Review Index:
Feedback

The AMD Ryzen Threadripper 2950X and 2990WX Review: Moving Forward

Author:
Subject: Processors
Manufacturer: AMD

Widening the Offerings

Today, we are talking about something that would have seen impossible just a few shorts years ago— a 32-core processor for consumers. While I realize that talking about the history of computer hardware can be considered superfluous in a processor review, I think it's important to understand the context here of why this is just a momentous shift for the industry.

May 2016 marked the launch of what was then the highest core count consumer processor ever seen, the Intel Core i7-6950X. At 10 cores and 20 threads, the 6950X was easily the highest performing consumer CPU in multi-threaded tasks but came at a staggering $1700 price tag. In what we will likely be able to look back on as the peak of Intel's sole dominance of the x86 CPU space, it was an impossible product to recommend to almost any consumer.

Just over a year later saw the launch of Skylake-X with the Intel Core i9-7900X. Retaining the same core count as the 6950X, the 7900X would have been relatively unremarkable on its own. However, a $700 price drop and the future of upcoming 12, 14, 16, and 18-core processors on this new X299 platform showed an aggressive new course for Intel's high-end desktop (HEDT) platform.

This aggressiveness was brought on by the success of AMD's Ryzen platform, and the then upcoming Threadripper platform. Promising up to 16 cores/32 threads, and 64 lanes of PCI Express connectivity, it was clear that Intel would for the first time have a competitor on their hands in the HEDT space that they created back with the Core i7-920.

Fast forward another year, and we have the release of the 2nd Generation Threadripper. Promising to bring the same advancements we saw with the Ryzen 7 2700X, AMD is pushing Threadripper to even more competitive states with higher performance and lower cost. 

View Full Size

Will Threadripper finally topple Intel from their high-end desktop throne?

Click here to continue reading our review of the Ryzen Threadripper 2950X and 2990WX.

  Threadripper 2990WX Core i9-7980XE Threadripper 2950X Core i9-7900X Core i7-8700K Ryzen 7 2700X
Architecture Zen+ Skylake-X Zen+ Skylake-X Coffee Lake Zen+
Process Tech 12nm 14nm+ 12nm 14nm+ 14nm++ 12nm
Cores/Threads 32/64 18/36 16/32 10/20 6/12 8/16
Base Clock 3.0 GHz 2.6 GHz 3.5 GHz 3.3 GHz 3.7 GHz 3.7 GHz
Boost Clock 4.2 GHz 4.2 GHz 4.4 GHz 4.3 GHz 4.7 GHz 4.3 GHz
L3 Cache 64MB 24.75MB  32MB 11MB 12MB 20MB
Memory Support DDR4-2933 (Quad-Channel) DDR4-2666 (Quad-Channel) DDR4-2933 (Quad-Channel) DDR4-2666 (Quad-Channel) DDR4-2666 (Dual-Channel) DDR4-2933 (Dual-Channel)
PCIe Lanes 64 44 64 44 16 16
TDP 250 watts 165 watts 180 watts 140 watts 95 watts 105 watts
Socket TR4 LGA-2066 TR4 LGA-2066 LGA1151 AM4
Price (MSRP) $1799 $1999 $899 $1000 $349 $329

One thing that is clear from the specifications above is how stale Intel's HEDT lineup now seems compared to Threadripper. While Skylake-X and the original Threadripper were announced at the same event, Computex 2017, and the i9-7900X launched before Threadripper, we have yet to hear any information out of Intel about a refresh.

Even more so, Intel's current top-of-the-line HEDT offering remains based on the Skylake microarchitecture. While Kaby Lake and Coffee Lake didn't provide IPC improvements, these high core count processors are still missing any frequency advancements made in subsequent generations and (slight) process node improvements.

On an architectural level, the 2nd Generation Threadripper processors are using the same Zen+ core that we first saw launched with the Ryzen 7 2700X. As a refresher, Zen+ is mostly a process node shrink, from 14nm to 12nm, but also provides some improvements to cache latency, support memory frequencies (DDR4-2933 in this case), and IPC. For more details, you can check out our review of the Ryzen 7 2700X and Ryzen 5 2600X from earlier this year.

While 2nd generation Threadripper doesn't introduce any more new architectural elements to the Zen family, it does have some stark changes compared to the original Threadripper in the form of segmentation between the Threadripper X and WX series.

View Full Size

The Threadripper X-series, consisting of the 2920X 12-core and 2950X 16-core processors are being targeted by AMD to more of a gamer/enthusiast audience, just like the previous Threadripper.

View Full Size

Constructed identically to the original Threadripper processors launched last year, the X-series consists of two Zen+ dies, each featuring 2 Core Complex (CCX) units. These two dies, each containing a dual-channel memory controller and 32 lanes of PCI Express, are connected through Infinity Fabric.

While this Infinity Fabric link does provide some detriments in the form of latency, these ramifications are well known at this point, as we discussed in our Threadripper 1950X review. This flexible configuration also allows the user to choose between a Unified (UMA) or Non-Unified Memory Access (NUMA) arrangement.

