, use clear premix with a biocide (and maybe a silver kill-coil if you're feeling fancy), and there's absolutely no reason your open loop shouldn't last as long or longer than any AIO, with the advantage that you can just put more water in rather than having to replace everything. Only disadvantage is, a good loop does not come cheap. It is a good investment though; one of those things that's truly "future proof", unlike people overspending on parts they're going to be replacing in a year just to get moar frames per second. If and when I get a new motherboard, RAM and CPU for a 7960x, the case, PSU and cooling loop is going to be a direct transplant. From everything I've seen as well, the TRX40 full-size coldplate will fit TRX50 as well, so I won't even need to swap that out. I also have dual 60mm-thick 560s for the radiators, and they're honestly excessive, so the $$$$ initial investment that's already paid for itself in terms of money earned using it, will be a "free" added bonus this time around.
Im still here i have hedt 3930k and it was about 500$ :D how times have changed.... Btw im still on that CPU.. its not the best but it still serves its tasks
4930K user here. They are solid chips indeed. Being someone who is doing some heavy rendering (ok, all the heavy rendering) ol' buddy is getting "demoted" to primarily gaming/ semi-server / whatever other mad scientist stuff I can do with it lol. (If I don't sell it. Have you seen what a working X79 parts go for now? "Affordable" HEDT is definitely missing and in demand.)
well hat of to you sir :D they were calling me a madman in 2012 you don need them 6 cores :D but i knew that intel would stall the development at that time... :D hahahab ive also changed 3 gpus on this beast and still goes strong on 1080p@@MaddJakd
@mit4c They loved calling us mad back then. I wanted a "best of both worlds" machine (plus film school lol) "It wasn't possible" they said. They still try to say that too, but I don't remember agonizing through 10 fps when I did get to finally play lol. The move now is moresoe because the footage is getting into that territory of "More core were required as is" and its finally fighting me for being stubborn. A new GPU or 2 and even only the most sensitive of frame hores will notice honestly, even back then. When this mobo used is going for the same price as it was fresh off the shelf back then, prettt sure more than just us crazies are seeing the utility in these "old" HEDT platforms.
Question: When LN testing, would it not provide better thermals to have a taller tube of LN over the CPU? Ideally you want the boiling layer over the CPU to be as thin as possible and since liquid pressure is directly a result of depth and not strictly volume of a container, would a taller cone not create more pressure and limit how large the boiling layer becomes under load? Of course with this it may be beneficial to create a cage to limit how much LN is thrown by rapidly boiling LN. OR, is the compression negligible and doesn’t change the gas layer much? Strictly asking about pressure via gravity and not pumps.
Thank you for reviewing these. I am *much* more interested in the 16-32 core processors (ideally the pro line with the more memory bandwidth) than anything greater than 32 cores (for that I would go with server products thanks to the 12 channel memory). That being said AMD *really* needs to work on pricing here. the 4 channel parts (non pro) I don't really see much of a use for, but even then for either the performance/cost metric for these are completely off. There's no way to justify a 300-1000% increase in cost for cpu alone and have performance be lagging by this much. realistically the prices should be halved if not more if considering the pro versions) i.e. I can see 100-150% increase over the same core count (16) 7950X desktop part and then scaling price w/ performance above that, but *NOT* what we have here.
Yeah it seems really dissapointing. You can easily build 3 7950X (48 cores) systems and configure them as a compute server for less money than it would cost to have an equivalent 48 cores build on the threadripper platform
I have to wonder how much of a difference there is between Windows and Linux using these CPU parts? I thought at least at one point Linux was doing a lot better because it got support to manage the multiple CPU dies effectively first. I suppose this brings up another point, at least historically, it has been considered folly to load up Windows with too many different things. I remember trying to just get my software development stack all on one Windows computer and at first it was taking several minutes to boot long before getting the full stack on as each new piece needed a reboot, and then Windows would split apart at the seems and BSOD on boot before I could finish installing my development stack. I had to spread across more than one computer just to develop software applications and my software development stack was fairly rudimentary, at least compared to what I was doing under Linux at the same time without any real issues getting it all going, plus with Linux I got it all setup with just a single reboot to switch over to the kernel on the hard drive. Especially with all of the crashes and reboots, just a fairly basic software development setup under Windows reminded me of an old IBM commercial where a guy trips and drops his deck of punch cards and then the next scene he is getting ready to hang himself. I would think this also limits down some the use cases for these CPUs for Windows users. In the Linux world, it is more common to load up a system with lots of different tasks and even lots of VMs and other containers such as Docker containers. Especially for home lab stuff, I could see getting a 24-core CPU like this because you need something reliable and with the I/O capacity to have some expansion cards besides a GPU and this is at an "entry level" price for such a setup. In other words the normal use case is you have a bunch of different things happening on the system under Linux and maybe some Windows VMs going (Windows in a bottle where it belongs), but then as a home system, you sometimes also game on it, especially as plenty of video games run under Linux natively or maybe you give a Windows VM direct access to a GPU. Trying to spread across a bunch of systems for a home lab setup can get really expensive, you need to have a 19" rack for it all, and then the power bill is insane as well as having to pay an electrician to wire up the power to deliver to all of the hardware in a 19" rack. You instead consolidate all that you can to one physical box and while it might end up being a beefy box, and it is unfortunate that Caselabs is not currently around to make a good case to stuff this all in, you can run it all off of the 115V outlet that was already there in your home and call it a day. You have to consider that a commercial grade 115V, 15A UPS maxes out at around 1,200W because it needs 300W for charging the batteries, you don't want to push more than 1,500W continuous through a 115V outlet, and for a home lab setup you need a commercial grade UPS because consumer grade just doesn't cut it. So with just the CPU taking up to 350W before conversion losses and the GPU taking up lots of power as well, maybe even more, that is over half your power budget powering just two chips in one computer. While some have pushed more into the cloud instead of having a home lab setup, on the flip side cloud services have been getting rather expensive, so then this pushes things back to running on a home lab setup, especially if you can get a CPU like this and have the flexibility to have 'elasticity' with 24 cores right there to either run broad and fast for parallel tasks or divvy up when you effectively need a bunch of different things happening at the same time. Going for a bunch of smaller systems has the problem that it is harder to go parallel effectively when you need to for a single task. I mean SMT (symmetric multi-threading) is the "quick and easy" way to go parallel and make quick work of that big task. Trying to get a bunch of small boxes to divvy up a big task and get it done quickly is a lot harder. So often with the many small boxes approach you end up with some things taking forever to do or maybe you can't get the resources together in place to get it done while the one big box approach it is easy peasy and you still have a bunch of smaller things going effectively using the one big box as a bunch of smaller boxes that can grab more resources when one (or a few) of the smaller box(es) need it. In other words 24 cores of 'elasticity'. At this you can just crank away at or near 100% utilization around the clock without the huge prices cloud providers charge for using this much CPU on their cloud these days. Considering the alternatives, an "entry level" 24-core $1,500 CPU as part of a whole system that costs a lot more is the 'cheap' option. The other ways you could be doing these tasks either with separate servers or in the cloud is far more expensive. And if you don't have this "workstation" capability, then you don't have the "sandbox" you need to figure out how to do this stuff. The idea being you figure it out on something like this and then you move on to the production system. If you have ever worked on developing for large scale corporate setups, you gotta have a "play pen" somewhere to figure out stuff. Granted often times "main stream" hardware is enough for the task you are trying to figure out, but not for everything, so it really hurts when you are missing this piece. It is a good way to end up with a production down scenario when your development area isn't good enough. And if your development area isn't reliable enough, often you are just straight copying from your initial development area all the way to massive production server systems, so any corruption that happens on non-ECC "main stream" hardware the bean counters decided was cheaper than real workstation grade hardware gets copied straight into production and then all hell breaks lose. I remember having a short conversation with Linus Torvalds about his non-ECC Threadripper system and he insisted it was fine for him to not have ECC memory of all people and I really seemed to have struck a nerve when I said he needed ECC RAM. Then I saw later that he was having corruption issues and basically had to admit that I was right. I suspect because things went so badly for Linus Torvalds of all people, this is why even the 'regular' Threadripper supports registered ECC DIMMs in this release.
