2nd Gen Threadripper 2990WX review: AMD’s 32-core CPU is insanely fast but not for everyone

Share on FacebookShare on Google+Pin on PinterestShare on StumbleUponTweet about this on TwitterShare on LinkedIn

 

AMD’s Ryzen Threadripper 2990WX crushes multi-threaded tasks like we’ve never seen before.

Stop, AMD. You had us at 32-core, 64-thread Ryzen Threadripper 2990WX consumer CPU.

You didn’t have to make it better by saying this weapon of thread destruction could be ours for $1,799 (available on Newegg or Amazon)—just $76 more than what Intel wanted us to pony up for its 10-core Core i7-6950X two years ago.

In Millennialese: That. Is. Just. Freaking. Insane.

But before you invest too much thought or cash into the concept of a 32-core CPU sitting in your PC at home, there’s a lot of caveats you need to know. Simply put: This CPU may be too much power for most of us.

threadripper 2 3

A Ryzen Threadripper 2990WX locked, cocked, and ready to crush multi-threaded benchmarks.

What is Ryzen Threadripper 2990WX?

If you’re wondering just how AMD went from an 8-core Ryzen 7 1800X to a 16-core Ryzen Threadripper 1950X to a 32-core Ryzen Threadripper 2990WX in the space of 16 months, while it took Intel three years just to go from a 6-core Core i7 to a 10-core Core i7, the magic is in the design.

threadripper parts

AMD has four new 2nd-generation Threadrippers on tap, but only one you can order today.

What makes a 32-core Theadripper even possible is the multi-chip design. Rather than the single contiguous or monolithic die approach that Intel takes, AMD CPUs are multiple chips joined together by the company’s high-speed Infinity Fabric. The original 16-core Ryzen Threadripper 1950X joined two 8-core chips together. With the 32-core Ryzen Threadripper 2990WX AMD joins four 8-core chips together.

2950x die topology

The 16-core Ryzen Threadripper 2950X features two active dies which each have their own memory and PCIe access.

This method comes with its own special penalty, though. Although the sTR4 socket for Threadripper is physically the same as the server socket used for AMD’s Epyc CPUs, sTR4 is wired to support four-channel memory using two of the dies, rather than eight-channel memory using four dies.

On Threadripper, that essentially means that of the four dies in the 2990WX, two are pure compute-only dies, without direct access to memory and to PCIe. Those two compute dies must talk through an I/O die that has PCIe and access to the memory. This design means an I/O die with memory access has 64ns of latency to memory, while a compute-only die has a latency of 105ns.

2990wx die topology

The new 32-core Ryzen Threadripper 2990WX features four dies connected by Infinity Fabric. Two of those dies must access RAM and storage through an adjacent die.

There’s also a reduction in bandwidth across the Infinity Fabric. On the new 16-core Ryzen Threadripper 2950X, which uses just two dies, the bi-directional die-to-die bandwidth is 50GBps. On the four chip, 32-core 2990WX, the bi-directional die-to-die bandwidth is halved to 25GBps.

For what it’s worth, that die-to-die bandwidth is apparently considerably lessthan a 7000-series server Epyc CPU’s 42GBps. What impact that lowered die-to-die bandwidth has on performance isn’t clear to us, nor is the reason for less bandwidth. It’s possibly due to how the CPUs are connected to the motherboards themselves. The server-focused Epyc CPUs feature 8-channel memory, with each die accessing its own set of RAM, instead of Threadripper’s shared approach.]

The Ryzen Effect

What that multi-chip design does do is enable core-count scaling at a rate unseen before. With Intel’s monolithic design, an 18-core CPU would require that every single die be nearly perfect and all of the cores functional to be sold. With Threadripper, to get to 32 cores, it just needs four functional 8-core dies. This all adds up to the crazy number of cores now available to consumers.

AMD has called this the “Ryzen Effect,” and you can see this on the chart below, where we mapped out Intel’s major CPU launches over the years and how many cores they’ve had. So what caused that spike in core counts in the last year and a half? It took Intel from 2008 to 2016 to go from 6 cores to 10 cores. It’s taken AMD 18 months to from 8 cores to 32 cores.

moar core wars

The Ryzen Effect is in full swing. Ever since the introduction of the AMD’s Ryzen CPUs in 2017, core counts of CPUs have taken off like an F-22 Raptor.