However, the real changes with 2nd generation Threadripper come with the introduction of the WX-series

View Full Size

Geared towards workstation and professional users, the WX-series Threadripper is a hybrid between the Threadripper and AMD's server CPU offering, EPYC. 

To hit core counts of 24 and 32, AMD needed to move to a 4-die configuration for the WX-series processors. However, this presents some interesting challenges. If AMD decided to move to a full EPYC configuration with a total of 8 channel memory controller, and 128 lanes of PCI Express, then compatibility among different Threadripper CPUs would be shaky at best on the motherboard level.

View Full Size

Instead, AMD is simply connecting the additional two Zen+ dies through Infinity Fabric, and ignoring their memory controller and PCI Express capabilities. These new dies depend on the rest of the processor for both memory and PCI Express access.

Due to this, the WX-series Threadripper processors must remain in a NUMA configuration, and present themselves as four individual NUMA nodes to an operating system, akin to a quad-CPU system. Additionally, the Infinity Fabric link between each of these dies is effectively running at half the speed of the 2-die arrangement found with the X-series processors.

While the highly expandable of the Zen architecture afford AMD the ability to create the first consumer 32-core processor at a relatively affordable price, there are some notable potential drawbacks to this approach, namely memory latency.

Review Terms and Disclosure
All Information as of the Date of Publication
How product was obtained: The product is on loan from AMD for the purpose of this review.
What happens to the product after review: The product remains the property of AMD but is on extended loan for future testing and product comparisons.
Company involvement: AMD had no control over the content of the review and was not consulted prior to publication.
PC Perspective Compensation: Neither PC Perspective nor any of its staff were paid or compensated in any way by AMD for this review.
Advertising Disclosure: AMD has purchased advertising at PC Perspective during the past twelve months.
Affiliate links: This article contains affiliate links to online retailers. PC Perspective may receive compensation for purchases through those links.
Consulting Disclosure: AMD is a current client of Shrout Research for products or services related to this review. 

August 13, 2018 | 09:32 AM - Posted by Power (not verified)

Testing with broken synthetics benchmarks?

August 13, 2018 | 10:36 AM - Posted by Allyn Malventano

This comment says nothing without elaboration.

August 13, 2018 | 11:58 AM - Posted by Power (not verified)

"In Geekbench's multi-threaded test, we see some unexpected results. It appears that the benchmark has issues handling massive amounts of threads, like the 64 found in the 2900WX, making the 16-core 2950X score higher in comparison."

August 13, 2018 | 12:19 PM - Posted by Allyn Malventano

Ok, so the benchmark had an odd behavior with 64 threads - something that can also negatively impact other applications (which is why there is an option to disable half of the threads on this CPU). Not only was the oddity disclaimed in the text, but omitting it removes some vital context in that higher threads, with current software, sometimes causes problems.

August 13, 2018 | 02:19 PM - Posted by Power (not verified)

Still, synthetic benchmark is very poor choice to convey the point. Reasonable person does not expect AMD to optimize CPU architecture to run synthetic benchmarks.

August 13, 2018 | 09:31 PM - Posted by throAU (not verified)

Actually a reasonable person _would_ expect both AMD and intel to make sure synthetics run fast, because that's what reviewers use.

You need to read the words not just look at the graphs.

Its early days, windows scheduler probably needs tweaks for proper thread:core affinity with these higher thread counts, especially with threadripper 2 and its NUMA configuration.

August 14, 2018 | 04:42 AM - Posted by Power (not verified)

"Actually a reasonable person _would_ expect both AMD and intel to make sure synthetics run fast, because that's what reviewers use."
I am not against using synthetic benchmarks, I am against misusing them. The results presented have value only to Intel's marketing arm.

August 13, 2018 | 09:48 AM - Posted by Balthazar2k4 (not verified)

You guys should really consider using something like RipBot264 for media encoding tests. Then you could fully leverage all of the cores offered through distributed encoding instead of hitting the current 6-8 core limit that x264/Handbrake has.

August 13, 2018 | 01:04 PM - Posted by Ken Addison

Going forward, we are definitely going to review our media encoding benchmarks. In the past they showed good, realistic scaling, but I think it's important to explore other options as the core count of consumer processors keeps increasing at such a quick rate.

August 13, 2018 | 10:32 AM - Posted by Randy (not verified)

"While there are some software optimization and compatibility issues, including gaming,....."

Lets just stop right there. AMD isn't even marketing this for gaming and you want to critique the gaming performance?

August 13, 2018 | 10:36 AM - Posted by Allyn Malventano

...so we should just ignore gaming results then - because they don't look the best? I suppose we should also *only* do gaming tests on gaming CPUs, ignoring productivity and encoding, right?

August 13, 2018 | 11:28 AM - Posted by Dapple (not verified)

The AMD crazies will be out in full force after the lackluster 2990WX results being shown by ALL reviewers today.

