Date: February 7, 2020
Author(s): Rob Williams
AMD’s latest Ryzen Threadripper has arrived, and it brings along some impressive specs with it. We’re talking 64 cores, 128 threads, 288MB of cache, and a price tag of $3,990 USD. To kick off our look at AMD’s 64-core monster, we’re going to start with our deluge of tests under Linux, including compiling, rendering, and science.
If you’re interested in Windows performance, you may want to take a look at our experiences with the 3990X here.
It’s been a lot of fun watching AMD and Intel duke it out over CPUs the past few years, though it’s become obvious at this point which one has proven mightier. It was the fall of 2017 when AMD released its Ryzen Threadripper series, with the top-end 1950X sporting 16 cores. A month later, Intel followed-up with its Core X-series, with the 7980XE one-upping AMD by offering 18 cores.
After the launch of Threadripper and Core X, it seemed like AMD decided to take its gloves off. In late summer 2018, AMD released its second-gen Threadripper series, which introduced 24- and 32-core models. Intel’s attempt at competing against these chips soon came by way of the 28-core 3175W, a Xeon custom-tailored for enthusiasts.
Even though AMD had been offering a 32-core CPU, Intel’s 3175W was priced with a huge premium, something that’s suddenly become a lot harder for the company to get away with. Intel still offers the 3175W for around $3,000, which makes AMD’s 64-core at $3,990 seem downright alluring.
As we’ve said before, AMD has done well at recalibrating the CPU market. At a time when quad-cores seemed to be de facto, AMD came along and pushed six- and eight-core chips to the masses. After Threadripper and Core X debuted, the benefits that end-users could derive from these many-core chips became obvious really quickly. More cores means heavier work gets done quicker.
While the second-gen Ryzen Threadripper series was impressive, it was really the 12- and 16-cores that we recommended to most people, as the 24- and 32-cores introduced a design that detrimentally impacted certain workloads. Over time, this design was worked around by many software vendors, so some workloads improved – but not all. The much-improved design of the third-gen Threadripper makes that concern go away, even if we have five dozen or so cores.
With Windows threading being a little questionable at times with high core-count chips, Linux becomes a fun playground for pushing a CPU like the 3990X to its limit. At a time when the 2990WX struggled in many Windows tests, our Linux tests seemed indifferent, delivering far better performance overall. The same is largely true here with the new 3990X.
|AMD’s Ryzen & Ryzen Threadripper Lineup|
|3990X||64 (128T)||2.9 GHz (4.3)||288MB||Quad||280W||$3990|
|3970X||32 (64T)||3.7 GHz (4.5)||144MB||Quad||280W||$1999|
|3960X||24 (48T)||3.8 GHz (4.5)||140MB||Quad||280W||$1399|
|R9 3950X||16 (32T)||3.5 GHz (4.7)||72MB||Dual||105W||$749|
|R9 3900X||12 (24T)||3.8 GHz (4.6)||70MB||Dual||105W||$499|
|R7 3800X||8 (16T)||3.9 GHz (4.5)||36MB||Dual||95W||$399|
|R7 3700X||8 (16T)||3.6 GHz (4.4)||36MB||Dual||65W||$329|
|R5 3600X||6 (12T)||3.8 GHz (4.4)||35MB||Dual||95W||$249|
|R5 3600||6 (12T)||3.6 GHz (4.2)||35MB||Dual||65W||$199|
|Ryzen w/ Radeon Vega Graphics|
|R5 3400G||4 (8T)||3.7 GHz (4.2)||0.5+4MB||Dual||65W||$149|
|R3 3200G||4 (4T)||3.6 GHz (4.0)||0.5+4MB||Dual||65W||$99|
As we said at the launch of the 2990WX, it’s important to know who the 3990X is for. It’s not for your everyday enthusiast, regardless of how deep their wallets are. Most people who would buy the 3990X based on price alone would likely see worse performance for their uses. If you’re a high-end gamer that dabbles in creation, then the Ryzen 3950X is a good all-rounder. For those with gaming as a paramount concern, the even higher-clocked lower core-count Ryzens should be considered (such as the 3800X or 3900X), and of course Intel’s higher-end gaming chips (like Core i9-9900KS).