I like when he says it suitable for a cheap workstation... I am VFX supervisor and generalist in big VFX studios and I wish I had that in my workstation...
the 7960x is more than double the price of the 7950x and it all depends on the complexity of the FX you are planning to render. I would say that for most case the 7960x is not worth it. The 7950x has a higher clock speed which mitigates the 8 missing cores a bit. BUT if you are planning to get more than 128 Gb of ram for large FX calculations then you have to go for the TR or the 14900K (up to 192Gb). In terms of efficiency and price I think the 7950x is the best. And you can keep your money to either compute super heavy thing on the cloud using AWS for example or upgrade in the future.
I'm here a day late, so I don't know if anyone has already mentioned this. Poisson is pronounced "pwa-soh" as is named for French mathemetician Siméon Denis Poisson.
I'd like someone to cover the fact AMD is absolutly gouging to price of these threadrippers in UK. The 7960x is £3000 in the UK (~$4000) and there are no sTR5 motherboards being sold anywhere.
AMD is not doing anything, your importers and retailers are gouging you... This happens every single time, with new releases everywhere... Someone should cover how people are still not used to wait some time for the market to settle...
Sometimes the hard call is clock speed vs core count, (the 7970X seems to have the same base clock, so that CPU is better, though $1000 more) The higher core count CPUs often have a lower clock, but then again, they do turbo around the same. The higher clock can be handy because multi-threaded applications often have a single thread operating as a traffic cop, and that traffic cop can be a bottleneck. If you are running 1 application/script at a time, and/or occasionally get pegged with a single threaded load (I had a compiler load like this before in the professional world), this can be especially a hard limit. The 7980X especially has a 2.5 GHz base clock, though the 64 cores would probably more than make up for that for many workstation loads. Of course, if you are running multiple applications at a time, the 7980X will probably have a larger advantage. The case where you are running 1 or a few applications at a time is where it can get tricky. Would the turbo make up for it for the lower base clock (it could actually if the other cores aren't active)? Do you want to rely on the turbo? Is it just better to buy a 2nd PC?
Chromium compile is good. One interesting bottleneck in compile is I/O. When someone said "SSD in RAID", I was like "naw, it won't make a difference", but it did shave off like 5-10 minutes at the time, which was pretty significant. This was before today's insane PCIe SSDs and was SATA. This might be a potential advantage of Threadripper. You can have a pair of PCIe SSDs RAID 0'd, but have a regular SSD for the OS to keep the OS from getting killed if you screw up the RAIDed SSDs. Remember: time can be money
If you guys added the phoronix test suite you'd become the gods of hardware testing (if you didn't feel that way already 😅) Maybe in collaboration with Level 1?
On a side note, the 7940x or the 7980xe is a better buy than the 7960x. The core clocks on the 7960x are not great. The 7940x and 7980xe will clock higher.
I'm wondering how any of The current generation xeon's match up against threadripper? I mean admittedly those are different class. Those are more likely to go up against Epyc which is a completely different price bracket. It would still be interesting to see.
Maybe not a popular question, but what cpu gpu combo would be the most efficient when it comes to power consumption while still having decent enough performance for most games out there these days?
The 24 core really wakes up when overclocked, sustaining 5.2GHz on all cores, all the time and it can do 128GB ram (or even more) at the full 6000MHz which normal ryzen cannot do. It's a friggen beast. I'd love one.
I know this is a completly seperate thing. I love youre information. I Iive in the UK and I have a 5700xt. I cant afford anymore. Love watching you. Id like to know,, i have an r5 3600 should I got to a 5800x3d or upgrade my 5700xt Whats best? Specigic question and I apologise for that.
i would go for the 5800x3d first, this will give you already a great fps boost and will be great for the coming years. And with that also upgrade to a fast set of ddr4 and sell your 3600 and old ram. Used ddr4 is pretty cheap these days.
I think the generic answer is that if you are maxing out your gpu then you should upgrade that first, and otherwise the cpu. GPU can uplift resolution, quality, and fps, while cpu only helps if it's holding back the GPU to a specific fps cap.
idk if done yet but would be cool if you guys could interview professionals who work with these software and give some insight and dumbing down for people like me who are intrigued. kind of a "meet the professional _______" series.
Did I miss some explanation in the video? Why is the font for the data labels in the charts different than normal? It's only the top result, too. If there's multiple, the rest are normal.
The tiniest of silver linings is that it's "only" $100USD more than the 3960X. I moved from the 3960X to a 7900X and will likely not move back in the foreseeable future. I got burned by the sTRx40 socket, and don't want to get caught like that again.