August 13, 2018 | 12:06 PM - Posted by Power (not verified)

Gaming test would be complete only if you include compile time. Pretty please.

August 13, 2018 | 02:16 PM - Posted by Rocky123 (not verified)

Personally I was surprised that the gaming results were so good for the 2990WX CPU once it was put into compatible gaming mode. The idea of having to reboot to switch modes is not all that big of a deal since reboot times with systems with fast SSD drives is only a few seconds I fail to see any problem with that.

The I guess the real show stopper here for those that want to have a CPU that can do heavy work loads and also game well is the 2950x 16 cores 32 threads and most likely overclock to 4.1-4.3GHz and slight chance for 4.4GHz on some chips would be a beast of a CPU for those that want to work and play with out having to spend 2K on a CPU alone.

If I was in the mode of wanting a TR CPU I think the 2950x would be my go to CPU.

One thing that makes me wonder though is why didn't AMD use the extra PCI-e lanes in the TR CPU's clearly they are the same setup as the EPYC CPU's and have 128 PCI-e on hand but they disable half of them on the active CPU complexes and totally disable them on the 2 extra CPU complexes.

They could have used those PCI-E lanes for inter core connections like they do on the EPYC CPU's and the main board would not have been any wiser to this since this is internal to the CPU itself and not something the main board would need to take care of. It seems like a missed opportunity to squeeze extra performance out of the TR CPU's and get that latency down a bit on the 4 tier CPU parts.

I under the as to why they could not up to memory channels since that is something that also is designed on the main board itself and could not be changed and keep the CPU's able to work with the current release mother boards.

August 15, 2018 | 03:18 AM - Posted by James

They don't disable any PCI express on the two die Threadripper. It is only 2 x16 links per die, so each die has 32 links and two die have 64 total. The 4 die Threadripper just has the links on the two extra die disabled.

Also, the die actually have 4 links each for on package routing. Only 3 of them are used at any time though. It was designed that way to simplify package routing on for Epyc. The links ued for routing on package are lower power than those meant to run off package. The maximum routing distance for an on package link is only a few centimeters at most. The off package links need to be more complicated and drive more power since they may have to route long distances through the motherboard.

In an Epyc processor, they can use one of the pci express links to connect to a die in another socket. In both Epyc and 4 die Threadripper, the die are fully connected within a single socket. They use 3 of the 4 on package links to route directly to the other 3 die so there is only one hop latency. Using another link, especially the off package link, would probably just waste power with little benefit. All of the cpus are only one link from any memory controller in the system. There is just a difference between being 1 link away from a memory controller and being 0 links away from a memory controller.

For a dual socket Epyc system, they still only have two x16 pci express per die. For single socket, all of them are available, so 4 die x 2 links each is 8 x16 links (128 lanes). For the dual socket system, they use one link from each die to connect to and different die in the other socket. So, each die only has one link available for IO. Since you have 2 socket x 4 die per socket x 1 link per die, you still have 8 x16 links (128 lanes) available in a dual socket system.

August 13, 2018 | 02:29 PM - Posted by Randy (not verified)

My point is this review seems disappointed in the gaming performance of a part that really wasn't meant for it......hence they supplied another part for it that actually does it semi well.

August 14, 2018 | 12:14 AM - Posted by Dapple (not verified)

Per Kyle at BitWIt, "AMD mentions the word game, gamer, or gaming over a dozen times in their own reviewers guide." He goes on to show one of AMD's supplied slides that is titled "Gaming on the AMD Ryzen 2990WX", which contains gaming benchmark figures for the 2990WX.

SO tell me again how it's unfair to critique the gaming performance of this cpu....

August 13, 2018 | 12:43 PM - Posted by Isaac Johnson

Wait, I'm not supposed to play games with anything not marketed for gaming? My entire life is a lie.

August 13, 2018 | 10:50 AM - Posted by terminal addict

While Kaby Lake and Coffee Lake didn't provide massive IPC improvements

They didn't provide any IPC improvements. Clock speed and core count increased. Zero IPC difference.

Your chart shows the 2950X on AM4. Wouldn't that be sweet! :)

August 13, 2018 | 01:10 PM - Posted by Ken Addison

Fixed! Thanks :) It would be quite difficult to fit the 2950X into an AM4 motherboard.

As far as IPC improvements, you are correct. I was fondly misremembering Kaby Lake, thinking it brought at least *some* IPC improvements, but looking back it was purely frequency scaling.

August 13, 2018 | 10:45 PM - Posted by Dark_wizzie

If we're comparing 4 core to 6 core, then the 6 core has more cache. So I think it's partially correct.

August 13, 2018 | 10:53 AM - Posted by Sunain (not verified)

MP3 encoding? At least try to get into the 21st century with that. Use something like FLAC / AAC / Apple Lossless encoding for a legitimate test. This is the same issue you had with your AMD-Ryzen-7-1800X

https://www.pcper.com/reviews/Processors/AMD-Ryzen-7-1800X-Review-Now-an...

