AMD Threadripper 3990X CPU Review | Code Compile, Render, & Compression

by birtanpublished on October 3, 2020

When I received the thread Ripper 3990 X on loan from a yet unnamed youtuber we quickly threw together an extreme overclocking stream and got the CPU to about five gigahertz with bearded hardware and on really zero notice since then we've been working on adding some new testing like a revisit of code compiled benchmarks and we've put together a full review of the 3990 x the AMD 64 core cpu is tough to benchmark at four thousand dollars it's also an uncommon purchase for most of our audience but it's still interested in from a perspective of seeing how the ultra high-end components scale and

Software the 3990 X isn't an easy benchmark since it begins nudging towards that server workloads which we don't test but we still have a full production suite of benchmarks for the 3990 X to see how it does in other applications before that this video is brought to you by Thermal Grizzlies conduct a not liquid metal conductor nada is what we've used in all of our liquid metal and delayed thermal tests capable of dropping CV thermals significantly and replacing the stock thermal interface over CPU thermals don't just allow better overclocks but

Also lower noise levels because the transfer efficiency is increased the mix of gallium and indium makes for a thermal conductivity of 73 watts per meter Kelvin outclassing traditional pastes significantly learn more at the link in the description below so first off we're just throwing this together before going to Taiwan it's a pretty complete review by our normal standards and that it's got all the tests we would do for a thirty-nine 70 X but the 39 90 X does sort of ask for a couple of new tests so we've added the code compile test for this one we're not really ready

Yet to talk about the full new methodology overhaul we do this every year around this time where we convert everything into new testing methodology new software new benchmarks all of that stuff new games but we're not ready to talk about it just yet except we pulled the code compiled test which is chromium it's done on Windows at least for now and we're using that as our new tool for the 3990 x and it was actually a very interesting one to test with a 64 core CPU other than that we have the standard set of things that we test for

Everything which would include Adobe Premiere Photoshop which everyone I think those will not really particularly scale well with a threader for CPU in general but we have premier on there which has scaled reasonable so far and then you've also got tests like compression and decompression which are interesting ones that show scaling in some places and not in others and a couple of other ones to look at blender for example for a tile based rendering application so that's what we've got to start with there's more that we'd like to do with this CPU as well like another

Overclocking stream an X OC stream and then maybe a chilled water stream something like that would be a lot of fun to do with it and then we've got some other production workloads that we'd like to run specifically on the 3990 x but all that's gonna have to wait till we get back from taiwan because we're trying to cram everything together now before we leave another quick note here so we won't really have much of a gaming focus here that should be predictable but the one reason we do still include some gaming benchmarks for stuff like this and the 3970 X the 3125

X is because specifically what you're testing for is not how well does it play games but does it play games well enough so when we have gaming benchmarks for these types of CPUs it's not to see is it at the top of the charts because it's not going to be a 9900 K will be in every instance currently so that's not why the game benchmarks there is really just to see if unlike a 2990 WX you might be able to run the games without severe issues like stuttering or just black screens or it doesn't even work and the older thread refer chips had a lot of problems with that going to try

And work with 3990 X more after our trip but for now what you need to know it's a $4,000 CPU the third 970 X is about 2000 or so it's supposed to be 1700 to 2000 the 3960 X is a fair bit cheaper the 1090 DX e cheaper still at about $1,000 in theory anyway and then the 3950 X which is a very good CP if you don't need the PCIe lanes and you don't need quite as many cores is in the 750 s so we're gonna be looking out a $4,000 part today almost no one in our audience will buy this type of thing but if you are buying it really the main competition you're looking at at least

That's unlocked is a 31 to 35 X we have that in our legacy charts which we'll be running today not in the code compiled chart though and that's a 28 core CPU which ranges in price these days fifteen hundred to three grams depending where you check and when and the boards are completely ridiculous to buy – one of the only ones on the market is the sr3 dark and that's like $1,800 so how even that ends up costing a bit more than the 64 court option from AMD and finally I'll give a quick shoutout here – it's a to others if you're interested in the Linux side of things Linux test

You know there's going to be interested in that for this part we're simply not going to do it for today maybe in the future but there's not enough time right now before a trip if you're interested in it check out fir onyx or a tech gauge both of them have Linux tests they're very in-depth and we'll leave that to them for now all right so let's get into it we'll start with the newer tests we'll look at some code compile look at how frequency scales across various thread intensities for workloads some thermal numbers which are very interesting for

