To some of you, it may feel like it's only been a few months since AMD took the HEDT market by storm. That's because it was only a few months ago when AMD took the performance crown with the launch of their Threadripper 3000 series processors. Apparently having the fastest HEDT processor wasn't good enough for AMD, and the lack of competition in this market meant they had to become their own competitor. Enter the Ryzen Threadripper 3990X, AMD's proof that you can solve all of your problems by throwing more cores at it!

In this review, we will be going over the various features of the new Ryzen Threadripper 3990X, how it performs against AMD's last generation flagship and it's performance relative to the current consumer processing kings on both the Red & Blue teams.

This review will focus specifically on the Threadripper 3990X, however it’s important to understand exactly where it sits in the rest of the Threadripper 3000 product stack. AMD is offering a 24 core, 48 thread 3960X as the entry level processor for this platform.  Next, we have the 32 core, 64 thread 3970X which sits in the middle of the stack. Prior to the launch of the 3990X, it was the king of HEDT and still offers insane performance for its asking price. Lastly, we have the main event, the Threadripper 3990X. Boasting 64 cores, 128 threads and a massive 256MB cache, it’s unlike anything on the market.

Zen 2 Architecture Overview

Looking at the numbers on paper, it’s easy to see why the 3rd generation Threadripper CPU’s are dominating the HEDT market, but it’s important to understand exactly why, and how AMD managed to pull it off. One of the key areas to their success was the topology of their 3rd generation Ryzen processors. As with the previous Ryzen generations, AMD is using an MCM or Multi-Chip Module design over the single, more traditional monolithic die. The biggest difference with this generation of Ryzen processors is that AMD has moved their central IO functions off the processor die, and on to its own I/O die.

Image courtesy of AMD

As you can see in the image above, there are multiple CCD’s or Core Chiplet Die’s that contain the Zen 2 CPU cores and their respective L3 caches. All of this is connected to a central I/O die that is responsible for providing each CCD access to PCIe lanes and system memory. You will likely notice that there are multiple memory controllers on this I/O die, and that everything is weaved together by small blocks marked with the infinity symbol ∞. This symbol represents AMD’s Infinity Fabric, a high-speed interconnect through which all these dies are woven together. This unique design also allows AMD to scale their processors in a very dynamic manner as they can simply pick and choose each CCD configuration for their processors. It also allows AMD to improve their yield rates and reduce waste by using the dies that didn’t perform to their original standards in a different capacity.

The current 3rd generation Threadripper product stack is a great example of this in effect. For an example, let’s say AMD designed a 32 core Threadripper with a 4x 8-core CCD configuration. Let’s also say that in the process of picking out the best 8-core dies, they noticed that some of the cores were unable to meet the standards that they designated for this 32-core processor. Instead of throwing the dies away and wasting silicon, AMD can modify the dies and design a 4x 6-core CCD configuration to create a 24 core Threadripper. The improved yields and waste reduction aside, this also plays an important role in filling a void in a market. Without this practice, you may end up with a wider product stack with a wider gap of prices and nothing to fit in between, making it difficult for consumers to find the best product for their budget. This is also true for the 3990X, which is made up of the very same dies used in AMD's EPYC server lineup of processors. 

TRX-40 Chipset Features

Now that we’ve covered the history of the Ryzen architectures and the internal design of the new Threadripper processors, it’s time to discuss the platform. Even the best processors in the world can be detrimental to your workflow and user experience without a strong, feature-rich chipset to make use of the processing power. Luckily the TRX40 chipset offers some of best cutting-edge features on the market.

Image courtesy of AMD

AMD was the first to bring PCIe 4.0 to consumers with the launch of the Ryzen 3000 series processors, and they made sure to provide plenty of PCIe 4.0 lanes with their Threadripper CPU’s and TRX40 chipset. With the CPU boasting 64 PCIe 4.0 lanes and the TRX40 chipset offering up to 24, that is a combined total of up to 88 PCIe 4.0 lanes, 72 of which is usable. It’s important to keep in mind that these lanes offer twice as much bandwidth as PCIe 3.0. For larger render farms requiring multiple PCIe x16 GPU’s or massive high-speed data centers that require a large array of NVMe drives, this platform has you covered.

Now that we understand the platform features, it’s time to move on to our testing.

Test Suite: Synthetic & Content Creation

 

These tests focus primarily on rendering and heavily threaded workloads. Keep in mind that the Threadripper 3000 series was designed with these types of workloads.

3DMark TimeSpy Extreme Physics (AVX2)

Despite this synthetic benchmark catering more towards gamers, the 3990X puts on a very good show. With PBO enabled, we see a near 20% increase in performance.

 

Blender Benchmark 1.0B2 (Measured in seconds)

 Here you can see the 3990X is rendering the project over twice as fast as it’s predecessor, the 2990WX. This lead extends further with Precision Boost Overdrive enabled. While the 3950X’s performance is respectable given its price and core count, both of the consumer flagships are far from the 3990X for this particular workload.

 

Chaos Group V-Ray Bench 1.0.8 (Measured in Seconds)

This is a great example of how much time you can save on a render. Even if you are only cutting seconds off a render, that adds up over time, which can save a tremendous amount of money. With PBO enabled, the 3990X finishes this project in less than half the time as it’s predecessor.