LAME is optimized for Intel processors, so it's an unfair benchmark to start, albeit a completely obsolete test for a CPU benchmark.
https://software.intel.com/en-us/blogs/2008/10/06/open-source-project-la...

Audacity and LAME MP3 encoding is NOT a common workload at all despite what Ryan said in the comments for the Ryzen 7 and as it is obsolete codec and encoding test. It just shows a bit of incompetence and a lack of listening to legitimate constructive criticism.

August 13, 2018 | 11:32 AM - Posted by Dapple (not verified)

This is just not true. You are speaking from your enthusiast bubble and forgetting the masses. Common mistake of internet schmucks.

August 13, 2018 | 12:10 PM - Posted by Allyn Malventano

"Audacity and LAME MP3 encoding is NOT a common workload"

Lemme stop you right there. Your 'legitimate constructive criticism' is flawed the moment you disavow commonly used applications in favor of those that aim to show your favorite product in a better light. If AMD's architecture has not caught up to decade-old optimizations (your link was from 2008), then that's AMD's problem to solve. Also, 'FLAC / AAC / Apple Lossless' are not MP3. Using Audacity to create an MP3 output is a common thing to do. We surely didn't pick the test because it favors one company over another. If AMD is not as fast at it, then it is what it is.

August 13, 2018 | 12:28 PM - Posted by Particle (not verified)

It's not like audio transcoding is a common task of the masses. They just consume media in whatever format is provided. A review for an HEDT processor doesn't have the average user as its audience.

For those of us who actually do transcode, his point is more or less on target. There isn't much reason to use MP3 anymore. Chances are good that anyone ripping a CD is going to make FLAC (or similar) master copies and produce from that whatever their media player(s) need.

Computers play FLAC. Phones play FLAC. Media players play FLAC. Not a lot of reason to play with MP3 outside of space constrained or legacy applications. I would expect Vorbis or Opus to be used there for the most part unless we're talking about some weird proprietary device like anything Apple makes. AAC for them. Still not MP3 though.

Sure, MP3 still exists and is used but you're overplaying its relevance in 2018 for the audience this review is relevant to.

August 13, 2018 | 01:14 PM - Posted by Ken Addison

To be fair, our MP3 encoding workload is actually taking a WAV recording of our podcast straight off of the mixing board and encoding it to MP3, as we do to distribute the podcast. This is where the idea for the workload came from.

While audiophiles have moved away from MP3, clearly there are still uses, particularly for creators looking to distribute in smaller file sizes. 

We will continue to reevaluate our audio encoding test going forward and make some changes to be more applicable to modern enthusiasts.

August 13, 2018 | 03:06 PM - Posted by Dark_wizzie

People can't really abx a good mp3 file to a lossless file anyways. MP3 is often the way to go. Well, Opus should be, but yeah.

August 13, 2018 | 02:03 PM - Posted by Sunain (not verified)

Exactly my point Particle. You are reviewing a HEDT processor. Review should focus on HEDT workloads, like Cinebench, Autodesk products rendering, video encoding with h264/h265/AV1, audio encoding using modern professional codecs like AAC/FLAC, Bluray authoring, ect.

The reason AMD didn't put any development into MP3 or LAME encoder is there is no point. It's obsolete and your benchmark is obsolete.

Allyn, unfortunately you don't understand that using obsolete software puts doubt on quality and legitimacy of the review. Why use out of date software?! I'm kind of baffled you are even trying to defend your point.

I don't favor one CPU over the other. According to your own benchmark for LAME, the 16 core Threadripper 2950x is exactly the same speed, 83 seconds, as an Ryzen 2700x and an Intel 18 core i9-7980x, which has the Intel optimizations with those Intel optimizations.

I'm just trying to help you guys increase the quality of your review from a readers perspective. If you don't want to take our advice to make it better and in turn, criticize your readers and commenters for helping, that's your prerogative.

Ken, you shouldn't be even distributing the podcast in MP3 at this point. AAC would have been a better test. Better quality at even lower bitrates and device compatibility is basically the same. Would also save you bandwidth when distributing the podcast.

August 13, 2018 | 02:26 PM - Posted by Rocky123 (not verified)

I think maybe what he meant to say is lame or whatever it was is optimized for Intel CPU's. With that said it is either up to AMD or the maker of the program to come up with updates that would help it be as optimized on AMD platform as well.

If that never happens then there are a ton of programs out there that are just as good or better that will work well on both platforms.

Either way the time it takes either platform to do this tasks is many times faster than it took to do back in 2008 so I hardly think you would be sitting there very long for things to get done or happen.

