Threadripper BUILDERS thread

Page 70 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

dnavas

Senior member
Feb 25, 2017
355
190
116
...the 4k footage clocks in at 67% CPU use vs the 1800X's 76%. Not sure how twice the number of active units provides so little in return, but there you are.

Well, in re-running the tests, I actually see the 4k section on the TR system hitting 71% at the front sections of the video . I can't explain why this is not that much faster than the 1800X. BUT, I can say that if I lower RAM speeds to 2933, CPU use goes back up to 77%. Well, what an interesting coincidence! Now, TR *should* have twice the memory bandwidth, so I'm thinking that what I'm seeing is really high IF use. That also squares with a theory I have on how the system is managing to run hotter than 68C Tdie -- the heat is from the IF end, and whatever algorithm AMD is really running in there discounts IF heat. It also explains why 3466 in this test crashes -- I likely need a little SoC voltage to keep the IF happy. 3466 should get me down to 66% in this test, if I can make it work.

For those watching from home -- running in NUMA is a net loss. The software supports it, and I already tried that. Which is odd, and a shame, but there you are. Maybe "channel" mode is not the right thing to set in the Taichi board? I was confused why I wouldn't use "die", but the ASRock BIOS .. yeesh.

Anyway, if I'm right (and it seems a reasonable theory, although it's just as easy to theorize this is really RAM and not IF :|), the next time AMD tells me that their IF subsystem ... well, let's just say that the rumors of no NUMA in single socket Rome is *very* interesting to me. And if AMD wants a test to run against, I'd be more than happy....

Ed: Got 3466 stable by "upping" my LLC on the SoC to 3 (from 5). Only good for 2-3% -- about half of what one might expect. Random observation: halving and then re-doubling the size of the 1080 clip (I've got it resizing into 4k) is actually good for another 4% speed improvement. :shrug: Oh, and idle is more stable in mid 30s.
 
Last edited:

dnavas

Senior member
Feb 25, 2017
355
190
116
Yeah, I dunno about custom WC -- that's a lot of work, and ... I do not have a good history with patience on the mechanical side of building stuff. I just absolutely destroyed the tip of my kryonaut paste trying to get it out so I could put the cap back on after installing the liqtech. Destroyed as in the tip is now in two pieces :| Do I really want to tackle custom water? Hmm :|
I think I'm going to keep after my missing 3-4 cores of performance, instead. Man's got to know his limitations....

I've noticed that the kernel use is sitting at around 25% on my Windows 10 TR box, and only about 5% on my Windows 7 1800X box (while running video in my editor). Something suspicious there, but haven't determined what it is yet. May need to install some tooling :|

The video that uses 67% w/25% kernel time uses only 50% w/10% kernel time if I set process affinity to every other "core". Sounds like contention to me. I'll be installing some tooling at some point to see if I can track down where. Although I don't have a great way to determine fps on the system, I can hear that removing every other core when the system is fully utilized does wind up with less throughput ... marginally.
 
Last edited:

dnavas

Senior member
Feb 25, 2017
355
190
116
So, umm, has anyone left HWMon open for a while, come back, and seen their threadripper indicate it's boosted all the cores to 4.7G at some point?! Because that just happened to me last night, and I'm wondering what I did! :lol: And of COURSE I haven't been able to reproduce it, are you nuts?! All I have is PBO active....

https://drive.google.com/open?id=1i31jUnZfpbBJhb4NG8u-sqxvO6JLEYN8

(okay, all except 'lucky' core 13)
 