This one and then we'll go through legacy benchmarks for the workstation applications that we have for now our first test from the new suite is our compile benchmark which has been widely requested we previously introduced code compile testing with GCC but we scrapped it and have instead introduced compile testing with chromium build time we'll talk more about the specific testing methodology and the parameters in our formal methodology unveil video after our trip to Taiwan but we wanted to show this in advance for the relevant 3990 x comparisons this chart doesn't have many

Parts on it yet but we have other charts that are more detailed later on for this one the 3990 x shows direct scaling and isn't held back before it's able to leverage the course we did find that the CPU really needs a lot of RAM for this though 64 gigabytes seems to do the trick our standard 32 gigabyte configuration was paging out to the drive and massively slowing down as a result the CPU in this Ram constrained scenario is too fast for its own good if it doesn't have enough memory and will actually perform worse than a 10 900 X with the bottleneck removed here the

Compile time with 64 gigabytes of RAM comes down to 22.3 minutes compared to 31.6 minutes on the 39 70 X 32 core CPU that's a time reduction of 29% so it's not a linear reduction or even really that close to a linear reduction but it's not bad overall against the 10 9 8 exe 18 core CPU the 39 70 X improves by over 20 minutes of difference reducing compile time by 42% the improvement to the much more expensive 39 is about 60% less time required although we'd like to see more price comparative benchmarks soon and we'll do that once we get back we can add the 39 50 X to

These charts once we return and for reference as well we ran the tests on them 10 900 X with 64 gigabytes of RAM and noted that there's no meaningful uplift it only starts to matter with the 3990 X for this one the 32 gigabyte ram allowance is not exceeded by anything except for the 3990 x it be worthwhile next to demonstrate the frequencies on the 3990 x under a few different circumstances as well as the temperature we'll start with a single-threaded maximum frequency running Cinebench RR 20 single-threaded against the 39 90 X the CPU ended up holding about 43 25

Megahertz on average with peaks at 40 to 50 megahertz single core and dips down to 40 300 megahertz ignoring one fall to 40 275 during work loads of limited thread counts this is where you can expect the frequency to rest for a single core load like Cinebench r 21 t we next ran a multi-threaded benchmark just for fun let's produce a completely useless chaotic chart with 64 lines on it on the screen to sort of illustrate the legibility of data we were dealing with if you feel overwhelmed looking at this that's how we felt when first looking at the log file after the first

Time of working with a 64 core 128 thread CPU one useful thing does come out of this though and that's the average all core frequency that we can plot on a new chart the average clock across all cores is about 31 30 megahertz to 31 60 megahertz with the maximum a single core boost during this time typically being 3800 megahertz which spikes to 4000 megahertz occasionally a quick thermal test is next this one is really just to establish what level of cooler is expected to keep the 3990 accent check we aren't doing a comparative test

Versus any other cpus for this so it's just a one-off to give an idea of cooler needs for this one we're using a refurbished Intermatic slick tech to 360 that we've drained and refilled it doesn't have gunk in it and so this one is one of the good ones remember that the lick tech series is actually one of the best solutions for thread Ripper when it comes to straight performance but that changes if gunk develops in your cooler so for straight performance if it's working well this is a good solution it's just not something we recommend purchasing at this point

But if you have one and you can clean it refill it with a better liquid then it's still a good option so the 30 I 90 X is interesting for thermal data the hottest CCD right about 61 degrees Celsius at steady state in a 21 degree ambient environment the first CCD reported an average TDI of 58 degrees at steady state with the next CCD about in line with the average but the third CCD plotting a significant CCD to CCD Delta in temperature the next two CCD is plot one in the lower ranks down below 50 degrees and one in the upper ranks just above 60 degrees ultimately at the end

Of the chart we end up with about half the CCDs in the cooler class and half in the warmer class of results this likely has to do with positioning of the dyes and also keep in mind that each CCD has eight cores in it and we can't see the individual core temperatures with AMD like we can with Intel so we're getting an aggregate with these CCD temperatures so either way this could be an instance where lapping the IHS could have meaningful decreases in the core to core Delta's we previously showed this in our 99 80 X eel app in benchmarks and it should carry over similarly here because

That temperature Delta CCD to CCD is typically representative of a difference in depth of the IHS blender is our first workstation application from our legacy benchmarks although this one's carrying over to the new stuff to because it's so good but under spawns one tile for every threat on the cpu which means that thread Ripper and the 64 cores of the thirty-nine 90 X scale in a predictable fashion but not quite linear versus the 39 70 X for our more intensive GN logo render and blender the 3990 x requires four minutes to finish each frame a reduction of 37.5% time required from

