Fourth dimension has finally come up. Today nosotros are bringing you lot our full review of the Threadripper 2990WX and Threadripper 2950X. Although these 2 CPUs share the same bones DNA, you lot should know they are very dissimilar processors, targeting completely different marketplace segments.

Since they were announced before this yr, everyone'south attention has been stock-still firmly on the 32-core/64-thread 2nd Gen Threadripper part, now known equally the 2990WX, coming in at $1800. There volition be two models in the WX series and for those wondering the 'W' signifies that this is a workstation serial and the 'X' the usual xtreme nonsense nosotros suppose. Forth with the 2990WX there will as well be a 24-core/48-thread model known as the 2970WX, though that model won't be bachelor until October.

Although the TR 2990WX has been receiving all the attending, we expect the 2950X to be the real hero of this new lineup in what basically is a refined 1950X at a $100 lower launch cost.

As was the instance with the 2nd Gen Ryzen 5 and Ryzen 7 models, these new Threadripper parts feature reduced cache and DRAM latency with support for slightly faster memory. Then they are based on the Zen+ compages which uses the 12PL process from GlobalFoundries.

The TR 2950X features the same layout as the 1950X and that means information technology comprises two active Zeppelin dies, each packing 8 cores, two memory channels and 32 PCIe gen 3 lanes. When using DDR4-3200 retentiveness the Infinity Textile throughput between these dies is roughly l GBps.

Every bit was the case with the 1950X, the 2950X tin exist configured in one of 2 means. Using UMA (Uniform Memory Access), which AMD refers to as 'distributed' mode in their Ryzen Master software, the processor acts as a single unit and this means threads and DRAM transactions are distributed evenly across the entire chip to maximise bandwidth, but in turn increases latency which isn't platonic for tasks such as gaming.

Therefore information technology is possible to enable NUMA (Not-Uniform Memory Access) which AMD refers to as 'local mode' in the Ryzen Primary software. They phone call this a local operating way every bit the processor is separated into ii domains and attempts to pair active cores with local DRAM, rather than accessing retentivity via a controller in a separate dice which comes with a rather hefty latency penalty.

The 2990WX on the other hand is a very dissimilar beast. Information technology consists of not ii Zeppelin dies merely rather four, enabling up to 32 cores. However on the X399 platform AMD has imposed some limitations to avoid cannibalizing their unmarried socket EPYC server CPUs.

The biggest of these limitations beingness that at that place are nevertheless simply iv memory controllers. Although there are two more Zeppelin dies, the additional two dies are compute dies, in AMD's words. This means they have no local PCIe or DRAM access, for that they must travel via the Infinity Fabric to the IO dies. As at that place are twice as many dies, the Infinity Fabric bandwidth is also halved, so now the throughput between dies is just 25GBps, assuming you're using DDR4-3200 retentivity.

MSRP prices reverberate pricing at launch. TR 1950X is down to $770 as of writing.
TR 1920X is $530 and TR 1900X is $333.

Because of this design that sees two of the dies without straight access to the DRAM, information technology means that different the 2950X, the 2990WX uses NUMA exclusively. AMD says that this quad-NUMA configuration has allowed them to create the world's first 32-core consumer processor, and simply as important, it has allowed them to do it while maintaining backwards compatibility with existing TR4 products.

There is an obvious drawback that'due south had me a bit concerned since we outset heard nearly this 32 core model. We always knew that 1st Gen Threadripper CPUs had the potential to offer up to 32 cores, then this isn't some kind of radical quantum for AMD with the 2d Gen series. The original Threadripper chips didn't accept 'dummy' dies every bit claimed by AMD. We e'er knew they were defective or disabled Zeppelin dies, these are after all just EPYC CPUs for the desktop. Though we don't mean to sound like we're downplaying anything hither, EPYC CPUs on the desktop is very much ballsy.

