Ryzen

[Stereodude]

I repeated the same case closed P95 Small FFT hour+ test today with PBO off with just the stock clocks/voltages. It settled at 64.5C and was running ~3875MHz and ~100W. So ~3% performance costs ~33% in power and ~15C.

 

[Handruin]

I with you on how much more power is needed for such a small boost. I was finding it fascinating how it is working (or lack there of).

Maybe I hit some bad silicon lottery and/or I'm missing a setting somewhere. If I run the same small FFT test with PBO off (Default mode), Coupled mode on, and everything else stock, mine also settles around 64C at roughly 92W and 95A but with a speed of around 3,630MHz.

 

[Stereodude]

I with you on how much more power is needed for such a small boost. I was finding it fascinating how it is working (or lack there of).

My guess is that by putting in really large numbers some divider settings are changed which reduces how granular the adjustment can be. It's also possible that such large values break the tuning algorithm. Or maybe it is just buggy.

Maybe I hit some bad silicon lottery and/or I'm missing a setting somewhere. If I run the same small FFT test with PBO off (Default mode), Coupled mode on, and everything else stock, mine also settles around 64C at roughly 92W and 95A but with a speed of around 3,630MHz.

That's the speed reported by Ryzen Master, not another utility right? I saw numbers slightly above and below what I posted on the graph in Ryzen Master for the cores. Like 3866 and 3883... I just sort of averaged them in my head and reported that.

It's possible the Gigabyte Aorus motherboard is slightly more aggressive even on the default "stock" settings. But I would expect it to show up at the expense of power (which it seems like it does slightly). What does Ryzen Master show as the PPT, TDC, & EDC at the default settings?

Here's my defaults right after booting with no Ryzen Master profile loaded:


Windows is set to the Ryzen Balanced power plan.

 

[Stereodude]

I also have 3 x 140mm Noctua for the case and the motherboard temp never moved from 38C.

Which 140mm Noctua fans are you using? I was thinking to perhaps swap out the stock Fractal Design 140mm fans for either the Noctua NF-A14 PWM or the BeQuiet Silent Wings 3 PWM high-speed. Either can move more air than the included Fractal Design X2 140mm fan though not by much per the specs. I'm skeptical of the airflow specs given the RPM difference.

 

[Handruin]

I went with the Noctua NF-A14 PWM chromax.Black.swap (which are essentially the same as the brown version per the specs). I've always been very happy with Noctua fans over the years in my other systems. I've never had one fail and I feel like they provide adequate airflow for cases with minimal noise. I have not done any kind of specific test with the same to check airflow rates or loudness readings so my positive points are just subjective. My case is definitely quieter now that I manage each of the three NF-A14's via the MB fan headers. I wasn't able to get the Fractal fans to be PWM managed via the fan hub in the back of the case. Had I tried them on the MB fan headers maybe they would have been better but I had minimal expectations from the 3-pin fans.

One warning I will give is that the cable length of the NF-A14 was barely long enough for my lowest front fan to reach my fan header. I know you have a different MB so my point is to check the length to your headers. Noctua says the cable length is 31cm long total because it comes with a single extension cable.

I can't speak to the comparison with the BeQuiet series; they may be equal or better. I admit to being a Noctua fanboy so I'm biased toward their products. Both the Noctua and BeQuiet fans look to be around the same price point and specs so you probably can't go wrong with either.

 

[Stereodude]

I think my HTPC has at least 1 BeQuiet fan in it. I will have to open it check what I ended up with in there. I know I bought several of them around the time I got the GeForce 1080 for it. I bought several different 140mm fans testing them previously in the HTPC. I put 140mm Noctua fans in the Fractal Designs R5 case.

I may get one of each and compare them for this new build.

Edit: Scratch that, the price of the Be Quiet! fan jumped by $4 at Amazon. I will not be buying one at the higher price. https://camelcamelcamel.com/quiet-Silent-High-Speed-BL071-Cooling/product/B01JME13SW

 

[Handruin]

I replaced my thermal paste with the Grizzly Kryonaut and reran some P95 tests. I can't say if it made much difference yet, if only slightly. I do note that when I run Prime95 with the default small FFT torture test (PBO off), I can sustain about 63C with the CPU fans at roughly 20% speed (~600RPM) and case fans at 35% (600RPMs) and the system is mostly inaudible except for a slight HDD whine. The ambient temp is still around 18C in my office.

 

[Stereodude]

What were the CPU fans spinning at before? Mine are close to 100% ~1100RPM at 64.5C and are definitely not inaudible.

 

[Handruin]

They were running somewhere around 800-900RPM for the same workload on the CPU. The case fans were about the same since they're set to monitor the MB temp, not the CPU temp.

 

[Stereodude]

They were running somewhere around 800-900RPM for the same workload on the CPU. The case fans were about the same since they're set to monitor the MB temp, not the CPU temp.

Maybe I will get some and try it to see how it compares to the stuff Scythe included. What size did you get? I saw some people on Amazon complaining they got what they presume to be fake stuff (due to the bad performance).

Edit: Turns out I already have some Hydronaut. Which has slightly lower thermal conductivity (which is better [I think]), but higher thermal resistance (which is worse).

