AMD Threadripper 3970X Review: 32 cores of unbeatable power
It blows away Intel's best.
By Gordon Mah Ung
PCWorldNov 25, 2019 7:55 am PST
Image: Gordon Mah Ung
AMD is on a rampage. Its 32-core Ryzen Threadripper 3970X is the monster rising out of the sea to level Intel’s prosumer chips. This is just weeks after the company’s 16-core Ryzen 9 3950X cleaned the clock of Intel’s top consumer chips.
Note that “prosumer” means Threadripper is designed primarily for graphics-intensive design and content creation tasks, rather than gaming. (You could play games with this chip, but it’s massive overkill.) AMD has an opportunity in this space, because the Xeon CPUs Intel makes for the same market are far more expensive. And this is just the beginning of AMD’s attack on the high end: On Monday morning, AMD confirmed that it will make a 64-core chip as well.
If you can “just get by” in the meantime with a 32-core Ryzen Threadripper 3970X that pretty much crushes all, put your seat belt on and keep reading for tech specs, benchmarks, and more.
What is Ryzen Threadripper 3970X?
Threadripper 3970X uses AMD’s phenomenal 7nm chips, but its most significant change is its topology. The original 32-core Threadripper 2990WX we reviewed was a little bit of a kludge. The chip was built with four 8-core chiplets, but only two of the chiplets had memory controllers and PCIe access. That meant the two compute-only dies had to go through the other dies to get to RAM or storage. You could think of the design as adding two more bedrooms to your two-bedroom home, but forcing those in the new bedrooms to go through the old bedrooms when leaving. You could imagine how off-kilter this could sometimes make the original 32-core Threadripper 2990WX.
With the new Threadripper 3970X, AMD has gone to a single 12nm IO Die that contains all of the access to PCIe 4.0 lanes and the memory controller. The IO Die connects to each compute die using a high-speed fabric, giving each of the CCD chips equal access.
Same Socket, New Connections
All this probably sounds great—unless you’ve bought into the previous X399-based, Threadripper series. That’s because even though the actual physical socket is the same with the new third-gen Threadripper, they are electrically different.
That means the new chip is incompatible with existing motherboards. Instead, AMD is introducing the new TRX40 platform, which works with the 3000-series and future Threadripper CPUs. Those X399 and 2nd-gen Threadripper chips? They’re essentially being pushed overboard.
In addition to the new CPU, the TRX40 increases PCIe lanes from 64 to 72. That’s an upgrade and then some, because the 3000-series Threadrippers also move up from PCIe 3.0 to PCIe 4.0. That basically doubles the theoretical bandwidth of each lane, meaning the new TRX40 platform offers the equivalent of 144 lanes of PCIe 3.0 bandwidth.
How we tested
For our test system, we loaded Windows 10 1903 on a Corsair MP600 PCIe 4.0 SSD and installed a Founder’s Edition GeForce RTX 2080 Ti GPU and 64GB of Corsair DDR/3600 CL15 RAM. We cooled the 32-core Threadripper with a Corsair H110i RGB, with its fans manually set to 100 percent and iCue disabled.
We’ll kick off our performance tests with the area of computing that always wrings the most out of a multi-core CPU: 3D modelling and rendering.
First up is Maxon’s new Cinbench R20 benchmark. Based off of the same engine in the company’s Cinema4D renderer, Cinebench R20 has been updated to support AVX2 and AVX512 and scales pretty perfectly with core count.
The result below pretty much sets the stage for everything you’ll see from Threadripper in 3D rendering performance across the board.
Despite being a beefy 32-core chip, the Ryzen Threadripper 3970X offers up surprisingly good single-threaded performance too. You can see that represented below, where it’s slightly ahead of the 12-core Ryzen 3900X and just about dead even with the 5GHz Core i9 chips.
Due to time constraints we didn’t refresh our 28-core Xeon W-3175X nor our 32-core Threadripper 2990WX for all of our tests, but the previous Cinebench scores we obtained are still valid. As you can see, even the $3,000 28-core Xeon W-3175X can’t hang at stock speeds against the new 32-core Threadripper. To be fair, running that Xeon W-3175X at stock speeds is almost criminal, but clearly the newest Threadripper rules the day. And yes, the Threadripper 3970X is outperforming the previous 32-core Threadripper 2990WX to the tune of 43 percent!
Firing up the “proudly CPU based” Corona renderer, we can see the 3970X outmuscle all other CPUs again. We should note there is some good news here for the new 18-core Intel Core i9-10980XE: It actually outpaces the 16-core Ryzen 9 3950X in a multi-threaded test. The Corona Render is an unbiased renderer, which means no shortcuts are taken to render a scene.