August 14, 2018 | 10:49 PM - Posted by Paul A. Mitchell (not verified)

Allyn, are you executing the same binaries when you run these programs on Intel and AMD? A lot of machine instructions can be optimized by high-level language compilers when the compiler knows the intended CPU. Thus, it is not unusual for one set of optimizations to work better on one but not the other CPU model. As one simple example, I've been collecting comparable measurements over the years, and in that data set there are newer Intel CPUs which execute floating-point divides much slower than certain older Intel CPUs. This is not some fluke, but a comparison that repeats itself, when overclocking and when not overclocking, and no matter how many times I run that same software. Thus, a single benchmark that makes heavy use of FP divide instructions will run faster on the older CPU, whereas a benchmark that makes heavy use of integer divide instructions will run faster on the newer CPU. So, which CPU is faster? Answer: the expected result depends on the mix of instructions being executed.

August 13, 2018 | 12:08 PM - Posted by Chris Brunhaver (not verified)

Kind of an odd selection of benchmarks.

Why the 2-pass handbrake encoding still? This hasn't been recommend for years. Please change to single pass "constant quality".

I think that mp3 encoding is somewhat misleading because there are a bunch of batch converters that can load all of the threads. Additionally, menntioning that we're in a tech bubble (using something newer than an old MP3 encoder) while reviewing a SOTA 32 core processor is an odd comment to make. Someone investing in an $1800 cpu probably knows what a flac or opus file is.

August 13, 2018 | 12:15 PM - Posted by Allyn Malventano

Batch encoding is a good point, but that's not what is desired from the test, which is measuring how long you would have to wait for a *single* MP3 output to be rendered. That particular workload does not benefit from a batch encode. Also, while workstation users may know what flac/opus are, if they want to make an MP3, they will still make an MP3.

August 13, 2018 | 02:29 PM - Posted by Rocky123 (not verified)

@Allyn Malventano I just have a quick question. Do you guys plan on doing more tests down the road once either AMD or MS come out with patches to make Windows behave better with the high thread count on CPU's like the 2990WX. I am sure you probably would or will for sure just had to ask.

August 17, 2018 | 03:40 PM - Posted by Ces (not verified)

Hey Allyn,

Is there a way to test heavy multi thread jobs with the 2990WX compared to the skylake?
Example, encoding mp3 while running cinebench?
or playing a game while at same time streaming to a fixed file (converting)?

August 13, 2018 | 01:00 PM - Posted by Ken Addison

Handbrake is actually running 1 pass, constant quality. The chart with the 2 pass encoding is the X264 benchmark utility. I'll work to add some clarification to the review as to encoding settings.

As far as MP3 encoding, this test used to show more difference in our CPU reviews. As we discussed in our Ryzen 7 2700X review, AMD worked with the plugin developer for LAME for Audacity to implement some modern compiler advantages he was not taking advantage of.

With the LAME 3.100 plugin, there is very little difference between CPUs. Going forward, we will have to re-evaluate the inclusion of this test, but it seems likely that it will be eliminated. 

August 13, 2018 | 12:20 PM - Posted by BubsFanspeedSpikes (not verified)

Phoronix is doing some Linux Testing and some cooler testing for Threadripper 2 and the results look good for TR 2 on Linux.

But damn when I go there to read my CPU usage spikes way high, and that's all 4 cores/8 threads of my i7 mobile CPU.
Something is up with Phoronix and I'm using IE11/Windows 7 but the my CPU usage on normal websites never averages above 5% usage once the page loads. Anyone else having issues with much hogher than normal CPU usage in the browser when visiting Phoronix(20+% CPU usage all cores/threads)?

August 13, 2018 | 01:16 PM - Posted by Ken Addison

The Phoronix testing certainly is interesting. It seems clear there are some stark differences in how the Windows scheduler handles large core/thread counts and bespoke memory configurations when compared to the Linux kernel.

While we don't really do a lot of Linux testing here, I hope to do some additional testing with Threadripper in the coming weeks!

August 13, 2018 | 03:47 PM - Posted by LockedInTheVMsBigHouseCome2020 (not verified)

What I really want to see for the TR2 WX variants is some Windows 7 tests with a Windows 7 OS instance installed in a KVM/Xen Linux Kernel Hypervisor hosted OS environment for any folks that want to run Windows 7 safely and securely after 2020.

So start with a Linux Kernel based Hypervisor product like KVM/Xen and install windows 7 there in a hypervisor hosted instance and see how that runs on TR2 compared to Windows 7 bare metal. The 24 and 32 core TR2 WX variants appear to be ready made for running Windows 7 inside a Linux Kernel based Hypervisor similar to the way many busineses ran Windows XP after it went EOL. And plenty of businesses are still running XP inside of a Hypervisor(KVM/Xen) and keeping that OS instance running safely and securely to this day.

Phoronix does do some thorough Linux testing but 2020 is getting closer and Hopefully Phoronix will be doing the same in addition to the regular Linux OS distro testing. But there is going to be loads of folks looking to use windows 7 even after 2020, just like Windows XP if any mission critical business software is still not certified for OSs later than windows 7.

Some Small businesses may want to look at the TR2 WX 24/32 core variants and save 10s of thousands running windows 7 in a Hypervisor Hosted OS Instance after 7 goes EOLs as opposed to having to purchase or rewrite any new mission critical software that's certified for any windows OSs newer than 7. XP was popular that way more for saving businesses the costs incurred in sepending for new mission critical software upgrade/certification costs that ran many time the costs of any new hardware.