Read the article HERE! gamersnexus.net/cpus/amds-cheap-threadripper-hedt-cpu-7960x-24-core-cpu-review-benchmarks
Check out our 7995WX CPU overclocking livestream: ua-cam.com/video/vU179_czCnU/v-deo.html
Watch our 7980X & 7970X review: ua-cam.com/video/yDEUOoWTzGw/v-deo.html
I will ALWAYS remember the moment the CPU pulled 1400 watts. The panic! 😂 I loled hard. Thank you for this momerable moment❤
ua-cam.com/video/vU179_czCnU/v-deo.htmlh30m21s
Would love to see a FL Studio 21 benchmark :)
Anybody working in a Startup/Software/Developer environment and used to the older Intel XEON workstations know this is extremely FAST and Efficient performance with triple the I/O available. And its the lowest end version. For Software Startups or small SFX/Graphic Design/Web Design companies this is a huge upgrade in a workstation.
hi gamers nexus just one thing and maybe not important, Poisson distribution is pronounced pwasOn. otherwise cool video
Thanks for the great content as always!
When Pat Gelsinger mentioned that he now considers AMD to be in the rear-view mirror I think he confused the rear-view mirror with the front windshield.
😂
Lmao so true ;-)
Rofl. Someone must have poured a barrel of whiskey into his Pre-Christmas Eggnog.
No... rear view... he's getting LAPPED 😆
Intel copium
This is a niche use case but I'm so glad you guys did some benchmarking using various molecular dynamics engines. I am so desperate to find actual benchmarks for MD packages, this is an absolute godsend.
That's awesome to hear! Was that the lifesciences section? We'll keep adding those!
It's always nice to see these tests, but for an avg. home user that do spent some time doing some of these tasks, no way a Threadripper system is worth the cost. It has to be pretty serious work for you before Its worth spending 3-4x or whatever the total cost runs up at compared to the top Intel or AMD dsesktop systems.
@@mrdali67 it depends on how much value you place in your computer in general. For me I'd go for threadripper if the compromise in general gaming performance weren't so drastic, purely because running particle simulations in Houdini is so slow with my poor 12400F and there are still games like Cosmoteer that eat threads for breakfast.
@@mrdali67These definitely aren't benchmarks for a typical home desktop. They're useful for something like the analysis workstations in my lab, though.
Very interesting benchmark, hope you guys also add MD benchmarks for gpu testing
The Chromium compile data is crazy seeing how much improvement has been made since Ryzen hit te market is almost incomprehensible.
Ow. So it's not a apples to apples comparison 😮?
And then you can speed this upeven significantly more on these new Threadrippers by setting some NUMA parameters, c.f. Phoronix. It's crazy how powerful they are for compilation workloads
These chips seem like a crazy good option for CI servers that run compilation jobs all day
@@zachansen8293they specify the version they compile btw
It’s interesting to see over the past 7 years AMD go from non competitive to competitive to outright domination
d o m i n a t i o n
That's right and I'm all for AMD doing the same in the GPU scene. They're still in that competitive phase imo
@@piev4166 Personally I want to see Intel become competitive in the GPU market so both AMD and Nvidia actually produce budget cards.
@@piev4166 hold my dreams 😅
@@samfisherxboxog8925 Keep dreaming. As soon as Intel gets truly competitive, their prices will be the same as the other two.
7:08 It’s so cool to see something like this because usually you’d have to dig in some specialized forum to get some comparison data for different CPUs.
I think it needs a little contextualization tough
CFD software are limited by memory bandwidth, the more the better. So higher number of memory channels and faster ram are the most important factors.
That’s why there is a significant jump in performance between the 14900K and the Threadrippers in the OpenFoam benchmark. It’s also clear that the 7960x has enough cores to saturate the memory bandwidth, so higher number of cores provide a very little benefit.
I’m not familiar with how the Rodinia benchmark works but apparently it’s used to compare different type of hardware (including GPUs), so it’s not a surprise to se some mixed results
Phoronix does HPC benchmarks if you're having trouble finding them
The pronunciation of Poisson at 10:41 gave me a good chuckle. For future reference it's closer to pwa-son (with my highschool French as reference)
Like fish
Like croissant
13:01 impressive to see the 14900k so far ahead of the 13900k. And people say they're the same chip!
I predicted when 12th gen came out that it would look like a prototype compared to 14th gen.
Boy, was I wrong!
APO or whatever could be the cause, which would be a shame to see. but gimping old products has been proven to work in the past to sell more new products *cough gtx780 *cough
@@limemasonAt the time we all thought Meteor Lake would be a desktop architecture too and not just laptop. In hindsight Intel should've never released a 14th gen to desktop at all.
And yet Intel will say it's a new chip with new stuff that only the 14900k can do ....:)))
@michaelscarportliterally
Spec benchmarking was *extremely* helpful. Hope that info stays available going forward b/c it really helps in terms of deciding whether the expected performance increase is worth paying for.
I hope you'll consider doing it on lower end hardware. If you don't need the threadripper IO and the extra performance doesn't have enough ROI, it's good to have a comparison between the 7950x and the 14900k. (And ideally a part or two lower down so that we can get a good sense of where the ROI stops making sense.)
P.S. Happy to talk to you about the details of the benchmark stuff if you think it would help you improve your coverage. (At least the details within my wheelhouse.)
P.P.S. Look up the pronunciation for "Poisson distribution". It's French and pronounced like Pois-son.
As impressive as this is, I'm glad to see that it doesn't stomp my 3 year old TR 3970X *too* hard (unlike what the 4090 did to my 3090's in Redshift rendering). My usual upgrade cadence for work machines is 5 years/4x performance boost (whichever comes first) at comparable budget levels, so I'm still good for a while longer.
@@zachansen8293lmao
Your 3970X is an old dinosaur compared to anything in the Threadripper 7000 lineup.
@jimmay is it true that a single 4090 will path trace render twice as fast as a single 3090?
@@chiefjudge8456 What are you talking about?? 3970X is the 6th fastest CPU in the Blender rendering test. And still faster than 7950X. Still relevant today I would say.
@@richardconway6425 In Redshift, it is nearly 2x the 3090. Can't speak to other rendering engines.
OMG Finally some Threadripper coverage... please keep them in with all other CPU reviews in the future.... Great vid
Steve, I really do appreciate you putting in a few game benchmarks. My system is about 70% video & audio editing / encoding / bluray authoring (Dolby Media Encoder, MainConcept's Encoders TopazLabs Video AI), & C# development (VS 2022), about 30% some gaming nothing amazing (BG3, WoW, etc.). So seeing it's not going to just wreck my experience is nice. I have been really looking at either the 24 or 32 core
Really happy to see the "Spec" Workstation Benchmarks beeing considered !
Thanks =)
I run a 3960X workstation so this is very helpful. BTW Poisson was a French mathematician - hearing it pronounced poison (as in toxic) while eating breakfast really perked my ears up!
To follow up on the how to pronounce part, Poisson, is not poison. Its stohastics metric, named after French mathematician Siméon Denis Poisson (/ˈpwɑːsɒn/; French pronunciation: [pwasɔ̃]).
