Ticker

6/recent/ticker-posts

AMD Ryzen Threadripper Pro 5995WX Review

 This is the most expensive AMD Ryzen CPU you can buy, which as you might suspect is also the fastest and most powerful Ryzen CPU currently in existence. Not even upcoming Zen 4 CPUs will change that, at least not when talking about core-heavy workloads.

The Ryzen Threadripper Pro 5995WX is a 64-core/128-thread workstation CPU that's actually been out in the wild for about 5 months now. AMD announced and released this monstrous CPU back in March, but it was available for OEMs only. Starting today, the Threadripper Pro 5000 series is going retail, meaning you can now build your own insane workstation if you have enough money, because damn are these things expensive.

The base model 24-core Threadripper Pro 5965WX comes in at $2,400, the 32-core 5975WX is $3,300 and the 64-core model which we are testing today costs a cool $6,500.

Internally we've been using the Threadripper 3970X and 3990X for video processing, all expensive CPUs in their own right, but still cheaper than the Threadripper Pro 5000 models. The 3970X was $2,000, which is about 40% less than the new model, while the 3990X was conveniently priced at $3,990, once again ~40% less than the newer Zen 3 version.

 

Threadripper

Pro 5995WX

Threadripper

Pro 5975WX

Threadripper

Pro 5965WX

Price

$6,500

$3,300

$2,400

Retail Release

August 8, 2022

Cores / Threads

64 / 128

32/64

24/48

Chiplets

8 x CCD - 1 x I/OD

4 x CCD - 1 x I/OD

Base Frequency

2.7 GHz

3.6 GHz

3.8 GHz

Max Turbo

4.5 GHz

L3 Cache

256 MB

128 MB

TDP

280 watt

If you're wondering why the 'Pro' models cost so much more, it's because they're legitimately more workstation oriented. For example, you get twice as many PCIe lanes from the CPU, 128 versus 64 while the current AM4 desktop parts have 24 in total, 16 for the primary PCIe slot, 4 four the primary M.2 and then 4 connecting the chipset.

Another big difference is memory support. Whereas AM4 processors support dual-channel memory, HEDT chips like the 3970X and 3990X utilize quad-channel memory. However, the Threadripper Pro series goes a step further by supporting octa-channel memory. That's eight DDR4 memory channels just like it's offered on Epyc server processors.

Essentially, the 5995WX that we're testing today is an overclocked Epyc 7713P, though the server CPU costs less at around $5,000 but also clocks much lower. Speaking of clock speeds, the 5995WX packs a base clock of 2.7 GHz with a boost clock of 4.5 GHz, 22% higher than that of the Epyc equivalent. This chip consists of eight CCD's and an I/O die, giving it 64-cores/128-threads with a beefy 256 MB L3 cache.

The 32-core/64-thread 5975WX features four CCD's and an I/O die, clocks to at least 3.6 GHz with the same 4.5 GHz boost clock. Then the 24-core/48-thread 5965WX also packs four CCD's and an I/O die, clocks no lower than 3.8 GHz underload with a peak operating frequency of 4.5 GHz and both the 24 and 32-core versions feature a 128 MB L3 cache.

All three models support 128 PCIe 4.0 lanes, 2TB of DDR4-3200 octa-channel memory and a TDP of 280 watts. This means AMD has officially killed off the HEDT version of Threadripper and replaced it with the workstation version called Threadripper Pro. So what's changed and why have they done this?

Simply put, demand for Threadripper didn't justify a more affordable version. It likely made sense back in the Zen and Zen+ days when mainstream desktop platforms topped out at 8 cores, but today 16-core models like the Ryzen 9 5950X do a good job of satisfying the needs of most power users.

Also back in the early days of Ryzen, AMD was desperately trying to gain market share in all segments, so offering compelling products at as many price points as possible made sense. Today though, AMD seems to dominate most fronts and with absolutely no HEDT competition from Intel, it's not a segment they need to fiercely compete in, and for most high-end users the Ryzen 9 5950X will suffice.

Now, when it comes to WRX80 motherboards your choices are still quite limited with under a dozen models to pick from. The boards that are most likely to appeal to our audience include the Gigabyte WRX80 SU8 IPMI (cool name we know), the Asus Pro WS WRX80E-SAGE SE WiFi (an even cooler name), the Asrock WRX80 Creator, simple and to the point, we like it, and then the board we'll be using today, the MSI WS WRX80, again simple and to the point product name.

We could have had our pick of these WRX80 boards but we went with MSI for a few reasons. We weren't impressed with the look of the Asrock and Gigabyte boards, those VRMs look okay for stock 280w usage, but a late change by AMD has seen the Threadripper Pro 5000 series support overclocking, and this means those boards will need to be refreshed if they want to properly support OC functionality for parts like the 5995WX.