Windows 7 is going to be treated like windows XP, Microsoft or not, because the cost of certifying for a businesses mission critical software on any new OS is prohibitive, and some business software still to this day uses XP and that will carry on for for any windows 7 OS certified mission critical software. If the Money is not in the Budget for new OS certified mission critical software then there is that Running windows 7 locked down in a Linux Kernel based hypervisor just like it was done for XP.

Maybe Wendell at level1 Techs will be doing more of that also because there will be loads of customers interested in that just for windows 7 as was done for windows XP. That mission critical software refactoring and recertification for a new OS is so damn costly that many small businesses have no other choice if the funds are not there to afford such a costly process.

AMD may just make more TR2 WX variants sales that way and it's a lot less costly for a small business to use TR2 currently compared to Intel's Xeon, or even Intel's consumer HEDT variants. Now after the Zen2 based Epyc/Rome variants are to market maybe first generation Epyc/Naples is going to become somewhat of a better deal. Just as first generation TR/Ryzen parts went on firesale after TR2/Ryzen 2 parts came to market, maybe next year there will be some great deals on Epyc/Naples SKUs in 2019.

August 13, 2018 | 09:44 PM - Posted by throAU (not verified)

Windows 7 is not a supported platform for any processors/motherboards released in the past 2-3 years.

Some boards/machines will not even boot/run on versions of Windows prior to Windows 10, version 1670.

I don't like Windows 10 either, but continuing to test with Windows 7 is pointless as almost no one will be running these machines with Windows 7 as a host OS, it is not a supported configuration and you can't actually legally purchase it any more anyway.

August 17, 2018 | 03:42 PM - Posted by Ces (not verified)

How about adding some sort of Thread arrangement? perhaps programs like ProcessLasso?
It has saved my butt a few times with gaming and other applications that didnt do very well with my TR 1950X.

August 13, 2018 | 12:29 PM - Posted by WhyMe (not verified)

There's mention of a half core mode for the WX, was any game testing done in this mode? Only asking because I'm curious as to why AMD decided 1/4 mode would be gaming mode.

I'm guessing dropping one of the only two I/O dies with direct access to the I/O isn't detrimental to performance but what happens when the enabled I/O die needs to send an I/O request to something attached to a disabled I/O die.

August 13, 2018 | 12:31 PM - Posted by WhyMe (not verified)

How do games perform in half mode, does half mode enable both I/O dies or one I/O and one compute die?

August 13, 2018 | 01:06 PM - Posted by Ken Addison

Half-core mode was not tested due to time, but I would expect it to behave similarly to the 2950X in gaming scenarios. I hope to dig into this more in the coming weeks!

August 13, 2018 | 01:21 PM - Posted by Funkatronis

Ken, would you expect half core mode to emulate most of the benchmark scores found on the 2950X, including Sysbench? If you were to take a look into the crystal ball, would most professional software catch up in 18-24 months to take advantage of increased core counts?

Thanks.

August 13, 2018 | 02:53 PM - Posted by Johan Steyn (not verified)

Guys remember these guys are Intel customers. Sad but true. I really hoped this was behind them. I don't know why I even bothered reading this.

August 13, 2018 | 03:13 PM - Posted by dragosmp (not verified)

Hi, thanks for the review. I was hoping I'd see memory access latency like you did in past Ryzen reviews - intra-CCX, across IF, now in the "far" cores. Hope they'd come at a later date

...and I could bet you'd test single-core MP3 encoding and get smacked for it. I get what you're doing from a scientific point of view, it allows for a comparison to older CPUs in a ST workload. I wonder though, unless you manually set affinity, if you're not actually measuring the scheduling hoping around latency mess that Win seems to have going on when a single thread is loaded out of many. I run a similar task at work and get way better results by setting affinity to one core, which I assume turbos like crazy, rather than leaving the work to hop around. Food for thought maybe, but I think you know what you're doing even if you maybe don't wall'o'text us explaining the reasoning every time

August 13, 2018 | 03:23 PM - Posted by Anonymous1337 (not verified)

ASUS KIT IS NOT FREE! Spoke to rep, says you can buy from Amazon or Newegg, once they are available.

August 13, 2018 | 04:09 PM - Posted by fla56 (not verified)

Hmmmm so despite very powerful 1-click overclocking via XFR and PBO being available this hasn't been looked at?

why not? compare it to the all-core turbo on Intel, after all this is how enthusiasts will set it up

August 13, 2018 | 05:07 PM - Posted by Paul A. Mitchell (not verified)

I am honestly very amused by many of the comments above. I must confess that my age is probably the main reason why I am laughing.

