AMD Ryzen Threadripper 2990WX

[HEXUS]

Helluva fast.

Read more.

[Dashers]

Thanks for the review.

I might be out of touch, but who is spending close to £2K on the CPU alone to then play games? I can see the point in these CPUs for people doing heavy multi-threaded processing (unless it's handbrake, natch), but for those who do gaming and work, I find the game-mode very off-putting.

My understanding is the last gen threadrippers had to be put into game mode too to get good framerates and this was to do with infinity frabric and the multiple cores being essentially standalone chips in a single package; where the Intel chips don't suffer from this as they are single processors.

The comparison I want to see is the entry level of the HEDT chips, as this is where people who want decent PCIe lane counts and what not have to buy in at. I'm willing to pay (a small amount) more to Intel for not having to use Gaming Mode and lower energy costs. I think Intel are probably aware of this too - the question can they make their price attractive enough to have their arguably more elegant solution come up with the very effective blunt object that is Threadripper.

Either way, I'm lagging behind on a measly 6/12C Haswell-E at the moment, which turns four this year. I haven't even looked at over clocking it yet, and I can't see a need to switch CPUs next year either, which at fives years I would normally do.

[Corky34]

It seems the WX's leans more towards the professional, content creating, and workstation end of the HEDT market than the gaming end, by all accounts the X parts are more balanced towards the gaming end.

I tried looking for 32 core Xeon gaming benchmarks so as to draw some comparison but either my Google-fu is letting me down or no ones ever ran gaming benchmarks on high core count Xeons.

[lilyx]

AMDs live launch thingy has just started for the 2990WX on twitch. Am I allowed to post the link?

[CAT-THE-FIFTH]

The 2990WX hits two issues - it seems that many applications under Windows can get utterly befuddled by 64 threads(run over 4 separate chips),plus running 4 Ryzen chips off only a pair of memory controllers seems to not help either. Linux benchmarks seem a bit better in that regard though,if you check them in the review thread:

https://forums.hexus.net/cpus/393554...ew-thread.html

However,for the most part the 2950X seems to make more sense unless you need lots of cores.

[badass]

What I don't understand is why Game mode is 8 core 16 thread? IIRC There are four 8 core CCX's in the package. Two of the CCX's have a dual channel memory controller.

Why not have game mode disable the 2 CCX's with no memory controllers and disable hyperthreading so each core has double the L1, L2 and L3 cache? Same 16 threads as well.

Even better, can this be fixed by setting processor affinity for games? If so, then no reboot required.

[Tattysnuc]

Wonder how games perform in a VM environment rather than switching to Gaming mode?
I understand that there's a (small) overhead of running software in a VM, but that would be interesting for me to see how it compares...

[Corky34]

What I don't understand is why Game mode is 8 core 16 thread? IIRC There are four 8 core CCX's in the package. Two of the CCX's have a dual channel memory controller.

Why not have game mode disable the 2 CCX's with no memory controllers and disable hyperthreading so each core has double the L1, L2 and L3 cache? Same 16 threads as well.

From what i understand there's also a 1/2 core mode but it seems AMD have called the 1/4 core setting 'gaming' mode, maybe it's a compatibility thing or performance? Do games play nice with 16 cores nowadays?

[scaryjim]

What I don't understand is why Game mode is 8 core 16 thread? …

As I read it, there are a number of options available for legacy modes and Hexus chose a 8C/16T mode, but I could be wrong. Also I note it doesn't say which cores are enabled/disabled so it could easily be doing that 8C/16T across 2 dies. 16C/16T across 2 dies might be better, it might not be. I'd assume AMD have done some testing, and that either performance or technical reasons have informed their decisions over legacy game mode settings...

… Even better, can this be fixed by setting processor affinity for games? If so, then no reboot required.

Nope, the issue is the game engine detecting the extra cores/threads. Some games won't even start if they detect more than a certain number of threads. Just setting core affinity won't do the job...

Wonder how games perform in a VM environment rather than switching to Gaming mode?
I understand that there's a (small) overhead of running software in a VM, but that would be interesting for me to see how it compares...

AFAIK virtualisation of GPUs is something that is still pretty patchy. I'm not even sure whether gaming GPUs support full hardware virtualisation.

Surely it's be a better idea for people who want to game to buy/build a machine for gaming. The WX Threadrippers are so obviously content creation focussed that it makes absolutely no sense to game on one when you can build a passable gaming computer for less than the price of the 2990WX...

[MLyons]

Surely it's be a better idea for people who want to game to buy/build a machine for gaming. The WX Threadrippers are so obviously content creation focussed that it makes absolutely no sense to game on one when you can build a passable gaming computer for less than the price of the 2990WX...

but muh e-peen.

[lilyx]

but muh e-peen.

I don't have an e-peen but I may have a couple of e-vries because I am having feelings of big love for threadripper

[Tabbykatze]

I have to admit, I struggled to read your review as it felt like there was a lot of onus on AMD being at fault for software not being able to handle that number of Threads efficiently.

I think TTL is the only reviewer I've seen so far (out of 3) whom has been balanced on the issues surrounding massive core numbers. Some reviews have acrually outright said AMD is not good in this and that multi threaded workloads. But Tom has actually gone insofar as to highlight that some tests suffer at high core counts wherein even the 7980XE is impacted.

The biggest elephant in the room that may be dressed over the coming years as massive numbers of cores become main stream (AMD have gone from a quad pair of cores (FX) to a fully fledged 8/16 in 6 years and 16/32 in 2 years and Intel in mainstream have gone from 4/8/10 to 6/8/18) in the last 2 years.

Benchmarking software needs to be updated to account for this. As noted, Cinebench looks like the only Benchmark to truly handle the massive numbers of cores and threads.