In Windows, threading success is largely going to depend on the software solution. In our Windows testing, we cover many consumer-oriented workloads, whereas for Linux, we gravitate more towards compiling, compression, security, and so forth, along with some shared tests, like Blender, HandBrake, and V-Ray. Threading improvements in the Linux kernel aside, a lot of the benchmarks we run in Linux have been designed with many-core CPUs in mind for a while, which helps greatly with scalability today.
Note that for this platform, the intended audience should be equipped with at least 128GB of memory; or, in other words, at least 1GB per thread. We don’t have a 128GB memory kit kicking around, so we instead used the same 64GB DDR4-3600 solution that the rest of the CPUs had been tested with. We were told by AMD that this would not detrimentally impact performance much, if at all, but that 128GB is still the suggested minimum.
If you’re looking to maximize memory bandwidth on this platform, you’ll want to try to go with four DIMMs rather than eight. AMD says that with eight DIMMs, the official supported DRAM speed is 2666MHz, whereas that increases to 3200MHz with four DIMMs. That said, there’s no hard roadblock that would prevent you from overclocking, but you shouldn’t expect anything beyond what’s officially supported – just in case.
And with all of that covered, we’ll quickly look over our testing methodologies, and then move right into a look at what the 3990X can do.
|Techgage Workstation Test System(s)|
|Processors||AMD Ryzen Threadripper 3990X (64C/128T; 2.9GHz)|
AMD Ryzen Threadripper 3970X (32C/64T; 3.7GHz)
AMD Ryzen Threadripper 3960X (24C/48T; 3.8GHz)
AMD Ryzen 9 3950X (12C/24T; 3.8GHz)
AMD Ryzen 9 3900X (12C/24T; 3.8GHz)
AMD Ryzen 7 3700X (8C/16C; 3.6GHz)
AMD Ryzen 5 3600X (6C/12C; 3.8 GHz)
Intel Core i9-10980XE (18C/36T; 3.0GHz)
Intel Core i9-9900KS (8C/16T; 4.0 GHz)
|Motherboards||AMD X399: ASUS ROG Zenith II Extreme|
AMD X570: ASRock X570 Taichi
Intel Z390: ASUS ROG STRIX Z390-E GAMING
Intel X299: ASUS ROG STRIX X299-E GAMING
|Cooling||AMD X399: NZXT Kraken X62|
AMD X570: Corsair Hydro H115i PRO RGB
Intel Z390: Corsair Hydro H100i V2
Intel X299: NZXT Kraken X62
|Chassis||AMD X399: Cooler Master MasterCase H500P Mesh|
AMD X570: Fractal Design Define C
Intel Z390: NZXT S340 Elite
Intel X299: Corsair Carbide 600C
|Graphics||NVIDIA GeForce RTX 2080 Ti|
|Memory||Corsair VENGEANCE (CMT64GX4M4Z3600C16)|
4x16GB; DDR4-3600 16-18-18
|Et cetera||Ubuntu 19.10 (5.5.0-050500 kernel)|
|All product links in this table are affiliated, and support the website.|
Our Linux configuration is simple overall. We’re using the latest version of Ubuntu (19.10) for our testing, which is fully updated, and upgraded to the 5.5 kernel. Out-of-the-box, Ubuntu 19.10 can’t boot on the latest Threadripper platform without the ‘mce=off’ boot flag, but once upgraded to the 5.5 kernel, that problem goes away.
Many of our tests are run with the help of Phoronix Test Suite, with tests locked to a specific version to generate repeatedly reliable data. We don’t do any real OS configuration outside of disabling sleep, and enforcing the performance power profile.
On AMD platforms, XMP is enabled, and that’s all that’s touched. There is no automatic overclocking performed by the EFIs on these platforms, as far as we can tell (when compared to reviewer guide performance). On the Intel platforms using ASUS, XMP is enabled, and “Yes” is chosen to let it optimize the settings, after which point the Turbo boost settings are reverted to Intel default behavior. This ensures proper memory timings and settings are enabled, and that no automatic overclock will take place.