Here's why: I began my love story with computers during the Summer of 1971 -- first year grad school. By the time I was professionally involved with computers, circa 1977, super minicomputers costing $250,000 came with one megabyte of RAM, expandable to two megabytes for an extra $40,000, and color monitors were $50,000 each. Cf. VAX 11/780

I also recall a good laugh I had several years ago with an IT specialist at Fry's Electronics: I quipped to him that the first IBM PC's CPU oscillated at 5 MHz. He promptly corrected me, saying it oscillated at 4.77 MHz!

If you had told me, way back then, that AMD would be mass-producing a 16-core CPU that oscillated at 4.0 GHz, I would have laughed in your face!

There may be some humans still alive today who do remember the many problems that came with programming 16-bit CPUs, when the sign bit really left only 15 bits of addressable memory space.

p.s. I think I will run the TR system I am planning to build at stock speeds, and focus my attention on storage acceleration: there should be plenty of idle cores available to process NVMe drivers.

August 14, 2018 | 04:09 PM - Posted by MarketingMajorYouAre (not verified)

I do not believe a word that you say as far as your educational credentials. Your excessive sycophancy has all the hallmarks of a sociopath with a pathological need for attention. We all know about computers being costly and the economy of scale that has occured over the years to bring down costs! So that's not news, but really stop with the marketing major types of excessive sycophancy.

Really Dennard scaling did not start to break down until 2006-2007 so around 4.77MHz scaling up to 4.0GHz was not that far off if you just plotted the Dennard scaling function on a chart, even with Dennard's scaling beginning to break down around that 2006-2007 time frame.

Go Laugh at that You Pretender!

August 13, 2018 | 05:31 PM - Posted by Particle (not verified)

An interesting note about 7zip compression results. Phoronix did some Windows vs Linux 2990WX testing and found the same general problem with the 2990WX under Windows 10. Under Ubuntu, however, compression results more than doubled from like 41,000 MIPS to 94,000 MIPS.

It's good to know that the issue isn't apparently an inherent weakness that makes the 2990WX unsuitable for that workload, but rather something more akin to thread scheduler shenanigans. If so, I expect such things will get patched in the future via Windows Update.

I have to wonder if the same core problem is impacting many of the surprising results from 2990WX that seem under par.

August 13, 2018 | 06:50 PM - Posted by ltkAlpha

There's one thing I'm still missing - and I apologize if it's explained in the review - why not enable one memory channel per die on the quad-die parts? As it is the 2990WX diagram looks... congestion-prone?

August 15, 2018 | 03:41 AM - Posted by James

The Zepellin die (all ryzen and epyc processors use the same die), has several features to make on package routing very simple, and therefore cheap; it only uses a 4 layer PCB. It actually has 4 on package links to connect die but only uses 3 of them. That was done just to simplify the routing. The memory interface pads are on one edge of the die such that they face outward on both sides. The die on one side are flipped compared to the other side. It would be a bit of a mess to try to route one channel from each die to the 4 channels that are routed on the motherboard. Also, for things that can take advantage of locally connected memory, going down to single channel could be a big memory bandwidth hit.

August 18, 2018 | 02:17 PM - Posted by Paul A. Mitchell (not verified)

Forgive me if this next question is really from outer space:

Is it even within the realm of possibility that AMD may be reserving 1 or 2 of the 4 dies for future embedded GPUs?

TR3-APU, perhaps?

I have no experience with circuit designs, as such, but it occurred to me that the wires connecting a CPU with the primary x16 PCIe 3.0 slot, could just as easily be confined to the TR chip's internal circuitry.

A TR with 2 x GPU embedded in the CPU would make for a very nice general-purpose workstation, even if those GPUs were not as powerful as many discrete GPUs.

August 13, 2018 | 08:21 PM - Posted by 0156 (not verified)

For me the biggest take away is that neither is slow by any means and at much less cost. These parts are all fast, it comes down to if you really need that little bit of extra performance or can we start spending money wisely?

August 14, 2018 | 02:15 AM - Posted by rukur

The linux benchmarks at phoronix look promising. Windows takes a bath vs linux as well.
Windows may been a patch to deal with the cores better.

August 14, 2018 | 06:40 AM - Posted by Prodeous@Work (not verified)

First of all, nice review. It does verify that currently my 1950x is sufficient for the work I'm doing and only few CPU's (that cost an arm and a leg) are only faster :) And even more so the 2990x not being a worthy replaement if one plans to do work and play games (and is unwilling to restart between the tasks).

Now i'm surprised by all the negativity here. And the unfortunate this benchmark and not that.

PCper spend a lot of their time, and I bet most of the weekends, testing these systems, providing impartial reviews, letting numbers speak for them selves.

And they do. 2990x is good for rendering, and high threaded tasks, not so good for gaming, and reduced threaded tasks.

I agree that optimization is the responsibility of both the developer and the hardware team. Looking at Blender, my go to 3D application required some work to get their Cycle OpneCL render engine to work properly on AMD GPU's. While CUDA worked well since Nvidia chipped in resources & GPU's to ensure there is proper compatibility.

Now after AMD investment in Blender, and using it as one of the key software to demonstrate CPU's, it ensured this software is equally optimized. And it shows that proper optimization (or at least some optimizations) brought up AMDs GPU's quite well. Bringing VEGA 64 performance improvements to nearly match the GTX 1080TI.

