AMD Threadripper 3970X, 3960X, and Intel Core i9-10980XE CPUs Tested: Intel Cuts Prices, AMD Redefines What's Possible -…
I have been a CPU reviewer for just over 18 years. In that time, I’ve been fortunate enough to observe a handful of truly game-changing launches. Today is one of them. The Threadripper 3960X and 3970X will never be mistaken for inexpensive processors, but AMD just redefined what’s possible in a high-end PC workstation.
There’s another reason this launch is unique — it’s the first time AMD and Intel have launched CPUs on exactly the same day. We have combined our reviews of Cascade Lake and the Threadripper 3960X and 3970X into a single document to preserve my sanity.
Intel isn’t standing still. The Cascade Lake 10980XE launches today, with a dramatically reduced price compared with the Intel HEDT platforms of old, and a new value proposition for itself at the $1,000 price point. Intel has finally trimmed its prices and it’s targeting the weakest point in AMD’s overall lineup — the gap between the 16-core Ryzen 9 3950X and the 24-core Ryzen Threadripper 3960X.
The Chips We’re Covering
Since this is a three-way CPU review I should start by introducing the contenders. There are three of them:
These two CPUs are built on AMD’s Zen 2 CPU architecture and use the same chiplet design as the Ryzen 3000 7nm APUs AMD launched earlier this year. They require a brand-new motherboard platform, TRX40, and are not backward-compatible with AMD’s earlier Threadripper platform. Similarly, TRX40 doesn’t support Threadripper 1 or 2.
Meanwhile, in Intel’s corner, we have the Core i9-10980XE:
The 10980XE and its fellow compatriots are built on Intel’s 14nm++ process. These chips have a higher TDP than in the past (165W) and are supposed to offer higher overall performance as a result. The major story with Cascade Lake is the significant price cut — roughly 50 percent compared with previous-generation CPUs. The 10980XE weighs in at about half the price of the Core i9-9980XE — which means it’s also much cheaper than either of the two Threadripper’s we will be reviewing. The primary competition for the Core i9-9980XE is the Ryzen 9 3950X, at $750.
The Core i9-9980XE and 10980XE are the most powerful Intel CPUs we have on-site; Intel did not offer a Xeon system for comparison against Threadripper. While we prefer to compare on price, we’ve always compared top-end products against each other, even when there were wide gaps in cost.
What Threadripper Makes Possible
The 3960X and 3970X are extremely expensive processors by desktop standards, weighing in at $1,400 and $2,000 apiece. I won’t blame you for raising an eyebrow when I say the following: The 3960X and 3970X are going to make ultra-high-end workstations much less expensive than they’ve ever been. Not many people play in this space — frankly, not many have reason to — but the ones who do are going to be very happy with what these CPU’s bring to the table.
Up until today, AMD’s Threadripper has focused on disrupting Intel’s workstation business by offering excellent overall performance-per-dollar. A 16-core Threadripper 2950X was always faster than the 10-core Core i9-7900X it was priced against, but it fell well behind CPUs like the Core i9-9980XE. The 9980XE wasn’t priced competitively — at nearly $2,000, it was over double the 2950X — but it was the unambiguously faster processor.
As of today, AMD and Intel have swapped these positions. Intel’s price cuts have positioned Cascade Lake as a potential price/performance contender, while AMD’s Threadripper 3960X and 3970X are redefining the performance envelope at the new top of the workstation market. Intel has no HEDT part it can position against these CPUs. While it does manufacture a 28-core chip that would conceivably give the 3970X a run for its money (and believe me, we’d love to see the comparison), the Xeon W-3275M (2.5GHz – 4.6GHz) is a $7,453 CPU that would be facing off against a $2,000 CPU, and it’s by no means certain that AMD would lose that battle.
AMD’s New Threadripper Platform: TRX40
Before we dive into performance, let’s talk a bit about the chips themselves. The diagram below shows how each Threadripper die connects to the central I/O block. Each die can read 51.2GB/s of data per second and write 25.6GB/s at a 1600MHz fabric clock (equivalent to DDR4-3200, as RAM speed scales with fabric clock).