I belive it was in stochastics class in 3rd semeser CS university.
Anyways.
P as in .. poison
O as in order.
I as in ardent ( damn french)
SS as in snake thats very hissy
ON as ON something.
Fish is a funny last name.
Poisson in French also means fish.
Makes sense that FFT would scale very well with core count, since it's a matrix factorisation and multiplication problem at its core, and that scales almost infinitely with core count. In fact, it's almost always worthwhile to do the computation on the GPU, especially with larger data sets.
When doing FFT, it may help to turn off extra hardware execution contexts if doing it on a CPU. FFT tends to fill up as wide as you can go, which is why it is so much better on a GPU. Multiple hardware contexts on the same CPU core is extra overhead and competition for resources. At one of my previous jobs where we were doing FFT heavy workloads on Xeon servers, the admins had Hyperthreading turned off specifically because the FFT heavy workloads ran faster this way.
I think that a side-by-side comparison with a low end Xeon is more appropriate than a comparison with an intel 14900K.
They'd probably need Wendell's help to find a Sapphire Rapids system to compare to, and help them understand how to test it.
It would be interesting but we all know Intel has no response there.
What they can get away with on consumer oriented desktops won't work at this scale
Intel sells the Xeon W line as a counterpart to AMD's Threadripper. But I think AMD is ahead of those in price/performance and allout performance thanks to the crazy 96 core.
@@titaniummechanism3214W series ismt HEDT though. Its like comparing cats to dogs.
@@randomguydoes2901 That's the question, though, isn't it? Intel cannot compete with the 64 core and 96 core parts, but the 24 core part may be different. But because GN did not test it we still don't know.
I know you guys don't care for the 7950X3D as it gets left out of a lot of comparison charts, even if it was used in certain charts in a video it will get left off of others in the same video. I would love to start seeing it included in more testing especially in all of these novel tests.
I totally agree! It's annoying especially when they keep including the 7800X3D when no other X3D parts are shown. If they show the 7800X3D, they should show the 7950X3D as well.
Thanks for the Monte-Carlo tests! Would also be really nice to see temperature graphs. We got 3960x and 280W is no joke to cool, it also puts a lot of stress onto VRM. Wonder how 350W will do in that department.
280W is what a stock 14900K runs at. I'm sure board partners can make it work on AMD's side as well
@@kekoraaaa 99% of use cases for 14900k is gaming at 2-4 cores full load. 7960x will be 90%+ load for 24/7, probably in hot case with several GPU's, so VRM heat might be an actual issue. I wouldn't compare these two.
man, your mane has reached 'glorious' levels. lookin' sharp !
I am, at some stage, expecting a shampoo endorsement. At the very least, conditioner.
Thanks for the code compile test. I'm a software engineer and I'm glad to see a linear trend when going up in tr products, you will get better performance.
I swear I totally didn't fast forward to see only the gaming benchmarks...
Always enjoy the informative content, and at last have subscribed to your channel. As an avid gamer, and engineer I've been looking to upgrade my home hardware and a 7960x might be just what is needed.
Also, FYI @10:42 "Poisson" is the last name of a famous French Physicist & Mathematician (known for Poisson Equation and other achievements) not to be confused with toxic poison, or poison ivy etc.
10:40 "The *Poison* result-"
Poisson is pronounced like 'Pwah-sohn' lol
No cpu was harmed 😅
as usual pure professional... thanks for another high quality review
Steve and Gamers Nexus team, your videos are a joy to watch.
Thank you!
I want a gaming PC and all roads led me here, so I got a good idea on what to get. Awesome channel. Very valuable.
16:30 I'm getting this cpu for my main pc for cs2. Planning on running a server, gaming, and compiling and running code on the third monitor
Watched beginning to end. I don't play non-Nintendo games so I usually skip to the conclusions of gaming benchmarks but this stuff I really like.
So you only play kids games?
@@thelunchbox420xoh yes you’re clearly not the childish one in this scenario 🙄
Incredible, incredible work 👍
Even if other people are doing these tests, you guys are making it easy to find and using your huge platform to share this info. THANK YOU! ❤
Threadripper is amazing.
Remember Intel charged 2000 US dollar for a HEDT desktop CPU in gen 9. In gen 10 they charged 1000 for the equivalent CPU. And how many cores was that I do not remember but it was less than 10. And now we get 24 cores for 1500 us dollar. I would say that is a win for the consumer. And that Intel back in the days Price gouged the HEDT market. Just because AMD could no deliver anything to compete
10 years ago you were paying this much for a 6c12t lmao it’s crazy
Sintel was price gouging the *18C/36T i9 9980XE* when they *dropped the price 50%* for the i9-10980XE !
Less than 10 cores on the $2000 9th gen HEDT???? How ignorant can you be? Lmfao. The 7980XE, 9980XE and 10980XE were all 18c/36t. You should probably use google to double check your information the next time you can’t remember.
@@phiphedude7684no you weren’t. The 3960x and 4960x Intel CPUs were $990 USD at launch……
Edit: And those are the top end skus, the 3930k and 4930k were 6 core CPUs of the respective generations and could be had for under $600 USD.
Yes its an exaggeration obviously but still pointing out how 6c12t used to be insanely expensive and now its standard
Thanks for doing this. I'd still really love to see some tests involving Davinci Resolve since many editors, even youtube editors are switching.
Love you guys, hope you keep doing the CFD tests too, as I know its a bit niche, but will be useful knowledge for me. Would be super interesting to see deltas between TR and Epycs!
I believe the name of the "Poisson" test and the Poisson equation comes from the French mathematician, so it's pronounced in a French way, kinda like "pwah-son", rhyming with "gone", and the ss is not a z-sound
My god, its so good to see such a great review.
Super onboard with the CFD testing!
Would love to see these high core count CPUs run some encryption benchmarks go see how many hashes per second they can handle
I would have loved to see you guys test your hypothesis for gaming + rendering! Please add that as a one-off test?
I’m looking to get rid of my TR 3970X for a 7960X. I do reverse engineering, and CAD work but the programs don’t seem to take advantage of 32 cores. It seems like a new generation w/ 24 cores will suit me much better. I game a lot as well so it’s nice to see that it seemed to be an improvement there as well. Thanks for this!!!
As an owner of a 3960X that makes a bit of CFD stuff, it was a nice video to watch and learn from it 😊
@GamersNexus, the DXO photo editing software can use all cores for photo AI correction. Example my father is using 120 photos to correct, and we calculate in minutes and seconds how fast it terminates the correction process.
Memory bandwidth tests would be nice.