Next up is the old POV Ray test. The Persistence of Vision ray tracer hearkens back to the Amiga days of yore, but an army of volunteers have kept this free and old ray tracer updated for modern times. As you can see, more cores means more performance, and none of the other CPUs can touch the 32-core Threadripper 3970X.
Our next test is V-Ray Next, which is a physically-based renderer that was used on such movies as Avengers: Endgame. V-Ray Next again confirms that the 32-core Threadripper is a monster CPU.
Our last 3D rendering test shows just how shockingly fast the Threadripper part is. V-Ray Next features CPU-focused workloads as well as a GPU-focused workloads. V-Ray Next also lets you run the GPU-focused workload in a hybrid mode, where you can combine the CPU and GPU or measure the CPU’s performance at the GPU’s job.
We ran this hybrid mode across all of the CPUs below, and also on the Founders Edition GeForce RTX 2080 Ti card. Although the Threadripper 3970X doesn’t beat the mighty GeForce RTX 2080 Ti, it comes surprisingly close.
What you saw above in 3D rendering is where the 32-core Threadripper 3970X was expected to do the best. High core counts don’t translate into the best performance in video encoding and editing, though. Memory bandwidth, special instructions, and other CPU features factor in just as much with video tasks, we’ve found.
For our first test, we use a nightly build of the popular and efficient HandBrake encoder to convert a 4K video file using the HEVC codec. Our first run takes that 4K file down to 1080p. The good news for the Threadripper 3970X is it’s the overall winner. The bad news: The 16-core Ryzen 9 3950X isn’t that far behind.
One thing we know with video transcoding and encoding is the codec, resolution and other video settings make a difference. In the chart below, we take the same workload but rather than select 1080p at 30 fps, we select 2160p at 60 fps. With this workload, we see the Threadripper 3960X open up a 33 percent lead over the Ryzen 9 3950X and nearly a 45 percent gap over the 18-core Core i9-10980XE.
Our next video test is Cinegy’s Cinescore 10.4 benchmark. This is a test designed to let broadcast professionals gauge commercial off-the-shelf component performance. It’s designed to fit in memory so as not to be storage-bound, and it tests CUDA performance and CPU performance using various broadcast codecs and resolutions. The overall score is based on each machine’s performance across all of the tests.
As you can see, the 32-core Threadripper opens up a stunning lead. The score is especially a surprise given that the gaps from 8 cores to 16 and 18 cores are far closer.
Our last set of video tests is a little more consumer-focused. For that, we take Adobe’s popular Premiere Pro CC 13.5 and export a video project shot on a 4K Sony Alpha mirrorless camera, using the Blu-ray preset with the maximum render option checked. The winner is the Threadripper, Although it doesn’t scale as perfectly as it does in 3D rendering, we do see a nice progression in performance as we go from 12 cores to 16 cores and 32 cores.
We did want to see whether more difficult tasks would result in better performance for the higher-core-count CPUs, so we take the same workload and output it using the HEVC codec instead of Blu-ray’s H.264 in Premiere. This is done with the Maximum Render option and using the High profile. The Threadripper again wins, but the margin is actually slightly less than it is at 1080p resolution and H.264.
Most people who pay for an expensive GPU tend to use that GPU for encoding, so we ran the same test, but rather than force Premiere to use the CPU for encoding, we used the GeForce RTX 2080 Ti. Although the Threadripper technically leads here, we feel the results aren’t conclusive enough to say whether it matters that much.
Keep reading for more benchmarks, including gaming.
Our next battery of tests looks at the CPU’s file compression and decompression performance. We’re particularly interested in the the free and popular 7-Zip Benchmark because the older 32-core Threadripper 2990WX had issues with it. The problems were severe enough that the 32-core scored lower than an 18-core Core i9-7980XE in the compression test, which pushes memory latency, cache performance, and out-of-order instruction performance. The good news: The new topology, as well as updates to Windows 10, seem to have fixed much of the limitations Threadripper suffered in the compression test.
Decompression performance, which stresses a CPU’s integer, branch prediction and instruction latency, is also as it should be, with the 32-core Threadripper 3970X simply singing.
Our last test uses RAR Lab’s WinRAR to measure compression and decompression performance. WinRAR has never been particularly friendly to Ryzen in the past, but these results are perhaps even more puzzling. Although WinRAR is a “multi-threaded” test, it seems to favor the lower-core-count chips with the 8-core Core i9-9900K and 9900Ks dead even with the 12-core Ryzen 9 3900X. The 18-core Core i9-10980XE is up next with, oddly, the 16-core Ryzen 9 3950X and Threadripper 3970X pulling even. The short answer is probably: Just use 7-Zip.
We have a question: Why would you buy a 32-core content creation CPU to play games? We certainly wouldn’t. But it’s a relevant question, as many people work by day and play by night on the same machine.