Asus was also caught out by this late change and their board doesn't support features such as PBO. So it would seem taking some extra time paid off for MSI as the upcoming WS WRX80 board supports both auto overclocking in the form of AMD PBO as well as manual overclocking control and with an 11 direct phase 105A design it's already setting records, pushing the 5995WX to an insane 5.15 GHz for a score of 116142 points in Cinebench R23, though please note we are talking about a LN2 overclock here. The point being though that this VRM is good to handle whatever you throw at it.

Having said all that, sadly for this initial review overclocking is off the table, and the reason why is a bit of an unknown. MSI says my pre-production board was tested before being sent out, and that both PBO and manual overclocking worked. However, after several BIOS updates and endless tinkering the OC features refuse to reveal themselves, and we can't work out why.

MSI will be providing a second board and hopefully that allows us to overclock the 5995WX, but for now we're stuck with the stock results, which we've got to say aren't half bad. We've also included a configuration with the power limits removed. For testing, Team Group provided four kits of their Team T-Force XTREEM ARGB DDR4-3600 CL18 memory, giving us 256GB in total, though for stability the memory was limited to DDR4-3400, not 3600. We didn't have eight memory modules of the same spec, so Team Group came in clutch, making this review possible alongside MSI and AMD, of course.

For comparison we've included the Threadripper 3990X which was tested on the MSI Creator TRX40 motherboard using the same Team T-ForceDDR4-3600 memory at 3400. Then we have a few desktop CPUs for comparison: a Ryzen 9 5950X tested with dual-rank DDR4-3200 memory, and the Intel Core i9-12900K using DDR5-6400 memory. The graphics card used for all testing is the GeForce RTX 3090 Ti. Let's get into the data…

Benchmarks

Starting with Cinebench R23, as expected the 5995WX scores big with 66403 pts stock. Then with the power limits removed that score crept up by 5% to 69581 pts. Interestingly, the 5995WX was just 6% faster than the 3990X, though this isn't terribly surprising given the 5950X was only around 9% faster than the 3950X in this test.

The single core gains are more impressive, here the 5995WX was 14% faster than the 3990X, but also 9% slower than the 5950X and almost 30% slower than the 12900K.

We're only looking at an 8% improvement in the 7-zip file manager decompression test for the 5995WX over the 3990X. But this time removing the power limits resulted in a slightly more impressive 9% performance increase.

For code compilation work the 5995WX doesn't appear all that much faster than the 3990X which makes sense as the 5950X was only slightly faster than the 3950X. So for this sort of workload the more expensive Zen 3-based Threadripper processors probably aren't worth the premium.

The Corona benchmark is finished in almost the click of the button when using the 5995WX, taking just 18 seconds. Still that meant it was 22% faster than the 3990X, despite only shaving 4 seconds off the completion time.

Another rendering benchmark is V-Ray and here the 5995WX was 15% faster than the 3990X, so not a huge performance gain, but about what we've come to expect for this sort of workload.

Premiere Pro is of particular interest to video creators and it's where Threadripper shines. We've got more to say about this which we'll save for later in the review, but for now let's check out the Puget Systems benchmark results.

Here the 5995WX was 28% faster than the 3990X, which is a significant performance advantage and this translates into serious real-world improvements. We believe the increased memory bandwidth and improved single core performance is largely responsible for the performance gains seen in Adobe Premiere.

For those of you using Photoshop 2022, high-core count CPUs are a waste for the most part. There might be plug-ins that can utilize more cores for specific tasks, but in general this is a lightly threaded application and therefore higher clocked, higher IPC parts like the Core i9-12900K are better here.

The last application benchmark is Blender Open Data and we'll also look at power usage in this test soon. Once again when it comes to rendering benchmarks the 5995WX is generally only a little faster than the older 3990X, here we're looking at 7% more performance, so nothing to write home about.

System Power Consumption

Power consumption is similar to the 3990X and it was interesting to see that removing the power limits for the 5995WX didn't see power consumption spiral out of control, in fact, we saw little to no changes. It's a shame we couldn't enable PBO or do any manual overclocking, but that's something we can look at in future content.

Gaming Benchmarks

Time for a few gaming benchmarks, just because you might also want to game on your workstation, rather than use an entirely separate system for that. The 5995WX is impressive in Shadow of the Tomb Raider, delivering slightly more performance than the 5950X and we have seen this in the past. The 3990X was faster than the 3950X, for example.

The Hitman 3 results are a bit more typical, where the 5995WX was slightly slower than the 5950X, a mere 3% decline in performance which meant it was still 18% faster than the older 3990X.

Rainbow Six Extraction sees similar margins to that of Hitman 3, the 5995WX was 4% slower than the 5950X, but 10% faster than the 3990X. However, removing the power limits allowed the 5995WX to match the 5950X.

