Reviews & Ratings
Intel’s 28-core Xeon W-3175X isn’t quite the Thread ripper killer, but damn, it gets close.
Intel’s crazy-ass 28-core Xeon W-3175X isn’t a CPU built for you, me, or most of us. Sure, Intel pitches it as a high-end workhorse: “Built for handling heavily threaded applications and tasks, the Intel Xeon W-3175X delivers uncompromising single and all-core world class performance for the most advanced professional creators and their demanding workloads.”
But make no mistake, the 28-core Xeon W-3175X is a chip made to do one thing:
make waves and push back at an increasingly assertive AMD. And it delivers—for a pretty penny. Fasten your seat belts, because you’re about to see some huge motherboards, a huger number of benchmarks, and the hugest desktop CPU price ever.
WHAT IS XEON W-3175X?
If you’ve read reviews ( go. pcworld.com/ci9r) of Intel’s previous 18-core Core i9-7980x or 18-core Core i9-9980x that replaced it, you already know something about the Xeon W-3175X. Like them, it’s essentially a Skylake-sp aimed at a highperformance crowd. It’s built on a 14nm process and has a stock TDP of 255 watts. To keep its cores fed, it features support for six channels of DDR4 memory in ECC or nonECC trim.
Perhaps its most important feature may be its unlocked status. As always, Intel doesn’t actually condone overclocking, the same way Lamborghini doesn’t tell you to break local speed limit laws.
What’s in the name? Why isn’t this a Core i9? Or, maybe a Core i11? Intel didn’t say why it chose to keep the Xeon name, but in the end it doesn’t matter much. With a list
price of $3,000, it’s how it performs that matters.
CRAZY-POWERFUL MOTHERBOARDS
One look at the first two motherboards designed for Xeon, and you know this isn’t for someone to run a SQL database to manage the inventory for a supermarket. Gigabyte’s AX1 features nothing less than a 28-phase power circuit ( go.pcworld.com/28ph) for the new Xeon W-3175X.
Hell, the Asus Xeon W-3175X motherboard weighs 10 pounds alone. Like the AX1 it, has enough auxiliary power connectors to light up a small city.
You don’t need two 1,600-watt power supplies to run a few virtual machines. No, the only time you’d need that amount of power is for over-the-top overclocking. Yes, it’s benchmark time.
HOW WE TESTED
For our tests, we basically made quad-core and eight-core CPUS leave the building. What? Just 18 cores? Get out! No, for this slug fest we wanted to find out the eternal question of what would happen if The Hulk fought Superman. Yes, nerds, you want to only know what happens when 28-core Xeon W-3175X faces off against 32-core Ryzen Threadripper 2990WX.
For our tests, we used the same configuration as our original Ryzen Threadripper 2990WX review here ( go. pcworld.com/290w). The only variance from the original review were updated motherboard drivers, updated GPU drivers and an updated BIOS and the latest version of Microsoft Windows 10 Pro RS5. While we used the same GPU, storage and OS for both, we did vary on memory and cooling.
The Threadripper 2990WX ran in quad-channel configuration with 32GB of dual-rank, DDR4/3200. The Xeon
W-3175X, ran 48GB of ECC DDR4/2667IN hexa-channel configuration.
For cooling, the Threadripper 2990WX used the same Enermax TR4 CLC cooler, while the Xeon W-3175X was cooled with an Asetek 690LX-PN.
The motherboard for the Xeon W-3175X was Gigabyte’s soon-to-be-released AX1, ,while the Threadripper used the same MSI X399 MEG from the original review.
One last note before we get too far: Since our last review, AMD released its “Dynamic Local Mode,” which helps keep threads on the most efficient cores in the CPU rather than let Windows randomly throw them at the wall. Normally DLM works really well, but there were moments when it slightly hurt performance. Because you have to enable DLM through Ryzen Master, we opted to run both DLM and the default configuration for our tests.
Finally, all of the tests were run at “stock clock” which oddly, means many things to many people these days. Basically, it’s the default setting on both boards.
3D MODELLING PERFORMANCE
Sometimes we apologize for using 3D modelling tests for performance testing because it’s not really what most people are doing. Well, if you buy a 28-core Xeon or a 32-core Threadripper 3D rendering is probably your jam.
The first result is from Maxon’s Cinebench R15. Although it uses a somewhat older version of the engine than the one in its Cinema4d product, it scales with core count and thread count exceptionally well.
