"Performance issues with AMD Ryzen Threadripper 3970X and 3990X"

As if it wasn’t hard enough trying to decide whether to abandon a lifetime of Intel-ing, and enter the Brave New World of AMD, I stumbled across this article:


The Threadripper 3960X is not mentioned, either in the headline or the article itself.

Was this because of an accidental oversight when the (undated) article was written, and the same problems are known to impact the 3960X as well . . .

. . . or was this a deliberate omission and — for whatever reason — the problems effecting the 3970X and 3990X do not apply to the 3960X . . .

. . . or has Steinberg not yet checked / determined whether the 3930X is similarly effected?


Do these problems impact Cubase and Nuendo equally, or is one program less susceptible to problems, given the identical CPU?

Thanks for any input, advice, etc.

— Alan

AMD Threadripper 32-core Full Support - Cubase - Steinberg Forums

That’s from a different user than me. Maybe ask how it’s going.

As for the article it seems reasonable that if this other poster’s 3970x is working then the article is either not up to date or the problem Steinberg saw doesn’t apply to all users. In either case it would seem that either it applies to the 3960x as well, in which case that might work too, or it never applied to it in the first place.

That’s (at least potentially) heartening news — thanks, @MattiasNYC for spotting and sharing that.

But I’m still really hoping that someone from Steinberg will stop by at least long enough to clear up the maddening ambiguity in the above-linked article’s headline and body text . . .

. . . and to bring us up-to-date regarding which specific issues were identified, which CPUs were and were not impacted, whether/how the issues were solved and, if they haven’t been solved yet, an appropriately vague and general estimation of how long it will probably take to overcome these issues.

If the TR 3960X is in fact ‘problem-free’ (at least iunasmuch as any CPU can be said to be), then it would seem to be the clear choice (for those ‘future-proofers’ who can stomach spending the extra $$$) over the 5950X — if for no other reason, for the sTRX4 socket, which will hopefully allow for at least one (Zen 3) CPU upgrade before its inevitable deprecation…

The next generation of Threadripper should come out this year. November is the last rumor I heard. So I’d personally just wait for that CPU to come out. It probably won’t be cheap though…

Oh, it absolutely won’t . . . :slight_smile:

But I actually need to build a new system ASAP.

And the reality is that, even if it does come out in November (which it probably won’t . . . ditto Alder Lake), we won’t be able to build new systems right then and there.

Whether we’re talking about the Zen 3 TR line or the Alder Lake line, there’s a good chance that the predictions of a 2021 release won’t pan out.

Whenever they are officially “released,” they probably won’t be available for purchase by jus’plain-folks until at least a good few months after their eventual release date . . . most of the initial production runs will probably wind up in the hands of the major OEMs . . . and, say, NewEgg’s pre-built-computer side-busines.

And even then, it’ll likely be the same deal as what’s going on right now with NVIDIA graphics cards: empty shelves, a glut of ridiculously-overpriced (and possibly used or misused) CPUs for sale by third-party-seller people/companies with clearly-made-up names, on eBay, Amazon, New Egg, Walmart, etc. . . . . . . meanwhile, every once in a blue moon, Tom’s Hardware or Ars Technica will type up a report about the long lines of hopefuls queuing in front of Best Buys across the country, most of whom walk away empty-handed.

Assuming that nothing horribly wrong is uncovered re the design and/or manufacture of the chips themselves, it’ll probably be a few more months before the next-gen MOBOs come out, the BIOSes/UEFIs are properly updated, and enough hardware/software conflicts are resolved to render purchasing the CPUs a reasonably safe proposition.

And I can’t wait that long . . . so I want to get the best that I can get right now, with an eye towards upgrading to the literally-just-around-the-corner New Tech (DDR5, USB4/TB4, etc.), once it has had a chance to settle in and have the wrinkles smoothed.

I would have stuck with the familiar and purchased a 10th Gen Intel chip now and wait for Alder Lake to become truly available, but the latter uses a new socket, so anything Intel that I buy now is a technological cul-de-sac. Ditto the 5950X.

That’s how I wound up looking at the 3960X. AMD has at least claimed that it will stick with the sTRX4 socket for a while . . . so, if I get the best-equipped 3960X-friendly MOBO now . . . by the time I can afford to upgrade to its Zen 3 successor (probably a year or two after its initial release), it should work**.

Now, somebody from Steinberg, please chime in and tell us what the 3960X’s Cubase/Nuendo true status is.