The 3970 x's 6.4 minute requirement that's also a reduction of 29.8% versus the overclocks 3970 axis result a 100% increase in cores obviously doesn't get you a 50 percent decrease in render time but at almost 38 percent stock to stock it's pretty damn close when considering all the other changes in the CPU aside from just the course compared to the 31 75 X at nine point three minutes the 3990 X finishes the render in about 57% last time the 31 75 X has advantage in overclocking but that's only enough to get it past the stock 3960 X and to begin approaching the

Stock 3970 X as for Intel's more commonly available 18 core CPU the ten nine eight exe finishes in about the same time as the stock 3175 X when the ten nine eight exe is set to four point nine gigahertz ours was a really good chip but it requires fourteen minutes when fully stock putting it just behind the 39 50 X you'd be better off with a 39 50 X if you don't need the PCIe lanes and if you want to spend less than say two to four thousand dollars on a processor this is probably your next best choice for quick reference we also ran our logo render on the primary

Production PC we use in the office except using CUDA only it's not a like-for-like it's intentionally different that resulted in a two point eight minute render time with the test w 100 GPUs for our blender GN monkeyhead render the stack shuffles a bit and changes from the results of our test design but this isn't too different for the HDD T parts the 3990 X requires three point seven minutes per frame to render this animation with the stock 39 70 X CPU at 5.6 minutes for its render time that's a stock to stock reduction of 34% render time benefiting the 3990 X

Overclocked in the 39 70 X reduced its render time by five point three percent but significantly increased cooling and power requirements so we don't really think it's worth it the 3990 x requires 46 percent less time than the 39 60 X furthering that it's not one-to-one scaling with cores and blender but that there are still improvements to be had we can next cover a common workstation application that's less likely to care but has been growing to support increasing core counts over the past few years using Adobe Premiere and Adobe Media encoder to render our

1080p convention floor reports the a.m. the threader for 3990 X CPU illustrates that we've encountered an immediate limitation of scaling against the threader for CPU Intel 3175 X is a technical leader here but it loses in value and we don't really recommend the processor given the combined board and price and limited victories and benchmarks we already talked about this one in our 3125 X revisit from a few months ago where we generally recommend the 3970 action step dr. 3175 axe current leads the 3990 acts by about an 8% reduction in total time

Required where appropriately it's about tied with the 39 70 X which leads the 3990 X by marginal marks in this test the 3990 X's additional core count isn't helping it and the higher frequency parts at or below 32 cores have the advantage more cores isn't always more better but in some applications it will be blender was one of those and we have others coming up to finish premier first our 4k 60 b-roll render shuffles the deck a little bit but not by much the 3990 ax ends up a worst performer than the 3970 acts or 3960 X CPUs you 3175 X completes in about 5.7 minutes or at

5.10 an overclocked with the 3970 X completing the render in 6.1 minutes or five point nine when overclocked overclocking thread Ripper really isn't worth it for these applications the 39 DX is between a 99 8 exe and 3970 ax and you probably shouldn't buy the CPU just for doing similar adobe premiere video production workloads to us there are cheaper better options like the 3 970 x4 39 60 X the 10 9 8 exe ends up a bit behind the 99 ADRC which is at least partially thanks to vulnerability mitigations baked into the silicon the biggest improvement has been against the

2990 WX which was one of the worst CPUs we've worked with for applications like premiere largely thanks to memory latency issues but we talked about that resolution in the 370 x review for another legacy test our compression and decompression results demonstrate clean scaling across the CPUs and also helps show memory bandwidth limitations on some products starting with compression we can see where the threader / CPUs begin to trip over themselves and choke on memory bandwidth until 6 channel 31 75 X pulls far ahead well d 39 70 x 3960 axon 3990 acts are all right around the

Same mark due to architectural bottlenecks they are clearly limited here the 3990 x doesn't offer any compression advantage in this particular workload versus the 39 70 X but the 3175 X stock CPU runs about a 12 percent increase in millions of instructions per second over the average threader for 3000 series part overclocking the 31 75 X provides a scaling benefit of 9% in this scenario thread refers limiting itself and has no room for improvement so it won't be I am the modifies its memory architecture or channeling again until

We see a jump in threader for performance for this benchmark decompression is a different story and this workload the 3990 ex runs just ahead of the 3970 ex stock cpu our current belief that this is a result the closeness that is of the frequency deficit in exchange for cores at which point we may be encountering diminishing returns because once you look at run to run variation they trade places and some rods but overall the decompression performance of thread Ripper is extremely strong as a comparative against the other parts versus the