According to AMD, it reduced IFOP (Infinity Fabric On Package) power consumption by 27 percent with Threadripper 3, allowing it to divert that power to the CPU cores. The move to the new TRX40 platform and the lack of backward compatibility for third-gen Threadripper has left a lot of first and second-gen Threadripper customers unhappy, but AMD has stated that the shift was necessary in order to support the increased power draw and long-term scaling needs of the company’s future CPUs. AMD has explicitly not made any additional long-term commitments to TRX40 like it had made to AM4, but the company’s overall history of chipset support is good.
One major upgrade to the TRX40 platform is PCIe 4.0 support. There’s a shiny, new PCIe 4.0 x8 lane connecting chipset and CPU, offering 4x the bandwidth of the old PCIe 3.0 x4 link. The new TRX40 supports up to 72 usable PCIe 4.0 lanes and up to 12x USB 3.0 ports. Technically the TRX40 has 88 PCIe 4.0 lanes, but 16 of them are reserved for the system’s use. Threadripper also now supports 256GB of RAM, with ECC support available, if motherboard OEMs choose to provide it.
And in the Blue Corner…
Since this is a combined review, I want to say a few words about what’s brewing on Intel’s side of the ring. Obviously there’s not as much raw excitement in the air, but Intel’s dramatic price cuts put the Core i9-10980XE in a very interesting position. It has an opportunity to carve out a slice of the market for itself as the overall stronger product, if it can take out the Ryzen 9 3950X.
Intel’s price cuts will be welcomed by fans of the company and Intel’s share of the workstation and ultra-high-end desktop space dwarfs AMD’s. Threadripper 1 and 2 were both excellent CPUs, in my opinion, but they didn’t exactly seize tons of market share. These price cuts represent a huge price/performance improvement for customers who are dedicated Intel users, and the company has a strong foothold at the $1,000 price point as we head into the performance section of our review.
Blender Render Fender Bender
Not all memory controllers are created equal, and I’ve historically had more trouble with Intel controllers refusing to run at high RAM clock than AMD CPUs. I’ve found that the Blender standalone benchmark (1.0Beta2) works fairly well as a RAM test and power consumption test both and have started running it for those twin purposes. It’s typically the first test I run on a system.
With the 10980XE, I installed Windows 10 and its various drivers, patched up, activated XMP, started the standalone benchmark… and the test crashed at the end of the 4th render scene (there are six, total). Confused, I repeated the test. Same result. Thus began an eight-hour test cycle of activating and de-activating XMP, manually adjusting RAM timing, and testing multiple Windows 10 installations. Part of the reason this took so long is because Blender isn’t a short benchmark. It took ~22 minutes for the application to reach the point where it would crash.
It’s reasonable to ask why I didn’t change course and run other tests. The answer is that I was focused on trying to establish a best-in-class configuration for the 10980XE and thought the RAM errors were the result of the controller being pushed above Intel’s spec. I have an Intel 7980XE that will not run at DDR4-3200 and Intel doesn’t formally certify anything above its maximum official clock. Since I knew I didn’t have time to go back and generate a second set of numbers if the first proved to have been built on a bad RAM configuration, I kept testing. And testing. And testing. Finally, I threw a 7900X in the motherboard to see what would happen — and the test completed perfectly.
Marco Chiapetta at Hot Hardware was kind enough to run the same test on his own motherboard, confirming that the 10980XE cannot complete the Blender Render standalone benchmark (1.0Beta2). This problem appears limited to the benchmark, other reviewers had no problem with the standalone version of the application and I didn’t observe any systemic crashes throughout my testing.
To be added later due to time pressure. Rest assured that i’s were dotted and t’s were crossed. Much like my eyes right now. It’s after 5 AM and I’m going to have to ask your forbearance, Dear Reader. Ditto on the fact that RAM speed isn’t listed for every CPU. Remind me to tell you about my adventures with what I’m calling my Blender Render Fender Bender when I wake up.
AMD wins every position in Cinebench R15, from single-thread (3950X) to multi-thread at 16-core (3950X), to multi-thread at 32-core (the 3970X). The Core i9-10980XE shows some signs of life in single-threaded applications here, with a 6.5 percent uplift over the 9980XE. Multi-threaded performance is essentially flat.