  • Like
Reactions: lightmanek

captain_nemo

Junior Member
Oct 4, 2018
1
5
16
Hi guys
After reading for about a year, I've finally built my threadripper system.
I was playing a little with idea of going with 2990WX but I would have probably never used it at its full potential and it's too expensive anyway.
I got
2950X,
2X G.SKILL TridentZ 16GB 2x8GB 3600MHz DDR4 CL15-15-15
Gigabyte X399 AORUS XTREME
Noctua NH U14S TR4-SP3 - Noctua
Samsung 970 Evo M.2 500GB
NVIDIA 1080
I've chosen gigabyte because I got badly burned with ASUS (twice), MSI doesn't ring my bell, and ASROCk was currently unavailable (at least in my country). Gigabyte low - middle range boards almost always were "plug and play" for me. I was a little afraid because there is little information about this boards on forums, and oficial reviews nowadays are a "little" vague. For example almost every review site goes about how close distances between super IO armor and memory slots are. TridentZ fit there with with like a two milimeters gap. Same goes with CPU cooling - but if wanted all 8 dimms I would have to find another cooling solution. The board itself is pretty well made, and building a system went like a charm (succesful boot on the first time), but BIOS is ... Let's just say it is there. Orange letters on the dark background are horrible enough. The information is scattered in submenus, settings are unintuitive.For example: you've got some setting set auto. Click it and nothing happens. You have to start writing and finally some values appear. I could go on but I'm not writing a review here :)
Given how unclear tr4 cooling situation is I went with Noctua. I would love to get AIO but... Custom WC - I don't have time, skill and I'm not sure if I'm willing to experiment on such a rig. Been building system for a few years now (those 300A's :)) but had never touched anything wet yet (computers wise):) I'll wait a little maybe there will be some AIO wich doesn't destroy itself. Was thinking about thermalright Silver arrow but I couldn't find direct comparison with noctua and it wasn't available.
I'm little worried about my voltages on auto. I don't plan to OC I just want productivity and longevity. If I set auto in ryzen master I get big voltage spikes. VID 1.550, soc voltage 1.100, cpu core voltage 1.537 and vcore 1.476 seem a little bit high. I couldn't find any official guidelines on threadripper 2 but everybody seems to agree that original threadripper safe limit is 1.450V. I know Vid voltages doesn't mean much but vcore is a little too high for me. If set manual I can get 4ghz@1.275V (hwinfo 1.308V) which seems stable and reasonable.
 

VirtualLarry

No Lifer
Aug 25, 2001
56,229
9,990
126
That big a difference in performance, between the different timings in the 3600 RAM? And I thought that only B-die could do 3600, no Hynix yet, or can it? The cheaper $137 DDR4-3600 kit is like CAS19.

Edit: The difference in price is ... pretty substantial. Like $100. $137 versus $226?
 

IEC

Elite Member
Super Moderator
Jun 10, 2004
14,323
4,904
136
That big a difference in performance, between the different timings in the 3600 RAM? And I thought that only B-die could do 3600, no Hynix yet, or can it? The cheaper $137 DDR4-3600 kit is like CAS19.

Edit: The difference in price is ... pretty substantial. Like $100. $137 versus $226?

3600 CL15 is better binned B-die. But I haven't been able to get better than 3466 CL14 running on Ryzen, so you probably won't be able to hit 3600 on it anyways (booting into OS doesn't count - I'm talking about stable).

However, you can probably get 3200 CL16 from the cheaper kit (it's Hynix chips) which would be "good enough" unless you're aiming to have the absolute best. But it may require tweaking of subtimings depending on what motherboard you have.
 

VirtualLarry

No Lifer
Aug 25, 2001
56,229
9,990
126
However, you can probably get 3200 CL16 from the cheaper kit (it's Hynix chips) which would be "good enough" unless you're aiming to have the absolute best. But it may require tweaking of subtimings depending on what motherboard you have.
Well, 3200CL16 would be way "good enough" compared to my current kits. I don't know if I'm knowledgeable enough about the sub-timings to be able to tweak them effectively, and make them compatible.
 

dlerious

Golden Member
Mar 4, 2004
1,772
718
136
Well, 3200CL16 would be way "good enough" compared to my current kits. I don't know if I'm knowledgeable enough about the sub-timings to be able to tweak them effectively, and make them compatible.
I'm running 3200 CL14. The biggest reason I went with the tighter timings is because I can loosen them up if I need bandwidth more than latency (for example 3466 CL15). The biggest downside with threadripper is with it being quad channel, costs go up pretty quickly.

My other problem was going with 16GB sticks - my ECC is only DDR4 2400 (64GB was about $550).
 

lightmanek

Senior member
Feb 19, 2017
387
754
136
Haven't checked out the level1techs forums yet, but Wendell just put up a video where he's running 2990wx with 128GB of memory at 2933 (non-ECC) and 2666 (ECC).
https://www.youtube.com/watch?v=5u6DY8On1XA

Excellent video from him, as always! Dedicated 128GB 2933 Threadripper kit from G.Skill and regular 128GB 2666 Corsair ECC. This is a very serious workstation with 2990WX CPU on a 'budget' compared to anything else currently available.
 

Paratus

Lifer
Jun 4, 2004
16,614
13,297
146
I'm running 3200 CL14. The biggest reason I went with the tighter timings is because I can loosen them up if I need bandwidth more than latency (for example 3466 CL15). The biggest downside with threadripper is with it being quad channel, costs go up pretty quickly.