This is the Threadripper part I've been interested in due to the "entry" level price of the processor and the base platform. Still expensive but there is already an upgrade path to the 96 core part on a motherboard I'd get even if AMD were once again to forget about workstation. Hopefully they stick with this socket for Zen 5 and Zen 6 parts.
As for the GN video editing workstation build, perhaps pair with both a fast nVidia card for CUDA based applications and an Intel ARC card for QuickSync? Best of both worlds? The unholy trifecta of red, green and blue teams?
adobe premiere cannot handle multiple gpus. it will somewhat randomly just use one gpu. it may be different for davinci resolve.
AMD has only released new CPUs on the first threadripper socket 😢
They will abandon this platform in a year.
This will be a fun revisit in 2033 to see how these wild workstation parts compare to normal desktop parts of that future date and whether games and other "normal" software will have become sufficiently multithreaded to utilize these currently-ludicrous core counts.
Except on @RandomGaminginHD
CFD is heavily limited by memory bandwidth. Will be interesting to see your Pro chips on the TRX90 platform for its 8 RAM channel glory!
Ah yes, we've come full circle.
You used to just get tons of IO on computers and it was just how it worked. Expansion cards for everything. You needed it for audio, networking, industrial equipment interfaces.. Etc.
Then it all got built in and locked down.
Now, hey, the professional market that needs that IO has the cash.
It should be said those that need it are in a niche. Almost everyone here isn't in that category.
Not really. Adding in a second netwrok connection, be it wired or wireless, is not too niche, neither is adding 1 or 2 NVMe adapters, that is 2-3 slots, with a gpu taking at least 2. Most mATX boards already peace out, forget mini-itx, and even most ATX are done deal if you want rear channels like me.
Then I just add an ascilloscope, which is too tall an order for most motherboards.
I just want a barebones board with just 3-4 NVMe slots, just surround sound out E-key M.2 slot. Then I am golden. My choices instead boil down to massive highend "GAMUHRR" boards or just finding workarounds and adapters and whatever BS.
Well HEDT seems to fit, sometimes. Just a 12 core is enough, I get bonus memory channels, and all the lanes I want. Sometimes there is even a mATX version of such boards, which allows me to keep my custom 2U case.
Funny my build has NVME, SATA, USB, NICs and PCIE slots; loads MORE i/o than PATA, ISA & serial
I seriously believe some of you underestimate how many machines it actually takes to "run all them simulations, and make the various media yall watch and such."
Every thing outside or ones usecase or what they "see" everyday is a "niche" now 🙄
You don't bend over backwards for a "niche." Server/ workstation/ HPC has always been here.
@michaelscarportunless them chips have way more lanes than it's advertised to, I guarantee just "aren't noticing" the bottlenecks... or not necessarily using everything simultaneously for it to scream at you.
If it was as simple "research your mobi better" we wouldn't be here today with non-pro Threadrippers again because that crowd (we 🤔) would have definitely sang at the top of the hills if we could have really made by with any odd consumer level platforms.
@@jamegumb7298 you have to admit that is niche though. Additional networking? Completely unessesary. Multiple NVMe? Professionals have network storage for large amounts. Storage servers aren't running thread ripper.
"Ascilloscope" (oscilloscope?) hooked up to a computer is VERY niche. Surround sound out? It's emulated now over USB headphones or use an external DAC.
But I used to have cards for all those things you mention. I get it. It's just not necessary for 99.95% of users. That's a niche. And niche markets are always expensive.
I appreciate that you test across a wide range of CPUs, but please put some visual highlight on the HEDT parts so they're easier to pick out. Most of us in the market for one of these cannot even consider the consumer parts regardless of how they rank in these benchmarks.
could you look into a re-review of the 7500f? it's been popular with budget builds in the eu as its an incredibly cheap entryway into am5, maybe having it into the charts alongside others would be useful
Great review, as this is the part I can most afford. I'm a 3D artist who will use the PCI-E lanes for GPU cards and GPU rendering. Would love to see you review the Gigabyte board to see if that is a good fit.
The thing that would really interest me is how much would running a VM with a given corecount of lets say 4 and 16 GB Ram affect the rest of the system for performance intensive activities like gaming that per se do not use or need these resources. These additional cores could be used to make a second PC that is being used as a server or something unnecessary.
With VMs, they don't actually lock the cores for their own use. If the VM is sitting mainly idle, it has barely any impact on the host machine's CPU usage. As a result, for a general purpose VM Host machine, you can easily get away with allocating to VMs 8 vCPUs per 1 physical core.
@@moist_ointment Sure, but my question is not even if the VM uses the cores and ram since I have more than enough left (8 cores and 48GB RAM. The question is if there are other "resources" that the VM uses/can use that will effect the performance of the system.
@jackchid6040 I would like to see VM benchmarks too, considering my job currently is to set up and maintain VM clusters, and we just best-guess the CPU based on assumptions.
But the workload of a VM on CPU, RAM, and Disk IO (disk IO can be a major factor when running a lot of VMs) all depends on the current demands of the VM at that time.
We comfortably run 40+ VMs on 16 core Xeons and are bottlenecked by only having 384GB of RAM of Disk IO long before CPU becomes an issue.
I prepared a large tank of copium prior to watching this video, as I already have a 7960X (but no motherboard as yet). Luckily, the copium went unused: my own guesses about the most appropriate Threadripper part for my workflow turned out to agree with GN's data.
I think it's the one actually that makes the most sense. It's got a few more cores, plus a lot more memory and expansion bandwidth, for a price that is attainable to those looking for a machine capable of that stuff.
Ultimately I think the high core count ones won't really make much sense unless you go for pro with its even higher PCIe and memory connectivity.
How do you explain the Blender chart at the lower end? I think there's a mistake there, how would an i5 possible do better at Blender than an i9?
RE: Squarespace ad...
I'm old enough to remember when FrontPage, Dreamweaver, etc were new on the block and considered "cheating" when building a website. (I understood the FrontPage hate at the time with its massive markup issues behind the scenes bloating pages in an era of dial up, but tools like Dreamweaver comparatively weren't nearly as bad IMHO.) Now, this is just what we do. SMH
If only I had known this was truly around the corner, I may have waited. Instead I have been trying to create a home lab with VMs using XCP-NG, on a AMD 7950X, with 128 GB RAM. Well we will see...lol might have to trade up.
I'll take your "old hardware" 😂
What are you doing anyways that you need that beefy hardware, for a Homeserver?
CoCalc? LOL
@@alexdomio4982 Naahhh, I'll get rid of my Phenom BE from 2010 before I'll get rid of my "old equipment".