Previous chart and looks good against Intel's parts for this one we had to expand the chart scale just to accommodate thread Ripper which is always a good thing for the CPU but a bad thing for the legibility if there's one production workload we can rely on for not scaling against cores its Adobe Photoshop for this test and we've noticed that the application scales most directly with frequency not core count and that's always easily demonstrated by highlighting the 99 hundred K and 97 hundred K both at five point one gigahertz where no benefit is visible

The 39 idx unsurprisingly doesn't do particularly well here it's capable of running Photoshop workloads just like it's capable of running gaming workloads but it's also about equivalent to a 10 900 x10 core CPU a 9700 k-8 core CPU or a 3900 X 12 core CPU and the 12 core is with an error of its own overclocked counterpart for this one if you're only doing Photoshop workloads like heavy filters transforms of large images smart filters or similar effects then you should be buying a 1900 K the 1098 exc with an overclock would be a good fit for a heavy Photoshop user who also

Needs some more cords for other workloads but you're really not buying a 39 d ax for this and before anyone cries about well why did you even test it with photoshop then we don't remove things from our benchmarks for any other parts so we're not going to do it for AMD's 39 DX either at the end of the day this does help illustrate that more cores doesn't always just get you chart-topping performance because that belief is what will cause people to waste a lot of money if they're buying for example only for Photoshop the last of our legacy tests is v-ray which

Closely and unsurprisingly mirrors the performance of blender the 30 are 90 X 128 thread CPU ends up at a render time requirement of 13 SEC a reduction of 30.5% from the 3970 acts at 4.25 gigahertz or a reduction of 35% from the stock 3970 acts as a refresher blender showed us about 34 percent to 38 percent render time reduction so this is in line with what we'd expect the 30 or 35 acts manages to mostly keep up with the 30 and 70 X's 20-second render time when running that for poen gigahertz but falls behind that stock so that's the 3990 X overall the CPU is in a class

Where it's difficult to make commentary on value you can of course always do it the 3175 X for example is poor value the 3990 X for a lot of people would also be poor value because most people not only don't need CPU with this many cores or threads but will also experience regression and performance because you are trading things off for that extra core count and all the threads that the 3990 X has so this is a scenario where these people probably exist more so than genuine users of a 3990 X or a server like xeon type part the people who buy this or who had caution against buying

It anyway might be the type who just want a really expensive part because they have a lot of money to burn there's nothing wrong with that and they just want the best or the highest end computer or whatever but if you are that person just be aware that in a lot of applications like standard applications a 3990 X is not going to be better so if you have specific workloads like certain types of file compression or decompression workloads although you get bottlenecks with some of those as you saw in these charts or if you maybe do a tile based rendering at which point the

3990 X is actually extremely competitive with even GPU rendering and it can have more memory than GPUs do so that's a benefit to or if you do things like code compile those are areas where we've seen a clear uplift and other benchmark sites that focus more on Linux or workstation would show you even more still but a lot of stuff won't show an uplift so don't just buy it because it's expensive and you want to sort by highest price and Add to Cart that's not a good way to build a computer when you have stuff like this out there there's probably better choices for you if you don't know

You need those those threads that the 390x has if you're more on the enthusiast side of things and you still want this CPU either for competitive overclocking or because you actually need it for whatever you're doing one of the enthusiast type things you could look into would be potentially doing some IHS lapping now that does void the warranty because when you send it back they're gonna know that you modified it unless you save all the nickel dust and then glue it back on but IHS lapping here we saw working with Joe on the 3970 ax would be the same idea a significant

Uplift in the consistency of the thermals across the chip le'ts with lapping the IHS you're not necessarily going for a direct massive gain and thorough performance there will be one but in a lot of cases it's not it's not really anything a lot of people will probably on a sanding the thin for hours and go really that's it but you're looking more for a reduction in the die to die or the the CCD to CCD Delta's so that's a project you could do that would have some benefit here as we've shown on the CCD thermal charts the they are about 10-15 degrees different sometimes

That's a lot better than we've seen with core decor Delta's on the ninety nine eighty XE but we can't see court decor here only CCD CCD so that's a project you might like now finally on the workstations side of things I also want to mention a lot of people whenever we have the 3175 acts and charts think that they're being super clever by saying it's not fair to benchmark an Intel server part a Z on CPU against an AMD non epic part shouldn't you be using the epic for that comparison the answer is no the 3175 X that X at the end of it is important the Xeon w31 75 X is billed as