My other problem was going with 16GB sticks - my ECC is only DDR4 2400 (64GB was about $550).

I’m also running 3200 CL14. I’ve managed to get it to 3333 Cl14 and seen some increases in a few bench marks I figure between CL14-16 and DDR 3200-3600 we’re talking only a few % improvement.
 

dlerious

Golden Member
Mar 4, 2004
1,772
718
136
I’m also running 3200 CL14. I’ve managed to get it to 3333 Cl14 and seen some increases in a few bench marks I figure between CL14-16 and DDR 3200-3600 we’re talking only a few % improvement.
Are you running with 128GB?
 

JDG1980

Golden Member
Jul 18, 2013
1,663
570
136
I have 4 of these still version one, and no hope of getting version4, I am just not spending any more time or effort on these, just hoping to give them bad PR for what they have done to me WASTING $600 ON 4 OF THESE.

You might want to look into filing a chargeback with your credit card company. I hear that AmEx is especially good about this.
 

dlerious

Golden Member
Mar 4, 2004
1,772
718
136
Just 32GB 4x8GB Samsung B-die. I’m not running anything that requires 8DIMMs.

Plus the price for 32GB was $480 at the time.
I was interested in that video mainly because 1st gen Threadripper struggled a little when trying to use all 8 slots, especially with dual rank memory at speeds above 2133. I don't remember what I paid for my FlareX 3200C14, but it was more expensive than the Trident Z (or V) - I grabbed it within a couple weeks of availability. I've been hoping to see prices closer to $10/GB for 3200.

Edit: This is the last kit I looked at https://www.newegg.com/Product/Prod...VScDICh0-ngKWEAYYASABEgKKBvD_BwE&gclsrc=aw.ds
 
Last edited:

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
Has anyone got the Liqtech II yet? Anything to report if you do?

My Enermax packed in the other day, but have arranged to get a Liqtech II when they come in as part of RMA.


In the meantime to keep things ticking over, got a Silver Arrow TR4* - but a mis-threaded pillar screw has derailed installation temporarily. New screwset should be with me next day or two.

*Also got two extra fans to install to make it a triple 140mm arrangement.

Will report back on how it goes for thermals.


Does anyone want to suggest a single benchmark we can all use to compare cooling performance? (I can estimate room ambient using the house heating thermostat.)
 

dnavas

Senior member
Feb 25, 2017
355
190
116
Has anyone got the Liqtech II yet? Anything to report if you do?

What would you like to know?
I have the liqtech ii 280 and am running it now. It arrived in good condition.
I remain concerned with how much Tdie jumps around (10C up for no reason I can discern), but I don't think that's related to the cooler.
In my estimation, this can remove about 250W of heat from the cpu with a 38C delta (from ~30C room temp to 68C Tdie) in push/pull configuration. Watts from the wall of about 375 -- I'm not sure where the diff between HWMon reported watts and wall watts lies. :shrug:

Test it in Prime95, and run with small ffts. Sucker produces heat like it's going out of style....
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
How long have you had it running?

Probably my main concern would be performance degradation - which would mean you'd have to keep a close eye (if not record) on idle/load temps (and clocks) to see if the former is rising/ latter dropping.
 

dnavas

Senior member
Feb 25, 2017
355
190
116
How long have you had it running?

Probably my main concern would be performance degradation - which would mean you'd have to keep a close eye (if not record) on idle/load temps (and clocks) to see if the former is rising/ latter dropping.
Well, the liqtech iis have not been out all that long. I've had mine for a week? or two? I don't think I'm going to see a lot of degradation in that timeframe :)

I've got bunches to say about TR, AIOs, and what I think about 'em, but that isn't answering your question, so I'm going to shut up now :)
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
Well, the liqtech iis have not been out all that long. I've had mine for a week? or two? I don't think I'm going to see a lot of degradation in that timeframe :)

You might wanna record your idle temps and load temps/clocks under prime (or similar) now, so you can compare in future to see if performance has dropped off.
 
  • Like
Reactions: dnavas

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
Noctua AMD Threadripper Air Cooler Roundup.
[H]OCP test the NH-U14S, NH-U12S, and NH-U9 on the 2950X and 2990WX.

Brilliant! Thank you for the link.

I guess next thing for them is to test Enermax just to see how much a functioning one would bring to the table.

Still waiting on my screw pillars then a triple spool SilverArrowTR4 becomes a fully operational technological terror.
 
  • Like
Reactions: IEC and DethGasp