Anyway, for this 2nd Gen Threadripper series AMD has enabled those extra dies to create the 24 cadre and 32 core models. The problem is memory bandwidth, as at that place simply isn't going to be enough of information technology. We however only take quad-channel memory admission, so retentivity bandwidth remains the aforementioned, but now we take twice every bit many cores to feed.

Likely this is going to brand an already niche product, fifty-fifty more focused, so keep that in heed. For testing we've got a truckload of information and while we tried to include as many CPUs as possible nosotros did run out of time to update the results with the Core i7-8700K.

The good news is we have results for a number of Intel's high-end desktop processors such every bit the Cadre i9-7980XE and 7960X, for case. Basically all systems were configured with 32GB of DDR4-3200 retentivity using XMP timings. So permit'southward go into the results.

Benchmarks

Okay well nosotros might too become this one out of the way showtime, Cinebench R15. As many of you lot are probably aware of by now given AMD leaked the results, the 2990WX achieves a score of just over 5000 pts in it's stock out of the box configuration. That makes it a whopping 52% faster than the Cadre i9-7980XE and at this bespeak y'all're probably wondering what the hell I was on near when I said there were some drawback to the design, but hold that thought for now.

Anyway in this particular synthetic rendering criterion the 2990WX has no problem blowing socks off. The 2950X is no slouch either though it does only improve upon the 1950X by a mere 5% margin.

Side by side upwards nosotros have some other rendering benchmark though this one is based on existent-world software. The Corona Renderer has been used to examination workstations with over 64 cores, so information technology scales very well. Here we again see breathtaking rendering operation from the 2990WX as information technology took just 41 seconds, assuasive it to complete the examination 40% faster than the 2950X, not perfect core scaling but still an impressive result. This also meant information technology was 28% faster than Intel current flagship Core i9 part. Also this fourth dimension the 2950X was just 4% faster than the 1950X, so another small-scale proceeds, just a gain all the same.

Moving on I fired up the Ryzen Graphic workload in Blender, this is a relatively quick examination for loftier-end CPUs and we see that with the 2990WX which took but eight.three seconds. This meant information technology completed the workload 36% faster than the 2950X and 31% faster than the Core i9-7980XE.

Again an impressive completion time for the 2990WX but it has to exist said for a doubling of cores we are only seeing a 55% boost in overall operation when compared to the 2950X and there is only a very minor clock speed difference between the two. So what about a workload that takes significantly longer than a few seconds for the 2990WX to complete.

Disappointingly the much more complex and therefore longer to complete Gooseberry workload was less favorable to the 2990X. Okay and then it nonetheless uprooted the 7980XE and kicked in information technology'due south pins, but it was also only able to reduce the completion fourth dimension by a 28% margin when compared to the 2950X. It was besides able to reduce the render fourth dimension by xx% when compared to the more expensive 7980XE, and then that's obvious a great result for AMD.

Still information technology's a troubling sign that in what should be an optimal workload for the 2990WX, nosotros're only seeing a 38% increase in performance, for a 100% increment in cores.

Okay then POVray is the last rendering criterion we're going to look at and this 1 bodes well for the 2990WX. Here it was able to reduce the rendering time by twoscore% when compared to the 2950X and this meant it was 65% faster, and so again not amazing scaling simply 65% is much amend than what we saw in Corona and Blender. It was likewise 57% faster than the Core i9-7980XE, so a massive win there.

I've included the heavy multitasking results from RealBench which runs prototype editing, video compression and rendering tasks simultaneously. The 2990WX saw a peak load of 70% but for at least one-half of the examination load was down around 20%, and then that's worth noting. Hither nosotros run into the 2990WX providing a surprisingly poor result taking 43 seconds to complete the workload.

This made the 32-core processor slower than even the 1950X. Hither information technology was the 2950X that impressed, matching the Core i9-7960X and 7980XE. This meant the 2950X was able to consummate the heavy multitasking test 6% faster than the 1950X, so again a great result.