Don’t worry, it’s compatible (mostly)

With the 2nd gen Threadripper, there was much hand-wringing that the CPUs might not work with the existing X399 motherboards. AMD has said every single X399 board available today will work once you’ve updated the UEFI/BIOS to support the newer chip. All of the X399 motherboards support “BIOS Back” features, which let you update a board’s BIOS via USB without requiring an older CPU.

While all of today’s motherboards will work, they might not all overclock the same—but even there the news is mostly good. AMD said the main issue is the massive power draw of the 32-core and 24-core versions of the chip, so some board vendors have beefed up existing boards by offering cooling kits.

threadripper 2 2

A snazzy new lunchbox-sized case comes with the new CPU.

32-core Threadripper 2990WX performance

For this review, we spooled up the Ryzen Threadripper 2990WX in an MSI MEG X399 Creation motherboard with Windows 10 Fall Creators Update and 32GB of DDR3/2933 RAM. For graphics we run a Founders Edition GeForce GTX 1080 and the latest Nvidia graphics drivers available. Storage is a Kingston HyperX Savage SSD. Both systems were cooled with closed-loop coolers. The Core i9 used a Corsair 280mm H110i, and the Threadripper 2990WX used an Enermax Liqtech 240 TR4 cooler with cold plate that offered full coverage for the giant Threadripper. Both coolers were set to maximum fan speed.

For fairness, rather than recycle older numbers, we updated the original 18-core Core i9-7980XE setup that we used in that CPU’s review with the same version of Windows, newer Nvidia drivers, and the newest BIOS. The last detail is key, as it’s been some time since the original Core i9 review, and we were curious as to whether its performance had improved with a newer BIOS.

The last time we compared Ryzen Threadripper vs. Core i9, Intel’s 18-core Core i9-7980XE took home the prize for performance (although not for value). This is the one to beat.

For context, we’ve included scores for some CPUs that were run on a previous build of Windows. The numbers haven’t shifted, so they’re still valid. We’ll note where you might want to dismiss results for older chips, or we’ll simply exclude them if we think they don’t apply.

threadripper 2 4

Here’s what you get with a new Threadripper: a torque wrench, carrying case, big sticker, and CLC adapter that will fit most Asetek-based coolers.

Cinebench Performance

Up first is Maxon’s Cinebench R15. This multi-threaded benchmark is based on the engine used in the company’s professional Cinema4D product. The engine is somewhat older, but superbly efficient. It scales well with core and thread count as well as clock speed.

The result speaks for itself, as Threadripper 2990WX dusts the rest of the pack. The 32-core Threadripper 2990WX is 52 percent faster than the previous champ, the Core i9-7980XE.

ryzen threadripper 2990wx cinebench nt

Maxon’s Cinebench R15 scales incredibly well with core count, and the performance of the Threadripper 2990WX is nothing short of phenomenal. 

If only it were as easy as running Cinebench and declaring a winner. Reality is a lot more nuanced, though, so we also run Cinebench with it set to use just one thread. This favors CPUs with higher instructions per clock, and also ones that can hit higher clocks.

ryzen threadripper 2990wx cinebench 1t

Running Cinebench R15 using a single CPU thread, you can see that CPUs with higher clock speeds rule the day, and it’s pretty much all about Intel.

The winner is the 8th-gen Coffee Lake-S Core i7-8700K, thanks to its high Turbo Boost clock scores. Intel’s 18-core Core i9-7980XE somein second, with other Skylake-X and Kaby Lake CPUs following. We don’t see AMD show up until we see the Ryzen 7 2700X in 7th place. Granted, the scores are fairly close, but those higher Turbo Boost scores clearly put Intel in the driver’s seat.

Blender Performance

Our next test is the open-source Blender 3D modelling and rendering application, which has seen some use in indie movies. It’s so popular, even NASA now distributes its 3D model for Blender.

The CPU rendering option favors more cores, and the performance of the Threadripper 2990WX again is a crazy 37 percent faster than the 18-core Core i9 chip.