An enthusiast part so it was marketed as an enthusiast overclocking unlocked part that's the goal of the chip it's not targeted for server users it's targeted for workstation enthusiasts like a hybrid of the two so just because it says Xeon doesn't mean it's server epochs a bit of a different story you don't have to be server with epoch either but it doesn't quite work so it's a fair comparison and it's certainly a fair price comparison and ultimately it doesn't matter because we're not recommending the 3175 acts in

Almost any use case anyway it's just not good value especially with the motherboard price factored in so you're left with basically choosing between a 39 78130 990 and for most people if you're genuinely a workstation user 3970 will get 90% of the performance and a lot of the applications it might better in some applications and then there's a select few like blender tile-based renderers or maybe code compiled where you start really showing a difference but it's not a linear one-to-one scale you don't get a doubling of performance because you've

Doubled your cores and hopefully our charts will help you determine if it's worth spending the extra 2 grand ultimately it's probably going to come down to the question of do you make money with your computer and if you're doing one of those applications where the scaling is pretty reasonable like 30% plus versus a 3970 ax and you make money with your computer and getting projects done faster will be meaningful to your ability to bring in more jobs then it's probably worth it and if you don't well you might be able to just run with a chip that cost half the price so

That's it for this one thanks for watching our thanks to the unnamed youtuber for sending out the 3990 X so that we could basically pull it out of a box and put it under liquid nitrogen that was insane but we'll try and do more of that soon probably with a different chip at that point or we'll have to send it back and then request it again but you all can continue trying to guess who sent it until they decide to come forward about it in the comments below post your guesses it'll be fun to see if anyone gets it right but I won't be able

To tell you so that's it for now subscribe for more go to store documents XS net to help us out directly or and exes we can get behind-the-scenes videos where we showed for example some of the early testing for this stuff before we even publish this review and then also check back for our factory tour series that's getting refreshed in the next week or two I'll see you all next time

Related Videos

because the Atari VCS is definitely something that's coming out we need to do a teardown on one to get some perspective and Atari if you don't know it&...
Hey how's it going guys Jack I'm out here with the toasty rose and today we're going to be doing a four hundred and fifty dollar PC that any of you ...
The biggest rule in testing coolers is to never trust anything don't trust the numbers don't trust the software don't trust the firmware and definit...
Hey what's up guys Jack and Matt here with the toaster brothers and today we're gonna be finding out if this 350 dollar laptop can game this is gonna be...
Everyone welcome back to another episode of ask GN sorry I'm not doing the intro with snowflake for this one we're flying to Taiwan and we need to get t...
Everyone welcome back to another hardware news recap for the week we are in Taipei right now working on a factory tour series and we actually have a point for y...
Hey how's it going guys Jack and Matt here with the toaster bros and today we're gonna be doing our first 20 20 thousand dollar gaming PC but guess what...
Coursers a 500 has been an embarrassing show of performance for the company's first revisit to air cooler isn't a long time but the upside is that it ma...
When I received the thread Ripper 3990 X on loan from a yet unnamed youtuber we quickly threw together an extreme overclocking stream and got the CPU to about f...
Every one so we're in Taiwan right now we're doing our factory tour series and while we were out here we stopped by Lian Lee's offices and they want...
Hey what is up guys Jack and Matt here with the wait what do you do it on sex you back your desk back to your desk sorry about that okay hi Jack and Matt here w...
We thought we'd found the limits of viewer interest when he posted our tour of a screw factory last year as poor Freight it was more of a screw shop than a ...
Everyone welcome back to another hardware news recap for the week we're in Taiwan where we venture outside to film things that's always a nice change of...
Hey how's it going guys Jack and Matt here with the toasty rose and today we're gonna be doing a 350 dollar gaming PC build it's gonna be really awe...
out of all the factory tours we've done until today we'd never gotten an opportunity to see how water blocks and open-loop cooling components like fitti...
Paints is one of the chief processes to dial in for product manufacturing and today's tour will bring us to a taiwanese factory replete with a mix of automa...
hey what's up guys Jack I'm Matt here with the tasty bros and today we're gonna be showing you our 2020 CES recap live from our CES Airbnb but befor...
except you can't wipe your ass with it well I mean we're going to take a gamble on the YouTube D monetization system and talk about the impact of corona...
everyone we're closing out our trip in Taipei for our factory tourists we have a lot of factory footage live already and we have more coming to the the chan...
hey what's up guys Jack and Matt here with the Thai stereos and today we're gonna be finding out is a six hundred dollar laptop worth it in 2020 it'...