All of the platforms we tested on had their EFI updated before testing, if one was available. While the current EFI available for our tested ASUS Zenith II Extreme from ASUS’ website will handle the 3990X no problem, we upgraded to an internal beta for this review, as it was expected to deliver better stability with DDR4-3600 memory.
Software compiling can be a tough test to get reliably scalable numbers out of, because depending on the project, there could be lulls in the process where a CPU could go virtually untouched for a while. We see the same thing sometimes in rendering, and especially photogrammetry.
No one is going to purchase a $3,990 processor for a compile that takes just 20 seconds, but it’s nice to see that the 3990X has managed to distance itself a fair bit from the 3970X in the Linux kernel compile test. Reduced gains are seen in the other compiles, but the end result with LLVM is pretty satisfying.
There’s a ton of rendering tackled above, but most of it paints the same kind of picture: in every single case, the 3990X soars to the top, albeit with different strengths depending on the test. In both Blender and V-Ray, we see exceptional scaling, with the latter in particular highlighting a nearly 70% gain. We admittedly wouldn’t expect to see that same gain in the real-world, but we happen to have real-world V-Ray performance en route for our Windows article, so we’ll see how the numbers agree.
We’re sure AMD will be pretty glad to see the Intel ray tracing results above, because when you’ve got more cores than the competition, chances are you’re going to do well. That’s even considering the fact that some of these tests utilize AVX-512 on Intel’s Core X-series. As we’ll see in our Windows results, the LuxMark test there doesn’t scale beyond 64 threads, even though it’s built with Intel Embree, which we can see here scales beyond that no problem.
Rendering is just one potential workload the 3990X has had thrown at it, so let’s check out some others on the next page:
We’re going to be tackling both video encoding and scientific performance on this page, and while that sounds like only two things, the reality is that there are a countless number of workloads available in each one, and loads of opportunity to see interesting scaling.
We unfortunately forgot about Blackmagic’s RAW Speed Test having a Linux build (again), so we’ll make sure that’s included next time. What’s new here video-wise is Intel’s SVT encoder libraries, all of which are open-source. For the latter half of the page, we’ll look at some scientific scenarios.
In our Windows HandBrake testing, we found that x265 performed a bit better with the 3990X over 3970X, whereas the roles were reversed in x264. With our Linux testing, we found… the opposite. Interesting how that works, huh? Ultimately, what we can derive from these first results is that video encoding is not going to see much of a gain at all in HandBrake, although that could change in time as development progresses.
It’s also worth noting that this was done with the 1.2.2 version of HandBrake, not 1.3.0, which is only because we couldn’t get the latest version working correctly in our install. We would kill for a standalone version we can just toss into an archive, without having to deal with flatpaks and the like – but we digress.
With Intel’s SVT, more cores matter. Unless we’re talking about VP9, apparently, as that doesn’t appear to be able to go over 64-threads. When that’s the case, it usually results in degraded performance, which lends some credence to our “know your workload” suggestion. We should point out the the 3990X will be slower in these thread-limited tests because it has an overall lower maximum clock compared to the lower core-count counterparts.
The 3990X hasn’t managed to double performance in any case so far, but has definitely made notable improvements overall. But, because of how the 3990X scales in encodes like Intel SVT, the 32-core would make more sense overall. That’s until these encoders can take better advantage of such large CPUs. For what it’s worth, we haven’t seen any stronger scaling in our Windows encode tests. In Adobe Premiere Pro, the codec matters a lot, with RED and ProRes alike both offering the best possible scaling.
Leave it to science to push many-core CPUs properly. In all of these tests, especially the top molecular dynamics tests, the scaling on the 64-core CPU is quite strong. The gains are less pronounced in the fluid dynamics test. If anything else is clear, it’s that it’s easy to go too low-end on your CPUs for heavy workloads. When you need to wait hours or days for results, you should probably get the fastest processor possible – budgets permitting, of course.
In reality, EPYC is much better suited for serious workloads like these, and while its top model still peaks at 64-cores, that platform offers the added benefit of an eight-channel memory controller, which would bode well for any workload that relies heavily on memory bandwidth.