More cores typically mean more performance
The clear winner here is the Xeon W-3175X, which crosses the finish line about 10 percent ahead of the Threadripper 2990WX. Some will see the proximity of the AMD CPU (as you can see, turning on AMD’S dynamic local mode slightly depresses performance) and think it’s a tie—but you should remember that we are talking about 64 threads vs. 56 threads. We had expected the Threadripper 2990WX to take the win here
instead of losing it.
The Threadripper’s loss to the Xeon isn’t just in Cinebench. In the Persistence of Vision ray tracer, the 32-core chip loses to the 28-core chip by about 8 percent. Again, DLM slightly depresses its performance.
Where it gets particularly ugly for Threadripper is in the Corona 1.3 Render. The Corona Renderer is an “unbiased photobought realistic renderer.”
Unbiased” refers to its rendering technique—not an allusion to any hardware leanings. The test embraces multi-core CPUS, so we expected Threadripper would do better than Xeon. Nope: t to be similar to the Cinebench performance but the Xeon absolutely hammers the Threadripper to the tune of 28 percent.
Next up is the Chaos Group’s V-ray renderer. Generally, the more cores you throw at it, the faster it gets. And yet the Xeon W-3175X again whups the Threadripper, despite the latter’s advantage in thread count. Possible reasons include each CPU’S particular micro-architecture design, the actual clock speeds each CPU runs at when loaded up, to the compiler used for each. We’ll try to circle back to find out more.
All that is academic, though, if the only thing that matters to you is less waiting for your V-ray or Corona-based render to finish.
Not everything is that ugly though. Using the latest beta version of Blender and Mike Pan’s BMW model, the Xeon again wins, but by just 5 percent. It’s close enough to be boring, unless, of course, you’re wondering why the extra 8 threads in Threadripper aren’t pulling their weight.
Our last multi-threaded rendering result is Blender’s new benchmark. It smashes together multiple popular benchmark models, runs them against Blender, and spits out a final result based on how long it took. There’s the option to run it against the GPU or the CPU, so we opted for the CPU test.
The result, again puts the 28-core Xeon on top by about 6 percent. Again, not a huge win, but it’s doing it against a 32-core CPU.
SINGLE-THREADED PERFORMANCE
If you bought a very expensive CPU to run single-threaded tests, you overdid it. Still, there’s value in seeing just how fast these particular chips are in much lighter loads.
First up is Cinebench R15 where, no surprise, we see Intel’s advantage in clock speeds show up: The Xeon comes in about 11 percent ahead of the Threadripper. Intel has long held the win for single-threaded performance.
If anything, we’re actually surprised the Threadripper is as close as it is, so in some ways, it’s a win for AMD too. We also see the Dynamic Local Mode actually help it slightly.
We an POV Ray using a single thread and, well, surprise, 10 percent.
COMPRESSION PERFORMANCE
Next up is performance of the two chips in compression, starting with WINRAR. We’ve been running it long enough to know that Ryzen just doesn’t like it. Besides testing AMD’S DLM mode, we also tested Bitsum’s Coreprio free utility ( go.pcworld.com/crpr), a DLM competitor that also fixes the mysterious problem in Windows that sees performance in some tests simply plummet. Most fingers point to problems with Microsoft Windows scheduler. Coreprio’s NUMA Dissassociator feature implements work discovered by Level1techs ( go. pcworld.com/lvl1).
First up is the decompression portion of 7-Zip which is mostly heavy in integer performance. Without the Coreprio utility, The Xeon has and advantage by about 10 percent. With the utility though, it’s mostly a tie. Intel fan, will of course point to 8-fewer threads means Xeon still wins right? But then, AMD fans will point to the dollar amount. So yeah.
Of more interest to us is the compression performance of 7-Zip. The developer has stated this portion of the test is particularly sensitive to memory bandwidth. As you know, the 32-core Threadripper has four channels of memory bandwidth spread among all of its cores.
The new 28-core Xeon has six channels of memory bandwidth. This theoretically gives each core about 27 percent more memory bandwidth at the speeds we tested and you’re likely seeing some of that here.
We’re saying likely because the memory bandwidth issue in Threadripper may not be as dire as it looked some months ago when we wrote this. With the Coreprio NUMA Dissassocator running we saw the huge gap of 58 percent for the Xeon versus just AMD’S DLM mode (red bar above) pull back to just 31 percent.