Doing information security research and relearning the basics of networking, filling in the gaps in my knowledge base from incorrectly taught garbage I didn't learn over the years. Basically, working on a Masters>>PhD degree in my spare time so that one day I can truly quit my "Day job" and enjoy living. 😉
Did you trade up? I kinda in the same boat. I'm waiting for a faster cpu to slot into the socket, not that the 7950x is slow but damn the 7960x can do ddr5 6000 on 48GB dimms. 4 x 48 and cpu overclocks to 5.2GHz all core rock solid. No harm to review but it dosnt show its full potential but I, just built the pc 4 months ago so cant really justify changing.
@@jondonnelly3 similar situation for me. The juice isn't worth the squeeze for changing.
Yeah I hope AMD new 8950x or whatever they will call it that I can just drop in and ebay the old cpu on. Maybe even X3D on all cores with 6GHz Probably too much to ask in one gen. @@renzokuken2g
Great vid as always Steve et all. I would love to know more about the Blender benchmarks. This is EXACTLY what I'm interested in, and I appreciate that you're looking at it and not just games etc.
"Even for those using EEVEE with CUDA, it's still impressive to see how much can be done with CPU rendering"
OK but why? Why focus on rendering for CPU benchmarks? Is the CPU really going to matter when you have a 4090 in there and you're rendering on that with Optix?
1. if we're using GPU rendering, I'd be much more interested in knowing how fast the CPU did a liquid simulation (which I think is CPU bound?) Why are the CPU benchmarks focused on The Thing That Will Be Running On The GPU?
2. I would expect most anyone with a recent NVidia GPU would be using OptiX not CUDA, but same applies
Thanks for the vid! and ps it's Poisson pwa-sson
This does not relate to this video In particular, but overall.
You are a comfy guy with your dudes.
Keep up the good work.
i don't know why i'm watching this. it 10pm over here and yet this is the thing i wanted to click most on. thanks Steve. back to you Steve
4am here lol
Solid video
For local code compilation for my development work, I wish I had a Threadripper system! Okay, I also wish the build server were an appropriate Epyc but alas... The speed is impressive - even more so on Linux with SNC/NPS tuning as per Phoronix.
I think the 32-core boy is more justifiable.
Well in the video it is shown that the 24 core performs close to the 32 core, but for a significant price decrease. I think that justifies this product a lot
@@Shanboss277problem is the total platform cost honestly.
Rdimm on non pro was a mistake, ecc udimm is good enough @@AjrAlves
Pro could be rdimm and that'd be fitting
I would really appreciate if alternating row coloration could be used in the benchmarking charts, on more than one occasion watching this video I had to pause and visualize which row reference was which chart bars
Need to remember that these target businesses where they may purchase a few hundred or thousand units. If there’s no benefit to CPU tasks at higher core counts (or diminishing returns)-particularly where GPUs take the majority of the load)-then saving a few grand times 500 units results in a large savings.
Would it be possible for you to include idle power going forward? The added IO makes this really attractive for homelab, but not if the platform as a whole pulls 150W sitting idle.
This!
Just a heads up, whenever people say things like "convolution" and "filters", they are nothing but matrix multiplications!
Convolution and filter both are types of getting smooth data from a relatively grainy set.
Five years back I bought a 2965X 16-core gen 2 threadripper. At first I was very pleased with it. Then the Enermax 360 gunk problems occurred. Ugh. After the E(holes) claimed they fixed it I got another one. Fixed? Not so much. My choices became going full custom loop (not interested in the maintenance) or going with an air cooler (but it was build in a PC-011 dynamic case and hardcore air coolers didn't fit).
Or use another 360 aio but the cold plates didn't give full heat shield coverage. I ended up with the Dark Rock cooler with a full coverage cold plate. While it does a fairly decent job, the 2965X at full boogie overwhelmed it. Not as good as the E(hole) 360 cooler (when it was working). It also meant I had to move my builtiful custom cabled installation to another case (Fractal R6). Bummer dude.
So! I noticed you didn't seem to be using a cooler with a full coverage cold plate. Are there any AIOs with such a cold plate? Can an air cooler even think about 300W+ cpu? How much performance was being left on the table with such a cooler?
Does one need a case that allows dual 420 radiators for the cpu?
I recently upgraded to a 7950x which comes close to doubling the 2965X scores in Cinebench and appears to render DaVinci twice as fast. Plus it has a more standard heat spreader so I could stick on a lian li galahad II trinity which allowed me to built in my pc o11 dynamic again.
I would *really* need a hard core video editor rig again before I'd pay the extra $$$$ for a thread ripper again.
Honestly the only maintenance I've had to do on the open loop keeping my 3960x cool, was a top up after swapping out an air cooled GPU for a liquid cooled one and stuffing a secondary pump in for a bit of redundancy because I could. Most maintenance problems come from people putting stupid coolants in with dyes that gunk up the works, a lack of biocide that gets algae growing in there, and mixed metals that just corrode everything.
Stick with copper or nickel-plated copper, use clear premix with a biocide (and maybe a silver kill-coil if you're feeling fancy), and there's absolutely no reason your open loop shouldn't last as long or longer than any AIO, with the advantage that you can just put more water in rather than having to replace everything. Only disadvantage is, a good loop does not come cheap. It is a good investment though; one of those things that's truly "future proof", unlike people overspending on parts they're going to be replacing in a year just to get moar frames per second.
If and when I get a new motherboard, RAM and CPU for a 7960x, the case, PSU and cooling loop is going to be a direct transplant. From everything I've seen as well, the TRX40 full-size coldplate will fit TRX50 as well, so I won't even need to swap that out. I also have dual 60mm-thick 560s for the radiators, and they're honestly excessive, so the $$$$ initial investment that's already paid for itself in terms of money earned using it, will be a "free" added bonus this time around.
Im still here i have hedt 3930k and it was about 500$ :D how times have changed.... Btw im still on that CPU.. its not the best but it still serves its tasks
4930K user here.
They are solid chips indeed.
Being someone who is doing some heavy rendering (ok, all the heavy rendering) ol' buddy is getting "demoted" to primarily gaming/ semi-server / whatever other mad scientist stuff I can do with it lol.
(If I don't sell it. Have you seen what a working X79 parts go for now? "Affordable" HEDT is definitely missing and in demand.)
well hat of to you sir :D they were calling me a madman in 2012 you don need them 6 cores :D but i knew that intel would stall the development at that time... :D hahahab
ive also changed 3 gpus on this beast and still goes strong on 1080p@@MaddJakd
@mit4c They loved calling us mad back then.
