It's time at last. Today we bring you our complete test of the Threadripper 2990WX and the Threadripper 2950X. Although these two CPUs use the same basic DNA, you should know that they are very different processors that target completely different market segments.
Since they were announced earlier this year, everyone's attention has been focused on the second-generation 32-core / 64-thread thread ripper, now known as the 2990WX and costing $ 1,800. There will be two models in the WX series and for those who are wondering, the "W" means that this is a workstation series and the "X" is the usual xtreme nonsense we assume. Along with the 2990WX, there will also be a 24-core / 48-thread model known as the 2970WX. However, this model will not be available until October.
Although the TR 2990WX has received all the attention, we expect the 2950X to be the real hero of this new range in a basically nifty 1950X at an introductory price $ 100 lower.
As with the 2nd generation Ryzen 5 and Ryzen 7 models, these new thread ripper parts offer reduced cache and DRAM latency with support for slightly faster memory. So they're based on the Zen + architecture that uses GlobalFoundries' 12PL process.
The TR 2950X has the same layout as the 1950X and therefore consists of two active Zeppelin chips, each containing 8 cores, two memory channels and 32 PCIe Gen 3 lanes. When using DDR4-3200 memory, the infinity fabric throughput between these chips is approximately 50 Gbps.
As with the 1950X, the 2950X can be configured in two ways. When using UMA (Uniform Memory Access), which AMD calls "distributed" mode in its Ryzen Master software, the processor functions as a unit. This means that threads and DRAM transactions are evenly distributed across the chip to maximize bandwidth. Turn increases latency, which is not ideal for tasks like games.
It is therefore possible to activate NUMA (Non-Uniform Memory Access), which AMD calls "local mode" in the Ryzen Master software. They call this a local mode of operation because the processor is split into two domains and tries to couple active cores to local DRAM instead of accessing the memory through a controller in a separate chip, which involves a fairly high latency penalty.
The 2990WX, on the other hand, is a completely different animal. It does not consist of two Zeppelin stamps, but four that allow up to 32 cores. However, AMD has imposed some restrictions on the X399 platform in order to avoid cannibalization of the EPYC server CPUs with a socket.
The biggest of these limitations is that there are only four memory controllers left. Although there are two other Zeppelin matrices, the additional two matrices are, according to AMD, arithmetic cubes. This means that they have no local PCIe or DRAM access because they have to get to the I / O chips through the Infinity Fabric. Since there are twice as many chips, the infinity fabric bandwidth is also halved, so that the throughput between the chips is now only 25 GBit / s, provided you use DDR4-3200 memory.
Due to this design, in which two of the chips are displayed without direct access to the DRAM, the 2990WX uses only NUMA in contrast to the 2950X. According to AMD, they were able to use this Quad-NUMA configuration to create the world's first 32-core consumer processor. It is equally important that they were able to do this while maintaining backward compatibility with existing TR4 products.
There is an obvious downside that has troubled me a bit since we first heard of this 32-core model. We always knew that 1st generation Threadripper CPUs have the potential to offer up to 32 cores. So this is not a radical breakthrough for 2nd generation AMD. The original Threadripper chips had no "dummy" stamps, as claimed by AMD. We always knew that they were broken or disabled. Zeppelin-These are after all only EPYC CPUs for the desktop. Although we don't want to sound like we're downplaying anything here, EPYC CPUs on the desktop are very epic.
However, for this 2nd generation Threadripper series, AMD has activated these additional matrices to create the 24-core and 32-core models. The problem is memory bandwidth because there simply won't be enough of it. We still only have quad-channel memory access so the memory bandwidth stays the same, but now we need to feed twice as many cores.
This will likely make an already niche product even more focused. So keep that in mind. We have a lot of data to test, and although we tried to include as many CPUs as possible, we ran out of time to update the results with the Core i7-8700K.
The good news is that we have results for a number of high-end desktop processors from Intel, such as the Core i9-7980XE and the 7960X. Basically, all systems with 32 GB DDR4-3200 memory were configured using XMP timings. So let's get to the results.
Okay, we might as well get this out of the way first, Cinebench R15. As many of you are probably already familiar with, given that the AMD leaked the results, the 2990WX scores a little over 5000 pts. That makes it a whopping 52% faster than the Core i9-7980XE and at that point ask You probably figured out what the hell I was doing when I said that the design has some drawbacks, but keep that thought going for now.
However, in this special synthetic rendering benchmark, the 2990WX has no problem blowing off socks. The 2950X is also not a slouch, although it only improves the 1950X by 5%.
Next we have another rendering benchmark, although it is based on real software. The Corona Renderer was used to test workstations with over 64 cores, so it can be scaled very well. Here again we see a breathtaking rendering performance of the 2990WX that only took 41 seconds and allowed him to complete the test 40% faster than the 2950X. This isn't a perfect core scale, but it's an impressive result. This also meant that it was 28% faster than the current Intel Core i9 flagship. Again, the 2950X was only 4% faster than the 1950X, another small win, but still a win.
When I moved on, I started the Ryzen Graphic workload in Blender. This is a relatively quick test for high-end CPUs, and we can see that the 2990WX only took 8.3 seconds. This meant that the workload was completed 36% faster than the 2950X and 31% faster than the Core i9-7980XE.
Again, an impressive completion time for the 2990WX, but doubling the cores only increases the overall performance by 55% compared to the 2950X, and there is very little difference in clock speed between the two. So how about a workload that takes significantly longer than a few seconds for the 2990WX to complete?
Disappointingly, the gooseberry workload, which was much more complex and therefore longer to complete, was less affordable for the 2990X. Okay, it still uprooted the 7980XE and hammered in its pens, but it also only cut the completion time by 28% compared to the 2950X. Compared to the more expensive 7980XE, the rendering time was reduced by 20%. This is obviously a great result for AMD.
Nevertheless, it is a worrying sign that with an optimal workload for the 2990WX, there is only a 38% increase in performance and a 100% increase in cores.
Okay, POVray is the last rendering benchmark we're going to look at, and it's a good sign of the 2990WX. Here, the render time could be reduced by 40% compared to the 2950X, which meant that it was 65% faster, which in turn is not an amazing scale, but 65% is much better than what we saw in Corona and Blender. It was also 57% faster than the Core i9-7980XE, so a huge win.
I have recorded RealBench's extensive multitasking results that allow image editing, video compression and rendering tasks to be performed simultaneously. The 2990WX peaked at 70%, but at least half of the test load was down about 20%. So that's worth mentioning. Here we see that the 2990WX delivers a surprisingly bad result and takes 43 seconds to get the work done.
This made the 32-core processor slower than even the 1950X. The 2950X, which matched the Core i9-7960X and 7980XE, impressed here. This meant that the 2950X was able to complete the heavy multitasking test 6% faster than the 1950X, which in turn was a great result.