Sure, 31 percent is still, umm painful when you consider it has more cores and this is a multi-threaded test, but it’s better than 58 percent
(green bar above). Expect more on this in the future hopefully. The short answer is: Xeon wins big still.
CONTENT CREATION PERFORMANCE
Not everyone who might buy these CPUS does only 3D modelling. There’s a good chance they will also do content creation tests, which traditionally lean heavily on the CPU.
Our first test tasks the free and popular Handbrake utility with converting a 4K, 4GB file using the app’s H.265 profile. Handbrake is multi-threaded but it typically won’t use all of the threads of a 32-core, or even 28-core CPU.
The big winner here is the Xeon, which comes out on top by 17 percent when DLM is off. When DLM is on, the Xeon is actually 21 percent faster.
What’s up? Well, there’s a good chance that where Handbrake maxes out is just in that zone where the Xeon is at its peak performance on clock speeds. Sure, there’s that memory bandwidth thing, but we honestly have not seen memory bandwidth make that much of a difference in most encoding tasks.
Our next test uses Adobe Premiere Creative Cloud 2019 to export a short video shot on a 4K Sony Alpha camera using the app’s Blu-ray preset for export. Because the resolution changes, we also check off the Maximum Render quality option, which improves visual quality when resizing.
Finally, we do the encoding on the CPU, which some video nerds claim gives you the highest possible quality over GPU encoding. The winner: Xeon by about 15 percent.
Those who actually use Premiere CC are probably slamming their fists on the table saying, “no one uses the CPU purely for a video encoding anymore!” So yes, we did also encode it out using the Geforce GTX
1080. The win still goes to the Xeon, but it closes to about 11 percent.
Our next test uses the recently released benchmark test by Puget Systems. The company is famous for its systems and also for its in-depth testing of workstation-level hardware. The test uses Adobe After Effects Creative Cloud 2019 to run through several popular tasks done in After Effects. If you have After Effects, you can download the benchmark here ( go.pcworld.com/pget).
Running the AE test on the Xeon and Threadripper, it was basically a dead-even tie between the machines (although Threadripper performance dropped slightly with DLM on). In our book that’s a win for AMD.
Although Adobe Photoshop tends to be pretty easy for any modern computer to run, we did want to see which CPU had the advantage in Puget’s Photoshop test ( go. pcworld.com/phbn). Like the After Effects test, it’s free to download from Puget Systems and again—we highly recommend you head over to Puget System’s website if you are interested in this level of professional hardware. It’s simply a treasure trove.
Photoshop rarely loads up the cores of a CPU so the chip with the higher clocks was probably always going to win this and no surprise, the Xeon comes out ahead by about 8 percent.
If you drive Photoshop exclusively, a machine with as many cores as a Threadripper or Xeon is probably way too much.
MULTI-TASKING PERFORMANCE
In the real world, few applications can use all of the threads in a 64- or 56-thread CPU, so we’ve been trying to measure performance when you do multiple things at once. We say try because multi-tasking can be inherently unreliable for performance.
Still, we’ve done this particular test enough that we feel the results are reliably repeatable. We run Blender while also simultaneously running Cinebench. The result for Blender is almost a tie, but the big, big win for Threadripper is in Cinebench where it simply blows the Xeon away.
It’s almost, hmm, like the 32-core Threadripper has an additional 8-threads of compute power sitting around to tap on
that the 28-core Xeon doesn’t have. Win: Threadripper.
One thing about the above test: It’s probably not that realistic for someone to do a Cinema4d render at the same time as a Blender render. So we also uses Premiere CC to encode a 4K video to the Blu-ray preset while also rendering out a scene in Blender. This may sound crazy to you, but if you’re an indie movie maker, it’s an entirely realistic workload.
For the most part it’s a tie, but the Xeon ekes out a little more performance in the Premiere encode. Enough to call it a win? No, more like a draw.
GAMING PERFORMANCE
If you bought a $3,000 Xeon or a $1,800 Threadripper to play a game 90 percent of the time—you’re doing it wrong. Still, you do want to know how it performs so we present abbreviated set of results culled from other gaming tests we ran.
The result is no surprise: At resolutions and game settings that make the graphics card power the bottleneck, it’s nothing to write home about. The Xeon has about a 5- to 7-percent advantage in frame rates, but let’s just call it a tie.