Cinebench R20 repeats the performance pattern of R15. The 3950X again leads single-threaded performance, though the 10980XE picks up about 5 percent compared with the 9980XE. The Threadripper 3970X simply destroys the benchmark, beating the Core i9-10980XE by 1.94x. Scaling between the Ryzen 9 3900X and the Threadripper 3960X is 1.91x — nearly perfect, at least in this test.
Dolphin 5.0 EMU
Dolphin is a Gamecube emulator, its benchmark tends to be sensitive to single-threaded performance and low cache latency. The Core i9-10980XE makes a surprise move here, winning Dolphin over any other CPU.
7zip 19.00 (64-bit)
7zip’s compression test is another win for the 10980XE over the 3950X, with the newer Intel CPU beating the 9980XE by an impressive 1.25x. The 3950X still beats both Intel chips on decompression performance, while Threadripper stays true to its name and rips the hell out of the benchmark, nearly hitting 4,000MB/s. Where the 2990WX falls down, the 3970X glides gracefully to a stellar finish.
Handbrake tests encoding in both H.264 and H.265. We encode a 4K video clip using the fast 1080p preset, with the video frame rate set to “Same as Source.”
You can definitely see diminishing returns in both encoders as core counts rise, though H.265 benefits more from higher CPU cores than H.264. The Ryzen 9 3950X continues to nimbly dance ahead of the 10980XE, though the latter does improve on its older sibling’s performance. The two Threadrippers aren’t as far ahead of the pack as they’ve been in other tests, but they’re still the fastest chips around.
Our Qt compile test is run using Microsoft Visual Studio 2019 on all systems. Antivirus scanners are deactivated for the compile test.
The Core i9-9980XE is decisively faster than the Ryzen 9 3950X, but the 10980XE is inextricably slower. The difference nearly allowed the Ryzen 9 3900X to catch the 10980XE — not a great look. The 2990WX turns in a weak performance relative to core count, while Threadripper streaks ahead, coming in below the 7-minute mark.
Corona Render is a photorealistic renderer that incorporates Intel’s Embree ray tracing kernels. It can be tested via a standalone benchmark.
The Threadripper 2950X manages to edge ahead of the 3900X here, with the 3950X well ahead. The Core i9-9980XE and 10980XE tie things up and would sweep the benchmark… if the Threadripper 3970X and 3960X hadn’t just smashed records.
V-Ray Standalone Benchmark:
The standalone benchmark for Chaos Group’s V-ray shows that the 2990WX isn’t always the kid at the back of the room eating paste, but it can’t beat the Core i9-9980XE or 10980XE. Neither of those CPUs, meanwhile, were capable of matching either of the Threadripper CPUs.
y-cruncher is a highly optimized application for calculating pi. We used optimized binaries for our Ryzen and Intel CPUs, though the Ryzen binary is actually optimized for Ryzen 1, not the Zen 2 architecture. We specifically used an AVX-512 binary to give the Cascade Lake platform an opportunity to put its best foot forward.
The 10980XE picks up about 5 percent over the 9980XE, and both beat the 2990WX. The 3960X and 3970X crush all other competitors, even without an optimized binary. AVX-512 gives Intel a boost, but it’s not enough to overcome the core count hit in and of itself.
The following benchmarks are tests suggested by either Intel or AMD. AMD suggested Keyshot in its Threadripper material, while Intel suggested Matlab and Sony Catalyst. Matlab is a particular pain to authorize — I had to create a new fake email address for every single trial account I wanted to test. This is why there are only Matlab results for three systems — the Core i9-10980XE and the two Threadrippers.
Keyshot is a 3D rendering application suggested by AMD. It runs extremely well on Threadripper when tested using the built-in Camera_benchmark scene.
The uplift from the additional cores is pretty reasonable — 1.2x better performance from a 1.33x core count improvement.
I was unaware when I initially ran this benchmark that it made use of a library that refuses to run AVX code properly on AMD CPUs. I will have more to say about this in an upcoming story. Consider the data below to be incomplete.
Matlab’s results are rather interesting. They show a mixture of wins and losses for AMD and Intel. If you add up the final times, they’re 38.83 seconds for the 10980XE, 40.71 seconds for the 3960X, and 31.84 seconds for the 3970X. Again, that’s roughly a 1.2x improvement for a 1.3x core count jump.
