Threadripper & Ryzen only builds (3rd Gen starts on page 50)

[fraine7]

Forgot to ask what cooler you are running on the 2990WX? I can order one and see if it has any effect, if not I can always send it back within 14 days

[Juraj]

My build is on page 8, it's custom water loop.

There isn't much choice right now:

-Single available AIO is Thermaltake Liqtech 360 II, but as you can read two pages ago, even this most recent iteration is having leaking issues. Big nope.
-Some upcoming AIOs like Coolermaster and Asus have full coverage of heatspreader...but this is bit false alarm. This is just additional copper plate on top of the same asetec above. It's better...but not as good.

So outside of custom water loop, Air coolers are actually the more sensible choice.

-Noctua UH14 TR4. At stock settings, would be ok. Potential 320W rate (at high RPMs).
-Silver Arrow TR4. Looks like it's finally here, has more heatpipes than Noctua and is considerably larger. A true alternative to non-existing NH-D15. Except no reviews... outside of two crappy youtube ones. Here is one:

https://geizhals.eu/thermalright-silver-arrow-tr4-100700418-a1848365.html?t=alle&plz=&va=b&vl=de&hloc=at&hloc=de&hloc=pl&hloc=uk&hloc=eu&v=l


And well that's it, nothing else.

[fraine7]

Very useful info, thanks so much!

Just ordered a Noctua UH14, I will look into custom loops at some point, once this animation is out of the way I have hundreds more images to get through so I just need something stable for now.

Huge thanks for your help so far!

[fraine7]

So I can render a still image for hours, no issues at all!

5 frames into an animation, complete system failure - I’m clueless as to what is going on

[fraine7]

Me again ;)

So I found a Reddit thread which discusses a 68c temp limit for the 2990WX - had no idea this existed before building this machine
(https://www.reddit.com/r/Amd/comments/9gidma/amd_precision_boost_overdrive_pbo_ramblings/)

As my CPU was hitting this AMD specified limit, the chip was failing and the motherboard had stopped communicating with it completely, I'm guessing some kind of fail-safe in the 2990WX causes this but it is not publicised anywhere and has driven me crazy for weeks!

I have now disabled every possible option for running this chip above the 3ghz base-clock, it has an auto feature which takes it up to the 250w limit but the temperature limit counteracts this and causes system failure.

Max wattage for the CPU is now capped at around 215w (250w limit) and the benefit is that my temperatures have yet to exceed 60c - and guess what, my animation is chugging along nicely.

Don't let any of these issues put you off building a machine with this chip - when it works it really is unbeatable value for money. Just be aware that there are a ton of built-in functions which try to push this chip to increase speeds, but then those same built-in functions have to fight against the temperature limit of this massive chip, and that's where the failures begin to happen.

Unless I can work out how to A) increase my total wattage allowed to be used by the CPU above 250w, and B) keep my temperatures below 68c (apparently this isn't currently possible due to AMD's own overprotective restraints) then I will have to use this chip at it's lowest and slowest speed of 3ghz.

So happy I can now leave the room and let my animation fly knowing I won't come back to black screens and an unresponsive box of heat!

PS - also had to lower RAM speeds from 2666mhz to 2133mhz - again resulting in slower rendering but at least it works

[Juraj]

I didn't know of this, hmm, surprising. Nonetheless, it should throttle, not crash.

To fully bypass PBO you can set fixed multiplier. But unless you fix it to 4.0 Ghz for which you require very solid cooling, you will loose single-thread performance as I didn't see option to set multiplier for different amount of cores (like I can do for my i9 with Asus uefi).

I think we moved into territory where top HEDT chips like 2990WX, i9 7980XE,etc. will simply require lot of juice (lot of VRM phases) and cooling. This is one place where these builds sort of fail compared to dual-cpu server type of workstations. Now Threadripper is actually very solid in terms of TDP, the 250W is actually 250W, my i9 stock is 280+ W (so much for claimed 165W, which only goes for base clocks that never actually apply).

The upcoming Z399 Intels (22c and maybe the mythological 28C ? But I doubt the latter) will be even worse. Regular cooling need not apply..

[fraine7]

Good to know.