For this test, both the Core i9 and Threadripper 2990WX were on the latest 2.78C version, but we also included the performance CPUs runs using 2.78B for reference on older CPUs.

ryzen threadripper 2990wx blender

Blender Performance puts the 32-core Threadripper 2990WX about 37 percent faster than the 18-core Core i9.

Corona Photorealistic Render Performance

Up next is the Corona Photorealistic Renderer. Available for 3ds Max and Cinebench, the renderer  is seeing popularity in architectural design and visualization. As with other 3D rendering tasks, Corona loves CPU threads, ergo the 32-core Threadripper takes the top spot.

The performance of the 32-core 2990WX, is about double that of the 16-core Threadripper 1950X. The 18-core Core i9 does rather well, though, so it’s entirely the upcoming 28-core Intel part will take this one away from AMD when released later this year.

ryzen threadripper 2990wx corona renderer 1.3

Corona is a real-world photo-realistic renderer used in popular applications such as 3ds Max. Corona is not optimized for any specific CPU, but it does love more threads, helping the 2990WX take the prize.

V-Ray Renderer

V-Ray is an advanced 3D renderer that’s notched some good wins in its belt, as it was used for some effects scenes in Doctor Strange, Captain America: Civil War, and Deadpool. The benchmark can be used to measure both CPU and GPU performance, but we’re looking only at the former.

Like most renderers, V-Ray just loves CPU cores, making the new 32-core Threadripper 2990WX the clear winner. That’s just smoking. For comparison, PC maker Puget Systems measured a dual 14-core Xeon E5-2690 V4 system with a score of 31. So yes, that’s a $1,800 consumer CPU eating the lunch of $4,200 worth of Xeons.

ryzen threadripper 2990wx vray 1.08 cpuIDG
Chaos group’s V-Ray is a realistic ray tracer seeing some success in Hollywood. The standalone benchmark tends to favor thread count over clock speed. The 32-core Threadripper 2900WX takes the prize, but not by the margin we’d expect.

POV-Ray Performance

Our last rendering test is POV-Ray—a ray tracing program that dates back to the Commodore Amiga in the 1980s. It’s obviously been updated along the way, and like everything you’ve seen before, thread count should count the most. No surprise, we see the 32-core 2990WX eat everone’s lunch yet again. That 18-core Core i9 is in distant second place.

ryzen threadripper 2990wx pov ray nt performance

POV-Ray agrees with just about every renderer we’ve tried: More cores are better.

POV-Ray also allows single-threaded performance testing. The pattern we see is very similar to what we saw in Cinebench: When it comes to single-threaded performance, Intel still rules. Haters go ahead and hate, but that Core i8-8700K is the clear front-runner, with a batch of other Core chips right behind it. Just like with Cinebench, we don’t see a Ryzen until 7th place. While these scores are all relatively close, they do illustrate yet again that Intel’s chips can hit those higher clock speeds.

ryzen threadripper 2990wx pov ray 1t performance

POV-Ray’s single-threaded results seem awfully similar to Cinebench R15’s in single-threaded mode.

VeraCrypt Performance

Thus far, we’ve done the easy, feel good tests. The tests where a giant 32-core CPU make you feel warm and fuzzy inside. Not all code and not all tasks are the same. One of those is VeraCrypt, an open-source encryption application that picked up where the now-defunct TrueCrypt left off. This particular test, by the way, was a test AMD used as part of it test suite for the Ryzen CPU launch.

Performance generally appears to scale with core count, as the 16-core Ryzen Threadripper 1950X offers almost double the performance of an 8-core Ryzen 7 2700X. So why is the 32-core Ryzen Threadripper 2990WX so far off the mark?

Is it the die-to-die latency? Is it clock-speed-deficient? We honestly don’t know as performance here is a puzzler.  But again, this is benchmark AMD had recommended, so we’re at a loss to explain what’s going on right now.

ryzen threadripper 2990wx veracrypt aes

VeraCrypt seems to scale with core and thread count—but not for the new 32-core Threadripper 2990WX.