Some reviewers have a large focus of their review of the 2990WX about gaming and it shows it as trash. It's not designed for gaming at all, it is trash relatively. I mean, you can play a game and probably satisfied with the reaults but it's not going to be an i7-8700K, or even a R7 2700X, experience in the slightest if not ever!

It looks like reviewers may need to take a leaf out of Server benchmarking for the large Multithreaded workload testing.

Great numbers in many content creation spaces but it is interesting that two of the CCXes become useless in some workloads because it is accessing the Chipset via another CCX. This is a failing in the design that i hope can be addressed one day. It means that technically 16 cores are gimped in some workloads.

[Corky34]

As I read it, there are a number of options available for legacy modes and Hexus chose a 8C/16T mode, but I could be wrong. Also I note it doesn't say which cores are enabled/disabled so it could easily be doing that 8C/16T across 2 dies. 16C/16T across 2 dies might be better, it might not be. I'd assume AMD have done some testing, and that either performance or technical reasons have informed their decisions over legacy game mode settings...

I assumed the same however I'd be really interested in how the different modes carve up the cores, like you say is it 8C/16T across 2 dies (that seems the most logical), or is it 8C/16T on a single die with some I/O requests going to a die with disabled cores, it could even be disabling six cores per die so the 8C/16T get evenly divided between the four dies.

[CAT-THE-FIFTH]

Great numbers in many content creation spaces but it is interesting that two of the CCXes become useless in some workloads because it is accessing the Chipset via another CCX. This is a failing in the design that i hope can be addressed one day. It means that technically 16 cores are gimped in some workloads.

There is probably a good reason AMD has not allowed 8 channel memory access on X399 - if they did so,it would probably compete with its own single socket Epyc platform.

[scaryjim]

... it could even be disabling six cores per die so the 8C/16T get evenly divided between the four dies.

This would fit better with AMD's usual approach of balancing the cores across the available CCXes, although I suspect they've gone for 4+4 across 2 dies. It'd be interesting to see power reading for game mode - if they are only using 2 dies they should be able to gate most of the inactive dies...

... it is interesting that two of the CCXes become useless in some workloads because it is accessing the Chipset via another CCX. This is a failing in the design that i hope can be addressed one day. It means that technically 16 cores are gimped in some workloads.

I'm not sure they're "useless" exactly; Hexus haven't given much of the TR2 context but other reviews go into much more detail - supported by many AMD slides - on how the various modes and accesses work. The design balances huge thread counts against memory access, and the non-memory-connected chips are not really any worse off than the other two half the time - if *any* of the dies needs to access memory it's not directly attached to it has to go via the off-die Infinity Fabric. All that does is increase the memory latency, and while that *will* slow some tasks down, those extra cores are still going to make a significant difference in well-threaded tasks.

For tasks which are latency sensitive AMD provided a local memory mode which keeps data local to the thread it's running on (as much as possible) - and since those tasks tend to be more lightly threaded it's not so big an issue. To call it a "failing" in the design is simply unfair - it's a carefully considered compromise.

For people who *need* that low latency access I strongly suspect there's a local memory mode on EPYC that would serve their purposes just as well (and 1P 32-core EPYC isn't that much more expensive...). And AFAIK the very high core count Intel's also have variable latency times (for the server Xeons with EPYC level core counts Intel even advisea 4-node NUMA set-up - the same as EPYC!).

There is probably a good reason AMD has not allowed 8 channel memory access on X399 - if they did so,it would probably compete with its own single socket Epyc platform.

it's simpler than that - the TR4 socket just isn't connected for 8 channel memory. Sure, there's enough pins - indeed it's physically the same as SP3 - but the way it's wired up is different, and there's no way to provide 8 channels of memory from a TR4 socket. The only way to provide 8 channel memory would be to allow mobo manufacturers to produce X399 motherboards with SP3 sockets on them, and that'd get VERY messy. The way they've done it makes the most logical sense, and I have to say I'm pretty impressed by what they've achieved here - 32 cores @ 3.3GHz with DDR4-2933 memory within a manageable TDP? Kudos

EDIT to add: @CAT - tbh I'm more surprised that they've provided up to 32 cores on socket TR4. The higher core count was one option for driving people towards EPYC for tasks like rendering, but instead they've bumped up the core counts, at a lower price, on a more affordable and more desktop-oriented workstation platform. That's interesting, imnsho....

[CAT-THE-FIFTH]

it's simpler than that - the TR4 socket just isn't connected for 8 channel memory. Sure, there's enough pins - indeed it's physically the same as SP3 - but the way it's wired up is different, and there's no way to provide 8 channels of memory from a TR4 socket. The only way to provide 8 channel memory would be to allow mobo manufacturers to produce X399 motherboards with SP3 sockets on them, and that'd get VERY messy. The way they've done it makes the most logical sense, and I have to say I'm pretty impressed by what they've achieved here - 32 cores @ 3.3GHz with DDR4-2933 memory within a manageable TDP? Kudos

EDIT to add: @CAT - tbh I'm more surprised that they've provided up to 32 cores on socket TR4. The higher core count was one option for driving people towards EPYC for tasks like rendering, but instead they've bumped up the core counts, at a lower price, on a more affordable and more desktop-oriented workstation platform. That's interesting, imnsho....

TR4 won't be wired up to drop costs and to stop it competing with Epyc.Intel blocks ECC on its HEDT platform but AMD does not. Also another consideration is dropping a pair of memory controllers will reduce power requirements but I don't think AMD really made TR4 to take a 4 chip CPU.

I think it is more likely the 2990WX exists because AMD wants a halo product to beat Intel but with a lower price,as the 18 core chips are nearly 500MM2 in size and any higher core count Intel chips will be even larger.