The gap that has haunted Ryzen since day one remains,
though. In fact, when you take the GPU out of the equation by lowering the graphics quality, we see the very familiar 15 to 17 percent advantage for the Intel CPU. If you are buying a big CPU and do plan to game with the fastest GPUS in the world and do 3D rendering, modelling and other content creation, the advantage will generally go to Intel. If you’re just playing games sometimes, then it really doesn’t matter that much.
OVERCLOCKING PERFORMANCE
The vast majority of our testing is based on baseline speed which is what most people will stick with. It is, after all, pretty scary to think about heavily overclocking a $3,000 CPU.
Still, it would honestly be a crime not to at least do some basic overclocking with Xeon W-3175X. It was snap to push the Xeon W-3175X to a 4GHZ all-core boost just by goosing the multiplier. We pushed all cores up to 4.1GHZ and then also set turbo ratios for higher clocks on lighter loads.
The results of a casual hour netted significant performance dividends.
While an all-core of 4.1GHZ sounds pretty weak, it’s something many can aim for and not feel squeamish about. But all the indicators are there’s a ton more headroom in the chip. Speaking with vendors planning to sell Xeon systems at CES, they suggested an all-core overclock to 5GHZ wasn’t far from reality, with the only limits being power and thermals.
That probably tells us why both of the
launch motherboards for the Xeon W-3175X feature dual power input.
To give you an idea of where that falls, the current HWBOT record for a single 28-core Xeon 8180 Platinum is 5,010. We kicked out 5,859 without breaking a sweat.
IT’S A POWER HOG
And yes, the Xeon W-3175X is a power hog. On stock, we saw it regularly pushing loads of 550 watts at the socket (we’d estimate 60 watts to be just the fans in the system). The Threadripper 2990WX system was far more ‘green’ down at 350 watts under full load. Overclocking our Xeon W-3175X to a mild 4.1GHZ, we saw power climb up into the 700-watt range, too.
Mind you: That’s with a single power supply. It’s generally recommended that if you want to attempt to push all cores to 5GHZ and up, you should run a second matched PSU to keep the power-hungry Xeon happy. After all, if you bought a muscle car with a 440-cubic-inch engine, you wouldn’t complain about the gas mileage, would you?
THREAD SCALING
The last performance chart we want to leave you with shows how the Xeon W-3175X performs when you scale from 1 to 64 threads in Cinebench. Rather than the actual result we’ll give you the performance advantage for the Xeon W-3175X over the Threadripper 2990WX.
With the original Core i9-7980x, the 18-core CPU would outpace the AMD chip on lighter loads but eventually get hammered
as the 32-core Threadripper 2990WX’S advantage took over. Here, at stock speeds, the Xeon W-3175X has a huge performance advantage across the board, especially with applications that sit in that middle ground.
This doesn’t necessarily mean all applications will follow suit, but we will note that our Handbrake test, which put the Xeon W-3175X ahead by 17 percent to 20 percent, typically only used about 28 threads.
The upshot, as the vast majority of our tests have shown, is that the 28-core Xeon W-3175X is faster most of the time over the 32-core Threadripper 2990WX.
AND THEN THERE’S THE COST
Our normal price guidance in the lofty echelons of high-performance chips is not to care about price or value. When you’re shopping for a custom-built, custom-painted PC that costs at a minimum $12,000, caring about how much the CPU costs is like haggling to get the floor mats on a $300,000 car.
Still, we do have to look at the value of the $3,000 Xeon W-3175X per thread. We looked up the list prices of AMD and Intel’s big socket chips and computed how much they cost per thread. The worst are the 28-core Xeon Platinum chips that go into servers, which is not a surprise.
Among the CPUS that might conceivably be used in a (very fancy) desktop, the $3,000 Xeon W-3175X is actually in line with most Intel CPUS. The best value still belongs to AMD, which is basically charging you half of what Intel charges per thread for its CPUS.
BOTTOM LINE
Here’s a funny story: When we originally received the Xeon W-3175X for testing at what we thought was a price of $4,000 we actually thought Intel had actually created a CPU at a price designed to actually make AMD’S Threadripper 2990WX look better. After all, with a 32-core Threadripper 2990WX going for $1,800, no amount of performance was going to really make it a product worth considering for anyone who doesn’t fly around on a private jet.
With a price of $3,000 and an actual demonstrable performance advantage in a lot of areas though, it’s actually a contender. It’s not a knock out by any means but for those who do want it all and don’t mind paying for it, it’s going to be really hard to find a faster CPU out today than the Xeon W-3175X.