HandBrake Performance

That unease goes beyond VeraCrypt performance. Our encoding test uses an older version of the excellent, popular and free HandBrake encoder to convert a 30GB 1080p MKV file using the Android Tablet preset. It generally increases in performance the more CPU cores you throw at it. But oddly, the 32-core Ryzen Threadripper 2990WX is sandwiched between two Intel CPUs with only 10 cores apiece.

Again, it’s a bit of a mystery to us why this is happening. Is it an overall lower clock speed? Is it that die-to-die memory latency? Is it because HandBrake just can’t use all of those threads? We suspect the answer is all of those, but at the end of the day, the 32-core chip is underperforming here.

Premiere Pro CC 2018 Performance

Next on our video test is Premiere Pro CC 2018. We take a two-minute video shot on Sony Alpha in 4K and edited for our website. For the test, we take that project and export it using in Premiere using the Blu-ray preset with Maximum Render quality checked, which aids when resizing media. We’ve done this same workload on other CPUs, and it generally scales well with core count. Here are the results from the previous Core i9 test.

core i9 7980x premiere cc 4k cpu encode

Video snobs say CPU-based rendering is superior, and if you do that—you’ll want more cores.

As you can see, the 18-core Core i9 takes the gold, with the 16-core Ryzen Threadripper taking the silver. Because the previous tests were all conducted with Premiere Pro CC 2017, we didn’t feel it was fair to compare directly. But we did rerun the same test on the updated Core i9-7980XE and the 32-core Threadripper 2990WX. The result? Just as with HandBrake and VeraCrypt, the performance is lower than expected. Based on the performance of the Core i9 between the 2017 and 2018 version, it looks like the Threadripper 2990WX is just a bit faster than the 10-core Intel CPUs.

ryzen threadripper 2990wx premiere blu ray

Premiere Pro CC 2018 puts Intel’s Core i9 in front of the 32-core Threadripper CPU.

Now hold on, folks…

If you’re ready to slam your mouse on the table and jab the power button, you do need to read on. While we’re not sure whether it’s the die-to-die latency or simple lack of efficient multi-threading in software at fault, we did want to see if the 32-core Threadripper 2990WX can make magic on another important segment of hard work: multi-tasking. More specifically, what happens when your throw heavy loads at the 2990WX all at the same time. You know the Core i9 actually beats up the Threadripper 2990WX when you run Premiere Pro CC by itself, but what if you also run a Blender rendering test at the same time?

While this may sound unrealistic to you, it’s not that crazy for a indie video editor to encode in Premiere while also rendering out a visual effect in Blender. In fact, if you could to that without dragging everything to a halt, you would.

Once you do that, the performance of the 32-core Threadripper 2990WX comes alive. In the chart below, the shorter blue bar and the shorter red bar win. In this case, it’s the Threadripper 2990WX by a healthy margin.

ryzen threadripper 2990wx simultaneous blender premiere

Running Blender and Premiere Pro together gives the advantage to the 32-core Threadripper 2990WX.

Taking the same project, we exported it to Adobe Media Encoder and then simultaneously had it encode the same two-minute video to Blu-ray, HEVC, 1080P YouTube, and 4K Vimeo. Maximum Render quality was selected for the Blu-Ray, HEVC and 1080P Youtube video. The 32-core Threadripper again pulled ahead, but not by the margin we expected. We think that’s because of how long the videos took to encode. The Blu-ray, Vimeo and YouTube videos completed fairly quickly, with the HEVC encode taking the longest. Once they were finished, this turned back into mostly a single video encode, which can’t use all the cores of the Threadripper.

ryzen threadripper 2990wx media encoder ray

Encoding four different files simultaneously gave the advantage to the Threadripper CPU. 

Our last simultaneous test run involved running Blender and Cinebench at the same time. Because both are heavily multi-threaded, we expected the Threadripper to win big, but it didn’t. It won huge. Not only did it finish Blender 40 percent faster than the 18-core Core i9, its performance in Cinebench R15 is about the same as the Core i9—when the Intel CPU isn’t running two things at once. Let me say that again: The 32-core Threadripper is as fast as the top-gun Intel CPU when running two jobs, compared to just a single job on the Core i9.

ryzen threadripper 2990wx simultaneous cinebench r15 and blender bmw 2.78b

One big obstacle facing Threadripper WX is lack of applications that can truly take advantage of 64 compute threads. Where the Threadripper WX makes up for that is in multi-tasking of multi-threaded tasks. Here we ran Blender and Cinebench together. The Theadripper WX’s performance equals the 18-core Core i9’s when the Intel CPU is running Cinebench alone.