 

Chaos Group V-Ray Bench 4 (Measured in kSamples)

This is where things get interesting. We see the 3950X out-scale the 2990WX despite it having twice the cores. This is because this benchmark favors both core count and clock speeds. This also explains why we see a dramatic 23% increase in performance when we enable PBO.

 

Cinebench R15 (Single-Thread)

Here we see the Core i9 9900K pull ahead for the first time, which makes sense given it’s 5ghz single core turbo speed under this kind of load. Surprisingly enough, the 3950X and 3990X are not far behind despite the clock speed disadvantage. The real shocker is that we see our first regression in performance when using PBO. We will discuss this further in our PBO & Overclocking section later.

 

Cinebench R15 (Multi-Thread)

There is nothing more satisfying than watching this classic complete in a matter of seconds. While single threaded workloads are not the 3990X’s strong suit, it is unmatched when it comes to mult-threaded workloads. We are seeing over 2x the performance of the 2990WX and when PBO is enabled, we are seeing over 3x the performance of the R9 3950X. 

 

Cinebench R20 (Single-Thread)

Now this was certainly a surprise. I expected the 3990X to perform towards the bottom here, and even expected PBO to perform slightly worse due to the lower single core clock speeds, but I did not expect the 3950X to overtake the Core i9 9900K on this text. I ran the test several times and confirmed that the 9900K was maintaining it’s boost frequency of 5ghz throughout the test. The 3950X did not even maintain its highest boost clock, settling at 4.6ghz during the test and still coming out on top with a 400mhz deficit. This warrants further study.

 

Cinebench R20 (Multi-Thread)

Now we see the 3990X stretch its legs. Unlike R15, R20 is designed to scale with high core count processors. Our PBO result is nearly 3 times as fast as the 2990WX and over 3 times faster than the Ryzen 9 3950X. 

 

Corona Benchmark 1.3 (Measured in Rays/Second)

Corona is another rendering benchmark that scales well with many CPU cores. We can see the PBO result is over 3 times that of the 32 core 2990WX. When you factor in the 2990WX’s $1400 price tag, the 3990X starts to look like a very good value for this type of workload.

 

POV-Ray 3.7 (Measured in Pixels Per Second)

This is the last test for our Synthetic/Content creation test suite. While the 3990X still showed a substantial lead over our consumer processors, it was not as pronounced as the other tests. Even with PBO enabled, we did not see a significant boost in performance compared to previous benchmarks. 

 

Test Suite: Gaming Benchmarks

 

We’ve limited our gaming test suite to just 4 titles since this processor is not really designed with gaming in mind. With that being said, we chose modern games, some of which have DX12 support in the hopes that it can highlight some performance benefits of having very high core counts. 

 

Assassin’s Creed Odyssey (High preset, 1080P V-Sync disabled)

This was an interesting result. We ran these tests 5 times instead of 3 as we were seeing weird per-test variance between each run. Surprisingly enough, the 3990X consistently had higher max framerates while also having lower minimum framerates than the 9900K and 3950X. It’s important to remember that minimum framerates are the most important metric when discussing a good gaming experience. The better your minimum framerates, the better the experience as you will have less jitter and jarring freezes. 

 

Far Cry: New Dawn (High preset, 1080P, V-Sync disabled)

Here we see the 9900K do what it does best. The Far Cry series has always favored fewer, faster cores and this shows with both the 9900K and 3950X leading the benchmarks. We also see the PBO result suffer a performance penalty due to the reduction in single-threaded performance.

 

Hitman 2 (DX12, 1080P High settings, V-Sync disabled, Fullscreen)

This is a great example of a well-optimized DX12 title. It scales well despite the lower core clocks and offers a very playable performance even on the older 2990WX.

 

Shadow of the Tomb Raider (DX12, 1080P, V-Sync disabled)

This is the last game in our test suite and yet another great example of DX12’s low CPU overhead. We even see PBO scale positively for the first time in our gaming results. If you intend to game on these HEDT processors, remember to take note of your games graphics API and be prepared to compromise on older API’s.

 

Thermals & Power Consumption

Thermals (30 Minute Blender Loop)

We ran Blender on a 30-minute loop to help fully saturate our loop and allow it to reach a thermal equilibrium. Our testing lab has a controlled ambient temperature of 23C. At stock, the 3990X idled at 41C. Average load was 52C and max load temperature was 69C. With PBO enabled, idle temperature remained the same at 41C, average jumped up to 68C and the max temperature spiked to a toasty 83C. As you can see, that increase in clock speed can get pretty warm very fast.

 

Power Consumption (Blender Benchmark Loop)

Both systems idled at 85W, with the stock 3990X hitting a max package load of 280W. The 3990X PBO result managed to reach a 599W package load under the exact same testing. When measuring the power draw from the outlet, the stock 3990X pulled 419W (system total) while the 3990X PBO result pulled 840W, almost exactly double the power. PBO is definitely not as efficient from a performance per watt perspective, but if you can spare the power overhead, it does provide a considerable performance boost in most heavily threaded workloads.