Lowering my clock speeds has allowed the system to stabilise and I can only think that the turbo features built in to the bios were pushing too hard. A throttle as opposed to a complete failure would have been fine, I would have been happy with that.

Could it be that my ram voltages were being pushed too high and causing the failure? The motherboard was displaying a red LED indicating CPU error but if the ram had failed then the CPU would naturally become unresponsive and force the motherboard to illuminate the LED.

I really don’t know enough about all the under the hood tuning features but it’s rendering now and until my knowledge is substantial enough to start playing around I’ll have to leave things as they are.

[vansan]

Finally my CPU arrived, built my rig and now I have statistics from 2 workstations with similar hardware:
AMD 2990wx, Asrock X399 Taichi, NZXT Kraken X72, 4x16 DDR4 2933 ram modules, Fractal Design Define S case.

No overclocking, CPU temperature is 62-64 on rendering, clocks 3.2-3.3 GHz, all stable.

[Juraj]

Happy rendering, hope the shit works :- ).

[Vuk]

Really sad to hear some people are still having problem with the 2990WX but I guess all will be ironed in the next few weeks, maybe months. Seems that the TR 2 launch has the same bad start as Ryzen when it was first introduced.
Juraj what time did you manage to get in corona render benchmark with the new machine?

[Juraj]

Stock settings (+/- 3.3Ghz) = 39 seconds, PBO auto (+/- 450W @3.95Ghz = 36 seconds). Manual PBO (500W @3.95- 4.05) also...36 seconds.

It does not scale that well. Because I have overkill cooling (1400W heh, the fans don't even need to spin at full overclock) I believe I can successfully ignore PBO and set the multiplier for x42 and get maybe 33 seconds (I did not set it this way because it's running on Beta bios and I need this PC to be super stable right now) But what is funny is that by this, it has close to 6200 CB15 score, super close to top platinum xeons, but those will reach easily into twenties.
So Corona can't utilize Threadripper as well. Some things to consider is that the benchmark isn't linear, it still have considerable pre-render phase and right now my 2990WX only runs by 30perc. during them, slowing stuff like UHD,etc..

So it's really good, but it's not great. It scales extremely well in some other "benchmarks" I run, namely Marvelous Designer and PTGui 360 stitching. I did not have time to run Vray. We're in super busy work phase, I had to set it up and immidiatelly get it to work for Veronik.

I would still suggest it 100perc. It's much more affordable even compared to OEM/ES 8173M/8180M at their best ebay price, and absolutely superior in single threaded operations which I believe will only improve with the upcoming dynamic mode (automatic process lasso/scheduling in Ryzen master utility). But there is some wasted potential and I can only wonder if it's due to memory (plausible, since 2133 vs 2933 yielded drastic difference) or partly Corona's architecture.

(I believe the dynamic mode would alleviate the issues with Corona's prerender phase as I believe this is where memory bandwidth becomes issue. This set of processes are not multithreading well from start to finish so as long as they will be ofshored to cores with direct memory access, perhaps it can shave some time on benchmark. It will not have any impact on longer rendering time).

[danio1011]

Hi Juraj - Just as another datapoint, my threadripper 1950X (OC'ed to 3.88ghz) hits 93% usage during the first few seconds of UHD Cache and then drops to 80-82% for the remainder.  For displacement it's lower....50-60%.  30% is pretty low but maybe it's the ratio of processing power to RAM changing how perceptible the bottleneck is?

Cheers,
Daniel

[Juraj]

I wish I would know what's behind it. In practical terms, it's difference of 2 seconds versus 3 seconds (i9 7980XE vs 2990WX), but it can potentially hinder interactive performance. Surely doesn't feel that way but still.
We're talking daily builds so maybe it's tied to it but I don't have time to try regular build.

It reminds me how parsing on nodes (2698v4 Xeons) during Distributed rendering is much slower than when run directly from 3dsMax and it mostly seems because it barely uses available performance in multithreaded fashion.
The whole precomp phase is bottleneck, but more so on Threadripper than i9.

[Shawn Astrom]

https://goo.gl/DPdd5x

I have the new threadripper 2 32 core and it's been great does run a little hot but nothing that can't be tamed with a good cooler.

- Shawn

[arqrenderz]

We need a new benchmark scene, we need to run it for at least 5-10 minutes and have a # of passes as the end result.