With that, we’ll get into a short summary of gaming performance on the Threadripper 3970X. Even though we’re showing only four results, we ran many other tests and at different resolutions. The only resolution we’re going to show here is 1920×1080, because higher resolutions you are more GPU-bound, and the differences get even smaller.
First up is the fairly new Shadow of the Tomb Raider at 1920×1080, set to Highest Quality and with ray tracing turned off. The results are surprisingly good for the Threadripper 3970X which outperforms the two Ryzen 9 CPUs and the 18-core Core i9-10980XE CPU.
Moving onto the mostly GPU-bound Gears of War 5, we see results more in line with what we expected: The higher clocked Core i9-9900K and 9900KS lead the way, with the Ryzen 9 chips nipping at their heels. Both of the larger-socket, high-core count Intel chips come in last. But differences aside, all of the CPUs here are producing decent performance, and you can’t make the wrong choice.
In Far Cry 5, however, the results were not good. This game has long run better on Intel CPUs, and the higher clock, the better the performance. We expected the 18-core Core i9-10980XE to do better, but its lower clock speeds basically pull dead even with the 12-core and 16-core Ryzen 9 chips.
The biggest loser is the Threadripper 3970X. At 76 fps, it’s basically half as fast as the 5GHz Core i9-9900KS. There’s really no reason for it to be so much slower when you consider that the Ryzen 9’s at least are in the race. We should also mention that when you start Far Cry 5, you get a “Ryzen and Radeon” screen so, yeah.
Clearly, Far Cry 5 runs pretty badly on the Threadripper 3970X. We considered using Ryzen Master to run the game in Game Mode (which switches off some of the compute dies), but it wasn’t available at the time we ran we our tests. We’ll consider revisiting this in the future.
Our last game is Tom Clancy’s Rainbow Six Siege, which has fairly low graphics requirements. This is probably pretty typical of most games, which see the Ryzen CPUs slower by only about 10 percent. Consider that we’re using a GeForce RTX 2080 Ti, and at a fairly low resolution. These results reinforce our belief that most of the time, the gap between Ryzen and Core isn’t enough for us to give up the Ryzen CPUs’ greater core count.
Before we close this off, we wanted to look at the performance of the 32-core Threadripper 3970X using Cinebench R15, ramping from using a single-thread to the maximum threads on the chip. This helps us visualize the strength of the CPU, as most applications simply can’t utilize 64 threads very efficiently.
First up, we see the percent difference between the 32-core Threadripper 3970X and its sibling, the 16-core Ryzen 9 3950X. As you can see, the smaller Ryzen 9 3950X actually has a small advantage on lightly threaded workloads—its higher boost clocks give it up to a 5- to 6-percent advantage over the big Threadripper. As you move further to the right though, just get out of the way
Besides the 16-core Ryzen 9, we also compared Intel’s new 18-core Core i9-10980XE against the 32-core Threadripper 3970X. Yes, some will say that’s unfair, because it’s a $2,000 CPU vs. a $1,000 CPU, but it’s also clear to to us why Intel slashed prices of the 18-core chip. Remember: Last year’s 18-core Core i9-9980XE sold for $1,999, and in a post-Threadripper 3970X world would make no sense at all.
So how does the Core i9-10980XE fare? Well, not better than the Ryzen 9 3950X, anyway. The mighty Threadripper simply outmuscles it.
Intel does have one more card up its sleeve, which we fortunately have numbers for: The 28-core Xeon W-3175X. The CPU was made as a showpiece, intended to put AMD’s 32-core Threadripper 2990WX in its place. For the most part it achieved that with its higher clock speeds. But the 32-core Threadripper 3970X is a different chip.
Although the comparison below is based on performance obtained with a slightly older version of Windows, they’re still valid. While the 28-core Xeon W-3175X actually fares better than the 18-core Core i9 on the right side of the chart where the core counts are closer, it actually performs worse than the 18-core Core i9, too. Losing by 37 percent is better than losing by 100 percent, but it’s still a loss.
There’s an argument that the Xeon W-3175X isn’t meant to be run at stock speeds at all. It’s a flashy chip that should be liquid-cooled or liquid-nitrogen-cooled for overclocking runs only. But there’s also an argument that at $3,000 for the Xeon and $1,800 for the motherboard to run it—it’s not even meant as a practical platform.
The 32-core Threadripper 3970X runs unopposed in our eyes. Intel’s new 18-core Core i9-10980XE basically has its hands full dealing with the 16-core Ryzen 9 3950X. And the 28-core Xeon W-3175X at $5,000 isn’t practical.
Some will complain that $2,000 is too much for this chip. But for anyone who actually needs multi-threaded performance that easily outperforms all previous high-core-count CPUs, it’s almost a bargain.
eBay CouponeBay Best Price Guarantee: Save 10% on any item