The 5995WX was on par with the 5950X in Watch Dogs: Legion making it 15% faster than the 3990X, but much slower than the 12900K.

We're also looking at comparable performance between the 5995WX and 5950X in Horizon Zero Dawn, meaning the new 64-core processor is 17% faster than its predecessor.

Cooling

If you're familiar with the Threadripper 3000 processors, then this new 5000 series won't offer any surprises, it's basically the same deal. The layout is much the same and the power consumption is much the same, therefore the cooling requirements are about equal.

For this review we used the be quiet! Silent Loop 2 360mm AIO with the be quiet! STRX4 mounting kit. Like most AIO's the Silent Loop 2 isn't purpose built for the massive Threadripper CPUs, instead it was designed with LGA1200/1700 and AM4 processors in mind. So it doesn't exactly cover the entire heatspreader of a Threadripper processor, but it gets most of it.

We did run into an issue with our build as the Dark Base Pro 900 Rev 2 case was unable to fit the Silent Loop 2 360mm AIO on the top of the case as the MSI WS WRX80 motherboard is simply too massive, and as a result we have four sticks of memory exactly where the radiator needs to go.

This means we were forced to front mount the radiator which normally isn't an issue, however the Silent Loop 2 water block is designed to only be oriented in a single direction with the be quiet STRX4 mounting bracket, and using the correct orientation the tubes for the AIO simply wouldn't reach with the radiator front mounted. So we were forced to install the block backwards in the bracket which sees it positioned well off center on the Threadripper IHS.

Therefore we had to get the Dremel out and crudely cut a section of the STRX4 mounting bracket, so that the block would be positioned in a more central location when installed backwards. Anyway, it was a bit of messing around but we now have the block centrally located with what appears to be good mounting pressure.

Of course, we tested thermal performance before we got benchmarking and everything looked good. After an hour of looping the Cinebench R23 benchmark we hit a peak temperature of 83C, which given the power usage is a good result. That said, it's unlikely to be optimal and if we look closely at the peak core complex die temperatures we see that only two peaked at around 80C, then two more which peaked in the mid to low 70s with what we assume are the four centrally located CCD's peaking at between just 57 and 61C.

What We Learned Using the Pro

The Threadripper Pro 5995WX is an impressive processor and although for the majority of our testing it wasn't a great deal faster than the 3990X, it can be around 30% faster depending on the workload. However even that percentage gain can be fairly meaningless in a workstation setting, so here's what happened in ours...

As we mentioned before, Steve and Tim's video editing workstations are currently using the Threadripper 3970X and 3990X and while both have been game changing, allowing us to stabilize dozens upon dozens of b-roll clips at once while also providing impressive encoding performance, the recent upgrade to a Panasonic GH6 mirrorless camera has caused a few editing related headaches.

The Threadripper 3000 series had no issues with 4K60 footage from the older GH5, but when we upgraded to the GH6 we wanted to start working with 5.7K60, and this proved troublesome. Of course, we could use proxies but they are a pain for editing, taking a lot more time, and overall just a more cumbersome method than we prefer.

We had been actively testing for alternatives, the Core i9-12900K was amazing at scrubbing through timelines for general editing, but as soon as we wanted to stabilize more than a dozen clips at once, the system became unusable and would often crash. Meaning we had to stabilize clips in batches, saving after each batch was complete before starting the next batch. It wasted a lot of time and the encoding process was also slower.

A better, but admittedly much more expensive solution would be to upgrade to Threadripper 5000 and the 5975WX would probably be the best option for our needs. The first thing we did with the 5995WX was to open one of the current projects we were working on and start editing. The experience was totally different. At no point did the 5995WX struggle with the 5.7K footage like the 3990X did, even color-corrected footage with transitions played back smoothly.

You could say editing is now "more playable," to use a gaming term.

In short, the 5995WX is a game changer for our workflow and as such it's difficult to put it into the context of percentage gains, it's not just faster, it's wildly more practical. You could say editing is now more playable, to use a gaming term.

We'd say we could easily justify spending $3,300 on the 5975WX, plus everything else you need to build the workstation PC, as it would pay itself off, both in terms of productivity and mental health.

On that note, it's difficult to give something like the 5995WX a value rating, if it can do for you what it can do for us then it's invaluable and will pay for itself. It's a workstation product, designed for professionals and with no retail competition for Intel, it's your best option. We're happy AMD is giving customers the chance to build their own workstation PCs and not forcing them to buy mediocre OEM systems, so that in itself is a win.

We expect placing the Threadripper Pro 5995WX with its rates over the coming months as well as years, as well as we hope we can have a look at the 5975WX quickly as we feel that version is better suited to our usage instance.


Post a Comment

0 Comments