Not sure why Zen 3 TR wouldn’t show up reasonably soon. And last time around the ones that ended up in OEM systems were the “Pro” models with 8 memory channels etc, not the “regular” TR. I don’t think availability will a big issue since AMD has had time to bin for these CPUs for some time, and the EPYC chips have been out for a bit. I mean, TR is basically a neutered EPYC.

And if you’re in the US and near a Microcenter they have had decent availability actually - pick up only though.

I have a 3960x. Unfortunately, I don’t have any project that put a higher load than 10% on the CPU, but using Performance Toolkit, I can verify that Cubase is using all 24 cores.

for information the 5900x and 5950x work perfectly for cubase and nuendo. attention must use 3600mhz or 3733mhz memory. below the performance of the cpu is not good

“not good” or not optimal?

we find with memories below 3600 mhz audio glitsch very quickly

then I will answer: not good

with 3600 mhz memory performance is optimal and audio glitches disappear as if by magic. we see the same problem with 3900x and 3950x processors

What do you mean by “very quickly”?

in fact it depends on the level of solicitation of the cpu. with vsti like Swam with ten tracks you can quickly have glitsch. after doing a lot of testing with other users affiliated with the audiofanzine site we realized that the real time side of an amd processor was dependent on the speed of the memory. if the frequency exceeded 3733 the audio was less efficient. similarly if the frequency was too low we had the same problems. the resolution of the audio glitsch was found when we changed the memory types with an optimized frequency. 10 tracks is very little. with vsti like halion the cpu resources are smaller so we can use more tracks. but it would be regrettable not to use the full power of amd cpu. having said that an i9 11900kf will do very well. i love intel and amd. I do not like the price of proceseurs sometimes prohibitive (in France). intel supports low latency better (64 samples for example). Amd allows huge projects but with latency at 256 or 512 samples or between 7 and 13ms.

The new Threadrippers won’t help here, since it’s entirely a software issue. Most (all?) DAWs have a 32-thread limit on Windows for legacy reasons (MMCSS being one at least). So I think the best bang for the buck is a 5950X right now.

Personally I got the 3970X because I run such heavy VI sessions with VE Pro, but at the moment it is necessary to disable SMT ( simultaneous multithreading, i.e. AMD’s version of hyper-threading) on that, because Steinberg needs to either fix the 32+ thread issue or at least allow limiting Cubase/Nuendo to 32 manually like you can do on REAPER, so the rest of the system could utilize the additional cores (physical or not).

Leaving SMT on will slow down Nuendo to an absolute crawl on my system, so that’s completely out of the question. You can manually limit CPUs from Task Manager every time you launch an application as well, but that seems insane to keep doing on a regular basis. And since the virtual cores from SMT are not full performance anyway, the 3970X’s 32 regular cores are basically the most raw compute you can throw at a DAW today I think.

As far as I recall the limit was a lot lower, and it was logical cores that were listed. Steinberg wrote that support article in a funny way, but it includes " the outcome of this is that on systems built around a CPU with more than 14 logical cores, processing threads exceeding this amount will be carried out as non-real-time (like UI-threads), introducing drop-outs and performance issues."

In other words it’s 8 physical cores and up that are susceptible to problems.

From what I’ve read around the interwebz the problem pops up on different CPUs with different core counts, but it isn’t limited to 16 cores and up, just more prevalent if I understand it correctly.

Also, the Threadripper CPUs supposedly had different problems which is why they’re listed in the article on Threadripper (by Steinberg) and not the 5950X. If it really was a pure software issue for everything 16-core and up then both would have been listed, along with all Intel CPUs with the same core count.

What at least used to be an issue with the TR lineup was their architecture. The way the core complexes were laid out along with how main memory was accessed used to be a problem in some situations. I think that was adjusted on Zen 2, but it still has a somewhat less efficient architecture as far as I know. Zen 3 TR should perform a lot better all else being equal due to a different architecture.

…as far as I recall…

Again, pulling from memory, the tweak is more about the memory fabric speed than the memory speed. The two are either locked in a 1:1 ratio or 1:2. Up to 3733 it’s 1:1 and so increasing memory speed to that point increases infinity fabric. But once you go above that memory speed infinity fabric drops to half the rate, and that is why there’s a slowdown.

Obviously there’s also a slowdown relative to 3733 as you lower the rate.

yes that’s it. you have to stay at a 1: 1 ratio. from this important point the frequency of 3600 mhz was for my equipment the solution with the best performance