So i'm happy you guys are you sing the software of all types as this is the software common individuals would use.

August 14, 2018 | 07:47 AM - Posted by Humanitarian

No video review this time? Or we saving that for the podcast?

Also, are you guys getting paid for every successful recaptcha completion? I'm logged in to a verified account, why do I need to enter 8, yes eight, pick the road sign captchas to post?

August 15, 2018 | 05:16 AM - Posted by themysteryvan (not verified)

hmmmm…that's EXACTLY what a robot would say if were trying to convince you they weren't a robot.

August 17, 2018 | 03:53 PM - Posted by Ces (not verified)

This one does not understand your implications meat bag.

August 14, 2018 | 09:25 PM - Posted by Paul A. Mitchell (not verified)

I believe this next paragraph was an excellent introduction; others may disagree, and that is their Right too:

"Today, we are talking about something that would have seen impossible just a few shorts years ago — a 32-core processor for consumers. While I realize that talking about the history of computer hardware can be considered superfluous in a processor review, I think it's important to understand the context here of why this is just a momentous shift for the industry."

I'm wishing AMD the very best, mainly because this type of competition is good for users, and good for the industry too.

August 15, 2018 | 04:11 AM - Posted by James

"Due to this, the WX-series Threadripper processors must remain in a NUMA configuration, and present themselves as four individual NUMA nodes to an operating system, akin to a quad-CPU system. Additionally, the Infinity Fabric link between each of these dies is effectively running at half the speed of the 2-die arrangement found with the X-series processors."

What are yoh refering to here? AFAIK, it is fully connected in the 4 die threadripper, just like it is in Epyc. In the two die variant, you only have one link between the two die and that is it. In the 4 die variant, they have 3 links in use each to connect to the other 3 die with a single hop latency. I don't think I would refer to anything as half speed other than the memory bandwidth. I suspect that windows does not have the necessary NUMA optimizations to handle such a configuration properly anyway. I would be running linux on such a system. It gets a lot of use in HPC and can handle, in some cases, thousands of processor cores with a wide variety of memory configurations.

The mp3 encode as a benchmark does seem a bit odd. The gamming benchmarks, while not really odd, are of little importance. If you are going to buy a $900 or $1700 dollar processor for gaming at 1080p, unless you are using a software renderer, it would be a compelete waste. For game developers, this might still be a good system, assuming you are a developer capable of making your game perform well with many cores available, or at least not crash on start-up. As noted, windows looks like a problem here. It might have been good to test at 4k, just to see if it is graphics card limited, or whether the cpu is the bottleneck. It could hit windows scaling issues though. Also, nvidia's driver is probably a near worst case scenario on any system that doesn't have a single, last level cache. It seems to have a lot of fine grained, thread to thread communication. Maintaining a single last level cache with good latency is a major bottleneck to scaling to more cores, so it would be better in most cases if it would just go away, and developers would optimize their software for multiple core clusters They have to do that anyway for the consoles with similar 4 core cluster architectures. I wouldn't be surprised to see cell phones go with core clusters also, due to better power consumption.

Well, off to look for linux compile benchmarks on Threadripper.

August 15, 2018 | 11:39 AM - Posted by Paul A. Mitchell (not verified)

I stumbled on this customer review of the TR 1900X at Newegg:

"- Large 20% memory performance difference between NUMA and non-NUMA settings.

"On my system, the NUMA setting (memory interleave on) for some reason reduces CPU performance by
about 15%, while boosting RAM performance by 20%."

...

"Wish amd could improve the memory controller and reduce CPU-RAM latency to competitive levels."

Perhaps this is something to consider, and compare, when benching TR2 CPUs.

August 15, 2018 | 12:00 PM - Posted by Paul A. Mitchell (not verified)

p.s. I believe der8auer at YouTube switched memory interleave ON when running 2 x ASUS Hyper M.2 x16 add-in cards with 8 x Samsung 960 Pro NVMe SSDs:

https://www.youtube.com/watch?v=9CoAyjzJWfw

fast-forward starting around 7:30 on the counter
for the BIOS setup in that video

@ 8:07 on the counter: "Memory Interleaving"

August 15, 2018 | 10:36 PM - Posted by Paul A. Mitchell (not verified)

Paul Alcorn's recent article is a fun and easy read:
"AMD Ryzen Threadripper 2 vs. Intel Skylake-X:
Battle of the High-End CPUs" (August 14, 2018)

Post new comment

The content of this field is kept private and will not be shown publicly.
  • Lines and paragraphs break automatically.
  • Allowed HTML tags: <a> <em> <strong> <cite> <code> <ul> <ol> <li> <dl> <dt> <dd> <blockquote><p><br>
  • Web page addresses and e-mail addresses turn into links automatically.

More information about formatting options

CAPTCHA
This question is for testing whether you are a human visitor and to prevent automated spam submissions.