32-core Threadripper 2990WX gaming performance

We actually considered not running any gaming benchmarks on the 32-core Threadripper 2990WX, because if you bought this CPU to play Fortnite, you made a huge mistake. There are far better choices you can make with either Core i7 or Ryzen 7 over the Threadripper 2990WX.

But you don’t care, so we ran them anyway. First up is Tom Clancy’s Rainbow Six Siege at a fairly low resolution (for a GTX 1080, anyway) of 1920×1080, and with a visual quality setting of Ultra. We ran the Threadripper 2990WX in both its default Creator mode and in the 1/4, legacy mode which switches off three of the four dies under the lid.

The result? With all of its cores hot the 2990WX is a yawner in performance, as it takes a hefty hit over the Core i9-7980XE CPU. But in legacy mode, it’s actually pretty close. And at more than 200 fps, it’s close enough not to matter at all.

ryzen threadripper 2990wx rainbow six siege 19x10 ultra dx12

While you shouldn’t buy an $1,800 32-core CPU to play games, it does rather well in legacy mode.

We also ran Rise of the Tomb Raider at 19×10 in DirectX 12 mode and legacy mode. The latter makes the Ryzen Threadripper 2990WX pretty dead-even with the Core i9 CPU. The message is, you still shouldn’t buy this class of CPU if you’re going to play games 75 percent of the time. If, however, you push pixels 75 percent of the time, it’s a no-brainer to buy the new Threadripper.

ryzen threadripper 2990wx rise of the tomb raider 19x10 very high dx12

When all 32 cores are hot, the Threadripper 2990WX can’t keep up with the smaller 18-core Core i9 in Rise of the Tomb Raider at 1920×1080. Switching to legacy mode evens the competition.

About that memory latency

We do honestly wonder how much of the Threadripper 2990WX’s performance is impacted when a die has to take a detour through another die. We tried to coax it out using AIDA 64’s memory latency test. We first ran it on default with all four dies on, then with two dies on, and finally with only one die on. We hoped that AIDA 64 would access memory from one of the compute-only dies and we’d see memory latency increase, but all three results were essentially the same. At this point, we’ll say it’s inconclusive but we’ll keep looking. Maybe it’s just not worth worrying about.

ryzen threadripper 2990wx aida 64 memory latency

Because two of the four CPU dies on Ryzen Threadripper 2990WX lack memory controllers, we expect to see a slight increase in latency in theory, but we had no luck measuring it.

Thermals and boost performance

The top-end 32-core Threadripper 2990WX has a rated Thermal Design Power (TDP) of 250 watts, which is the maximum heat it will dissipate before hitting a wall. The 2990WX also has a total socket power of 250 watts, which it’ll hit a few milliseconds after doing any heavy multi-threaded loads. What this means is that you need to keep the CPU well fed with power and generally pretty cool.

Overall, we saw temperatures in the range of about 113 degrees Fahrenheit (45 degrees Celsius) during much of our testing on our Enermax Liqtech TR4 cooler, so it doesn’t require anything too exotic. We actually saw similar performance with the optional  $99 “stock” Wraith Ripper heat sink.

We also want to point out that it was very easy to push the Threadripper 2990WX out of its top boost speeds. Applications that normally would not bump CPUs out of boost speeds, such as Valve’s Steam, would do it to Threadripper. For example, with Steam installed, minimized and essentially doing nothing, the highest boost clock we saw was 3.4GHz to 3.5GHz. Once we exited Steam, we saw saw boost clocks of 3.9GHz to 4GHz. Fortunately, Steam was only installed and running for our gaming tests.

In fact, even using the Ryzen Master software to monitor Threadripper was enough to kick the CPU out of its top boost speeds limiting it to 3.5GHz rather than 4.1GHz. AMD officials were able to replicate that experience, but the company said its own results with Steam weren’t as bad as what we reported.