I wanted a "best of both worlds" machine (plus film school lol)
"It wasn't possible" they said. They still try to say that too, but I don't remember agonizing through 10 fps when I did get to finally play lol.
The move now is moresoe because the footage is getting into that territory of "More core were required as is" and its finally fighting me for being stubborn.
A new GPU or 2 and even only the most sensitive of frame hores will notice honestly, even back then.
When this mobo used is going for the same price as it was fresh off the shelf back then, prettt sure more than just us crazies are seeing the utility in these "old" HEDT platforms.
Question:
When LN testing, would it not provide better thermals to have a taller tube of LN over the CPU? Ideally you want the boiling layer over the CPU to be as thin as possible and since liquid pressure is directly a result of depth and not strictly volume of a container, would a taller cone not create more pressure and limit how large the boiling layer becomes under load? Of course with this it may be beneficial to create a cage to limit how much LN is thrown by rapidly boiling LN.
OR, is the compression negligible and doesn’t change the gas layer much? Strictly asking about pressure via gravity and not pumps.
Thanks Steve!
Thank you for reviewing these. I am *much* more interested in the 16-32 core processors (ideally the pro line with the more memory bandwidth) than anything greater than 32 cores (for that I would go with server products thanks to the 12 channel memory). That being said AMD *really* needs to work on pricing here. the 4 channel parts (non pro) I don't really see much of a use for, but even then for either the performance/cost metric for these are completely off. There's no way to justify a 300-1000% increase in cost for cpu alone and have performance be lagging by this much. realistically the prices should be halved if not more if considering the pro versions) i.e. I can see 100-150% increase over the same core count (16) 7950X desktop part and then scaling price w/ performance above that, but *NOT* what we have here.
Yeah it seems really dissapointing. You can easily build 3 7950X (48 cores) systems and configure them as a compute server for less money than it would cost to have an equivalent 48 cores build on the threadripper platform
Thanks Sreve!
can you include davinci resolve for video editing? premiere is such a mess and davinci resolve ususally can use alot more hardware more efficient.
I have to wonder how much of a difference there is between Windows and Linux using these CPU parts? I thought at least at one point Linux was doing a lot better because it got support to manage the multiple CPU dies effectively first.
I suppose this brings up another point, at least historically, it has been considered folly to load up Windows with too many different things. I remember trying to just get my software development stack all on one Windows computer and at first it was taking several minutes to boot long before getting the full stack on as each new piece needed a reboot, and then Windows would split apart at the seems and BSOD on boot before I could finish installing my development stack. I had to spread across more than one computer just to develop software applications and my software development stack was fairly rudimentary, at least compared to what I was doing under Linux at the same time without any real issues getting it all going, plus with Linux I got it all setup with just a single reboot to switch over to the kernel on the hard drive. Especially with all of the crashes and reboots, just a fairly basic software development setup under Windows reminded me of an old IBM commercial where a guy trips and drops his deck of punch cards and then the next scene he is getting ready to hang himself. I would think this also limits down some the use cases for these CPUs for Windows users. In the Linux world, it is more common to load up a system with lots of different tasks and even lots of VMs and other containers such as Docker containers. Especially for home lab stuff, I could see getting a 24-core CPU like this because you need something reliable and with the I/O capacity to have some expansion cards besides a GPU and this is at an "entry level" price for such a setup. In other words the normal use case is you have a bunch of different things happening on the system under Linux and maybe some Windows VMs going (Windows in a bottle where it belongs), but then as a home system, you sometimes also game on it, especially as plenty of video games run under Linux natively or maybe you give a Windows VM direct access to a GPU. Trying to spread across a bunch of systems for a home lab setup can get really expensive, you need to have a 19" rack for it all, and then the power bill is insane as well as having to pay an electrician to wire up the power to deliver to all of the hardware in a 19" rack. You instead consolidate all that you can to one physical box and while it might end up being a beefy box, and it is unfortunate that Caselabs is not currently around to make a good case to stuff this all in, you can run it all off of the 115V outlet that was already there in your home and call it a day. You have to consider that a commercial grade 115V, 15A UPS maxes out at around 1,200W because it needs 300W for charging the batteries, you don't want to push more than 1,500W continuous through a 115V outlet, and for a home lab setup you need a commercial grade UPS because consumer grade just doesn't cut it. So with just the CPU taking up to 350W before conversion losses and the GPU taking up lots of power as well, maybe even more, that is over half your power budget powering just two chips in one computer.
While some have pushed more into the cloud instead of having a home lab setup, on the flip side cloud services have been getting rather expensive, so then this pushes things back to running on a home lab setup, especially if you can get a CPU like this and have the flexibility to have 'elasticity' with 24 cores right there to either run broad and fast for parallel tasks or divvy up when you effectively need a bunch of different things happening at the same time. Going for a bunch of smaller systems has the problem that it is harder to go parallel effectively when you need to for a single task. I mean SMT (symmetric multi-threading) is the "quick and easy" way to go parallel and make quick work of that big task. Trying to get a bunch of small boxes to divvy up a big task and get it done quickly is a lot harder. So often with the many small boxes approach you end up with some things taking forever to do or maybe you can't get the resources together in place to get it done while the one big box approach it is easy peasy and you still have a bunch of smaller things going effectively using the one big box as a bunch of smaller boxes that can grab more resources when one (or a few) of the smaller box(es) need it. In other words 24 cores of 'elasticity'. At this you can just crank away at or near 100% utilization around the clock without the huge prices cloud providers charge for using this much CPU on their cloud these days.
Considering the alternatives, an "entry level" 24-core $1,500 CPU as part of a whole system that costs a lot more is the 'cheap' option. The other ways you could be doing these tasks either with separate servers or in the cloud is far more expensive. And if you don't have this "workstation" capability, then you don't have the "sandbox" you need to figure out how to do this stuff. The idea being you figure it out on something like this and then you move on to the production system. If you have ever worked on developing for large scale corporate setups, you gotta have a "play pen" somewhere to figure out stuff. Granted often times "main stream" hardware is enough for the task you are trying to figure out, but not for everything, so it really hurts when you are missing this piece. It is a good way to end up with a production down scenario when your development area isn't good enough. And if your development area isn't reliable enough, often you are just straight copying from your initial development area all the way to massive production server systems, so any corruption that happens on non-ECC "main stream" hardware the bean counters decided was cheaper than real workstation grade hardware gets copied straight into production and then all hell breaks lose. I remember having a short conversation with Linus Torvalds about his non-ECC Threadripper system and he insisted it was fine for him to not have ECC memory of all people and I really seemed to have struck a nerve when I said he needed ECC RAM. Then I saw later that he was having corruption issues and basically had to admit that I was right. I suspect because things went so badly for Linus Torvalds of all people, this is why even the 'regular' Threadripper supports registered ECC DIMMs in this release.