There are some cases where the 10980XE is significantly ahead, and a few where AMD makes up ground. As an Intel-provided test, I would expect this to favor Chipzilla. The point Intel is making here, however, is a valid one. There are specific applications/calculations or tests where Core CPUs show very strong performance relative to Ryzen.
Sony Catalyst 2019
Intel’s Sony Catalyst Edit test measures how long it takes to render, export, and encode three audio clips into a single 4K video.
This test shows the 10980XE turning in impressive performance, but it’s still outperformed by both of the Threadripper competitors.
For gaming, I’ve left the Core i7-9700K to represent lower-end CPUs in this rarefied air. All of the CPUs we’re going to discuss here can game. None of them are gaming CPUs, specifically. You’ll get just as good or better performance from a chip like the 9700K than you will from an ultra-high-end CPU.
All of these games are perfectly playable on all of these chips, though Warhammer II on the 2990WX may take some finagling.
Ashes of the Singularity: Escalation
First off, check the title, not the graph. I’ll fix it when I’m alive again. Ashes of the Singularity: Escalation is always a bit of an odd duck, and today is no exception. The 3970X and 2990WX(!) are nipping at the 9700K, while the 10980XE is sharply below the 9980XE for unknown reasons. The 3960X also fell off sharply at 4K for unknown reasons.
Deus Ex: Mankind Divided
Deus Ex: Mankind Divided cares not for your CPU choice. Deus Ex: Mankind Divided has an MSAA implementation that hits the GPU like a sackful of bricks.
The 10980XE picks up some frames on the 9980XE in Hitman, but the Ryzen 9 3970X is still a hair faster. The 2990WX shows a significant weakness in Hitman, but the 3970X has no such problem.
Shadow of War
The 10980XE again picks up a bit of performance over the 9980XE, but all contenders are closely grouped here.
Warhammer II is not a friend to AMD CPUs, but it isn’t overly fond of high core counts, either. The 9700K wins the test, but the 3700X, 3900X, and 3950X account well for themselves. The 10980XE picks up a few frames above the 9980XE but remains behind the various AMD cores.
Shadow of the Tomb Raider
Everyone is closely clustered in this test, at all three resolutions. There’s no dominant winner, though the 2950X can be proud of its performance against the 10980XE and 9980XE.
With the launch of the Threadripper 3960X and 3970X, AMD has taken a decisive leadership position in the high-end desktop market. I’m not going to say that Intel doesn’t build chips that could take on the Threadripper 3970X — it does — but it doesn’t sell them in this space or for these kinds of prices.
Intel’s dramatic price cuts make Cascade Lake objectively a far better deal than any previous Intel Core X CPU. If you bought into Skylake X with, say, a 7900X for $1,000, a 10980XE for $1,000 may be the best upgrade path for your machine (assuming, of course, that you can use the extra cores). At the same time, however, it’s hard to ignore the fact that AMD’s 16-core is now beating Intel’s 18-core in a number of tests. For now, Cascade Lake’s 50 percent price cut will make plenty of people happy, but the Ryzen 9 3950X is nipping at the heels of Intel’s HEDT product line.
If the 3950X represents the dog nipping at the back of Intel’s heels, the 3960X and 3970X are the 24-ton and 32-ton elephants sitting on its head. [That’s quite a scene happening here. -Ed] These chips offer amazing, unbeatable performance that Intel cannot match in the short term. It’s not clear when Chipzilla will field a product that can rival what AMD has dropped in-market today.
Last year, the 2990WX had too many flaws and shortcomings to be considered a serious product. The 3970X fixes them. Its gaming performance is excellent. Its application performance is phenomenal. Intel has been driven back to a barricade built on 1080p gaming and lightly threaded workloads that are well-tuned for its CPU cores. Given Intel’s long-standing dominance of the market, that barricade is a good deal stronger than it may look to AMD’s most adoring fans — but the company has been driven backward, make no mistake.
Cascade Lake will not be driven from the field. While its performance was a bit odd in places, the Core i9-10980XE delivers some nice additional single-threaded performance. But the excitement and energy in the PC desktop market right now is being driven by AMD. The Threadripper 3970X and 3960X represent one of the largest leaps forward for high-end desktop computing that I’ve ever seen any company deliver.