Ryzen Master

The new Ryzen Master utility gives you a glimpse of the cores in the crazy 32-core Ryzen Threadripper. Stars mean better CPU cores and the circles mean the next best CPU Core. We did run in to an issue where the CPU not boost when running the utility though.

32-core Threadripper Performance analysis

Take today’s multi-core CPUs, which all react differently to thermals and power loads, and throw them software that varies in thread parallelization efficiency and you basically get a big question mark on what to expect in performance.

To give you a general idea of what to expect, we take Cinebench R15 and run it repeatedly while increasing the threads from 1 to 64. While Cinebench isn’t going to tell you quite the same thing as a single-threaded game or Photoshop or an application that may effectively use 8 threads, it can tell us in general what you can expect out of a CPU under certain light or heavy loads.

For comparison, we took the 32-core Threadripper 2990WX and compared it to the 18-core Core i9-7980XE. The result: On the left side of the chart, the higher boost clocks and slightly better efficiency of the cores in the Core i9 give it a significant edge in performance. In fact, the Core i9 actually has a hefty performance advantage all the way up until you get to 28 threads, or the equivalent of a 14-core CPU.

18 ryzen threadripper 2990wx cinebench thread scaling

We used Cinebench R15 to measure sequential performance from 1 to 64 threads to gauge performance in multi-threaded applications. The 32-core Ryzen Threadripper 2990WX just crushes the 18-core Core i9-7980X at the top end, but it can’t hang with the Intel CPU on lighter loads.

Threadripper 2990WX wins the heavy loads

As expected, once you start to climb above 28 threads, the sheer number of cores and threads in the 2990WX outgun the Core i9. Because the chart above doesn’t give you a sense of just how much of an advantage each has over the other, we generated a chart based on the percent. As you can see, the Core i9 offers up to 20 percent more performance on those “light” loads of up to 28 threads. Move to the right though, and the Threadripper 2990WX can exceed the performance of the Core i9 by 50 to nearly 60 percent.

ryzen threadripper 2990wx cinebench thread scaling percentage

To give you an idea of how the previous Cinebench chart translates, we calculated the difference as a percentage. You can see the Core i9 has a big advantage on lighter loads, but the Threadripper crushes it on heavy loads.

What should you buy?

Frankly, if you tend to have workloads that aren’t going to scale and you don’t intend to multi-task heavily, the higher performance of the Core i9-7980XE might actually make more pragmatic sense. But if you do high-performance, heavily multi-threaded loads or multi-task heavily—you can’t look at the performance of the 32-core Threadripper and walk away.

The money shot

The literal money shot is just how much value you’re getting out of the Threadripper platform. It’s crazy to think that an $1,800 CPU can be a good value, but it truly is if you consider how much you’re paying per thread. AMD has amazingly introduced a new CPU with more cores than consumers have ever seen, and it has lowered the price at the same time. If that isn’t a deal, we don’t know what is.

21 bucks per thread

When you look at how much you’re paying per thread for a CPU, AMD’s the big winner. Perhaps worse, how does Intel sell a consumer 28-core CPU at a competing price, when the server version is almost $9,000?

Conclusion

AMD Ryzen Threadripper 2990WX

First let’s make it crystal-clear: The 32-core Threadripper 2990WX ($1799 on Newegg or Amazon) is not the CPU for most of us. Not by a long shot. For those of us who play games, edit some photos, browse the web and even do occasional video editing, an 8-core CPU is plenty, while a 16-core CPU is overkill. A 32-core CPU is double-overkill and honestly a waste. A Ryzen 7 or Core i7 is the more sound investment.

But for those of us who actually do push pixels around for a living, this new 32-core Ryzen Threadripper is Thor’s hammer falling right into your hands with a crackle of  lightning and thunder. For these heavy-hitters, it’s well worth the price.

(pcworld.com, http://bit.ly/2O76Plg)

Comments

comments

Share on FacebookShare on Google+Pin on PinterestShare on StumbleUponTweet about this on TwitterShare on LinkedIn