I like when he says it suitable for a cheap workstation...
I am VFX supervisor and generalist in big VFX studios and I wish I had that in my workstation...
I am looking at a new Houdini build, do you think the TR 7960x would be worth the extra price vs 7950x or 14900?
the 7960x is more than double the price of the 7950x and it all depends on the complexity of the FX you are planning to render.
I would say that for most case the 7960x is not worth it. The 7950x has a higher clock speed which mitigates the 8 missing cores a bit. BUT if you are planning to get more than 128 Gb of ram for large FX calculations then you have to go for the TR or the 14900K (up to 192Gb).
In terms of efficiency and price I think the 7950x is the best.
And you can keep your money to either compute super heavy thing on the cloud using AWS for example or upgrade in the future.
I'm here a day late, so I don't know if anyone has already mentioned this. Poisson is pronounced "pwa-soh" as is named for French mathemetician Siméon Denis Poisson.
I almost commented one the movie commercials I saw on this video.
I’ve wanted a Threadripper, since they first launched. I just don’t really have the need nor can I justify the cost.
I'd like someone to cover the fact AMD is absolutly gouging to price of these threadrippers in UK.
The 7960x is £3000 in the UK (~$4000) and there are no sTR5 motherboards being sold anywhere.
It's $2580 AUD here in Australia, even after GST (that's our VAT). That's a WILD price difference.
AMD is not doing anything, your importers and retailers are gouging you... This happens every single time, with new releases everywhere... Someone should cover how people are still not used to wait some time for the market to settle...
For the last effing time. AMD cannot control retail pricing. There are laws specifically prohibiting it, for better and for worse.
No it isn't. I got a 7960x for £1460 and a motherboard for £699.
Sometimes the hard call is clock speed vs core count, (the 7970X seems to have the same base clock, so that CPU is better, though $1000 more)
The higher core count CPUs often have a lower clock, but then again, they do turbo around the same. The higher clock can be handy because multi-threaded applications often have a single thread operating as a traffic cop, and that traffic cop can be a bottleneck. If you are running 1 application/script at a time, and/or occasionally get pegged with a single threaded load (I had a compiler load like this before in the professional world), this can be especially a hard limit.
The 7980X especially has a 2.5 GHz base clock, though the 64 cores would probably more than make up for that for many workstation loads. Of course, if you are running multiple applications at a time, the 7980X will probably have a larger advantage. The case where you are running 1 or a few applications at a time is where it can get tricky. Would the turbo make up for it for the lower base clock (it could actually if the other cores aren't active)? Do you want to rely on the turbo? Is it just better to buy a 2nd PC?
Chromium compile is good. One interesting bottleneck in compile is I/O. When someone said "SSD in RAID", I was like "naw, it won't make a difference", but it did shave off like 5-10 minutes at the time, which was pretty significant. This was before today's insane PCIe SSDs and was SATA. This might be a potential advantage of Threadripper. You can have a pair of PCIe SSDs RAID 0'd, but have a regular SSD for the OS to keep the OS from getting killed if you screw up the RAIDed SSDs. Remember: time can be money
Nice some else also uses the O11 XL for productive computing
If you guys added the phoronix test suite you'd become the gods of hardware testing (if you didn't feel that way already 😅) Maybe in collaboration with Level 1?
I could really see for a lot of people, the split could end up being laptop and then TRX5 parts for desktop.
The "7960x" is going around 250 USD on ebay.
RIP old 7960X
On a side note, the 7940x or the 7980xe is a better buy than the 7960x. The core clocks on the 7960x are not great. The 7940x and 7980xe will clock higher.
I'm wondering how any of The current generation xeon's match up against threadripper? I mean admittedly those are different class. Those are more likely to go up against Epyc which is a completely different price bracket. It would still be interesting to see.
id Tech 7 scales damn well. Would be great if you guys could test those Threadrippers in DOOM eternal 🤓 Thanks!
Ah yes, let me see if i should get one in 2027
It's getting close!
Are you guys planning to do a comparison with the Xeon W7-2495X or possibly the W7-2595X when it releases in a month or two?
What's the performance numbers on those gray streaks?
Maybe not a popular question, but what cpu gpu combo would be the most efficient when it comes to power consumption while still having decent enough performance for most games out there these days?
I am going to buy one to play 1080p and a pair it with a 980ti. Thanks, Steve.
Hi @GamersNexus! ETA on 2023 case roundup video?
the lack of AMD R9 3900x in the reviews bothers me. yet you have multiple low end chips for comparison.
Im so glad you finally found a modern CPU that uses more power than Intel's 13900k and 14900k! AMD finally did it, they topped ALL the charts.
The 24 core really wakes up when overclocked, sustaining 5.2GHz on all cores, all the time and it can do 128GB ram (or even more) at the full 6000MHz which normal ryzen cannot do. It's a friggen beast. I'd love one.
You coukd do with adding a machine learning benchmark to see how fast CPUs are at building models using large data sets.
I know this is a completly seperate thing. I love youre information. I Iive in the UK and I have a 5700xt. I cant afford anymore. Love watching you. Id like to know,, i have an r5 3600 should I got to a 5800x3d or upgrade my 5700xt Whats best? Specigic question and I apologise for that.
i would go for the 5800x3d first, this will give you already a great fps boost and will be great for the coming years. And with that also upgrade to a fast set of ddr4 and sell your 3600 and old ram. Used ddr4 is pretty cheap these days.
I think the generic answer is that if you are maxing out your gpu then you should upgrade that first, and otherwise the cpu. GPU can uplift resolution, quality, and fps, while cpu only helps if it's holding back the GPU to a specific fps cap.
@@jjbankert Im playing at 1440p atm and all is fine. i am owrried about dx12 and mesh shader/ I kno my 5700xt doesnt support mesh and other stuff.
idk if done yet but would be cool if you guys could interview professionals who work with these software and give some insight and dumbing down for people like me who are intrigued. kind of a "meet the professional _______" series.
Did I miss some explanation in the video? Why is the font for the data labels in the charts different than normal? It's only the top result, too. If there's multiple, the rest are normal.
8:09 where is the 7970x CalculiX score?
The tiniest of silver linings is that it's "only" $100USD more than the 3960X. I moved from the 3960X to a 7900X and will likely not move back in the foreseeable future. I got burned by the sTRx40 socket, and don't want to get caught like that again.
Damint Steve! I released 6 hours ago :) You could not let me have 1 day! just one day. Nice review as always