The previous pages have been focused on specific scenarios, like science, rendering, and compiling, so to wrap things up, this page is going to revolve around some more general scenarios, like encryption, compression, and memory bandwidth. For good measure, a chess engine even makes an appearance.
We’ve been lacking more single-threaded tests in the past, but have some SciMark-sourced ones here. We’re still not sure if they are ideal, but they serve the purpose for now. As always, if you have a suggestion of another test we should implement, please leave a comment.
The first results from this page seem to agree with V-Ray Bench from the first page. We’re seeing nearly 70% performance improvement with both the OpenSSL and Blowfish tests. We wish we’d see this kind of scaling all of the time.
When the second-gen Threadripper came out, we quickly learned that so many cores would wreak havoc with certain software. At launch, the 32-core 2990WX didn’t gain much at all in 7-zip in Windows, but Linux accelerated it just fine. Interestingly, as we move forward to this 64-core launch, we’re not seeing much improvement at all. It even seems to suggest that this software is simply not going beyond 64 threads, but we used HTOP to validate that it was. This could be an example of software that needs to be tuned to scale higher than it’s currently able to. Higher memory bandwidth wouldn’t hurt, either.
It doesn’t seem likely that any chess fanatic would run out to purchase a 64-core CPU for chess engine use, but isn’t it great to see that such a niche function scales incredibly well? This scaling isn’t quite up to par with the OpenSSL result earlier, but +58% over the 3970X is nothing to balk at.
Memory bandwidth for the 3990X is similar to the other Threadrippers, but it lags a bit behind them overall. Intel’s top-end Core i9-10980XE shows some great strength here, although the roles reverse a bit in the Windows version of this test (run through SiSoftware Sandra). In both OSes, the 3990X sits behind the other Threadrippers.
Both of these SciMark single-threaded tests love Intel’s CPUs, with both of those included here keeping to the top of the chart. As mentioned at the top of the page, if you have ideas for other single-threaded tests in Linux, please let us know. The 3990X, almost not surprisingly, falls behind the others, despite having a higher single-thread clock speed than some of them. What you may lack in single-threadedness, you gain back handsomely with threads.
Across the few dozen performance charts here, we’ve seen AMD’s new 64-core Ryzen Threadripper 3990X perform great overall, at least if the workload calls for many cores, and can take proper advantage of them. As was the case with the 24- and 32-core Threadrippers in late 2018, the 64-core 3990X is definitely not a one-size-fits-all processor.
As talked about in the intro, even if money is no concern, you would not be better off buying the 3990X over a smaller Ryzen in all situations. You need a reason for the 3990X, and at the forefront, that involves uses like 3D rendering, mathematics, or even in some cases, compiling. Video encoding can also be included in this list, as long as the codec complements many cores. You should be good with the pro formats, but consumer formats will likely struggle to take advantage of more than 64 threads.
For those who can take advantage of as many cores as they are given, the Ryzen Threadripper 3990X is an outstanding processor. In most cases where multi-threadedness is properly supported, the gains of the 3990X over the 32-core 3970X are huge – certainly not double, but oftentimes hovering around 50%, and going up to 70%.
The fact that we’re not seeing doubled performance may be why there is no 48-core SKU in this updated Threadripper lineup. Both that and this 64-core would rub against each other too much. That said, AMD’s EPYC lineup has multiple 48- and 64-core options, all of which are priced above this 3990X.
It should be obvious at this point whether or not 64 cores is for you, and whether or not your needs justify its cost. Fortunately, if this chip doesn’t do the trick, AMD has the more modest 24- and 32-core options available.
When all is said and done, how could we not give this beast an Editor’s Choice award? It offers unmatched performance in many ways, and is a technical marvel at the same time. It was only a few years ago that AMD pushed eight-cores to the mainstream, and now we have a 64-core for the ultra-high-end. That kind of momentum is incredible, and means we might be waiting a little while for a return punch from Intel.
Copyright © 2005-2020 Techgage Networks Inc. - All Rights Reserved.