Edit2: Based on this: https://www.hwcooling.net/en/the-test-of-27-thermal-compounds-part-2-en/2/ It's probably not worth messing with it.

 

[Handruin]

I bought the little one with 1g of thermal paste just to try. I don't know how to tell if it's fake; it came in a sealed package with Thermal Grizzly's name all over it that makes it look official.

I should have looked to see what came with the NH-D15 for thermal paste. I probably got a tube of the NT-H1 I could have tried.

 

[sedrosken]

I find it interesting that in Stereodude's latest screenshot, it appears to be idling at ~300MHz all-core. That's actually way lower than the lowest idle I've seen from a modern chip (the N3050's 480MHz is what I'm comparing against) and I find that interesting -- is that kind of power-savings only available at the crazy high end, or can I expect something similar out of my 1600AF once it arrives? I know I don't get access to Ryzen Master -- I can't remember if that's restricted to the 3000 series in general or just their non-budget line, but I'm very interested to see what I can get out of mine.

@Handruin I think the NH-D15 came with NT-H1, but I don't remember if my NH-D15S came with it (edit shortly before posting: it did, I don't know why I didn't use it). I seem to vaguely remember that I've got Arctic Silver 5 between my IHS and heatsink. I'll probably dig my box out and see if I've still got mine and that's likely what I'll use on my 1600AF.

 

[Stereodude]

I find it interesting that in Stereodude's latest screenshot, it appears to be idling at ~300MHz all-core. That's actually way lower than the lowest idle I've seen from a modern chip (the N3050's 480MHz is what I'm comparing against) and I find that interesting -- is that kind of power-savings only available at the crazy high end, or can I expect something similar out of my 1600AF once it arrives? I know I don't get access to Ryzen Master -- I can't remember if that's restricted to the 3000 series in general or just their non-budget line, but I'm very interested to see what I can get out of mine.

I'm not sure what features the 1600AF has. The 3950X will actually go even lower. It will straight up put cores to sleep and just show "sleep" for most of them.

 

[sedrosken]

Well, considering it's just a 2600 hiding behind the old 1600 name (the AF moniker is an unofficial one to differentiate it from the original 1600 whose P/N ends in AE) I'd probably say anything it can do, the 1600AF can do as well. I'm not sure what that means, and likely won't until I get it in the post and test for myself, but I was mostly asking rhetorically there. It'd be nice to see it put most of my hardware to sleep, but I'd never see it anyway -- idle for me is around 10-20% CPU usage here on my 3570K since I use this as a bit of a home server as well.

I'm not sure what I'm doing with this once I've got the Ryzen hardware built in here -- I have a few ideas, ranging from just selling it to recoup some costs to building a server out of it. I might even give it to my dad, he's still running my old Phenom II x4 945 machine. But I digress.

Like I said, it'd be cool to see it put cores to sleep, but if I ever get a server running and don't need this machine up 24/7 anymore, it likely won't ever be just going to sleep -- I'll be turning it off.

 

[Handruin]

@Handruin I think the NH-D15 came with NT-H1, but I don't remember if my NH-D15S came with it (edit shortly before posting: it did, I don't know why I didn't use it). I seem to vaguely remember that I've got Arctic Silver 5 between my IHS and heatsink. I'll probably dig my box out and see if I've still got mine and that's likely what I'll use on my 1600AF.

I checked and the NH-D15 does come with NT-H1. I should have just tried that first.

 

[Handruin]

Well, considering it's just a 2600 hiding behind the old 1600 name (the AF moniker is an unofficial one to differentiate it from the original 1600 whose P/N ends in AE) I'd probably say anything it can do, the 1600AF can do as well. I'm not sure what that means, and likely won't until I get it in the post and test for myself, but I was mostly asking rhetorically there. It'd be nice to see it put most of my hardware to sleep, but I'd never see it anyway -- idle for me is around 10-20% CPU usage here on my 3570K since I use this as a bit of a home server as well.

I'm not sure what I'm doing with this once I've got the Ryzen hardware built in here -- I have a few ideas, ranging from just selling it to recoup some costs to building a server out of it. I might even give it to my dad, he's still running my old Phenom II x4 945 machine. But I digress.

Like I said, it'd be cool to see it put cores to sleep, but if I ever get a server running and don't need this machine up 24/7 anymore, it likely won't ever be just going to sleep -- I'll be turning it off.

The sleep state will happen when you have a CPU with numerous cores like these. Not all tasks/processes will make use of multithreading and the OS is getting smarter to stick/pin threads to specific cores and let others sleep. During the times when you're using your system without maxing out the CPU, putting certain cores to sleep should happen. This isn't specific to you shutting down the system to save power, it's meant as an efficiency thing during normal usage.

 

[Stereodude]

Not all tasks/processes will make use of multithreading and the OS is getting smarter to stick/pin threads to specific cores and let others sleep.

To add to that it's supposed to stick them to the cores that can run the fastest out of all the cores in the CPU.

I saw this yesterday while the system was mostly idle while the SATA drives were going through their RAID initialization. About 12 of the cores were asleep and there were 4 active cores in CCX 1 in CCD 1.

 

[sedrosken]

So would this be something I might see on Linux, perhaps when using a different CPU scheduler? I currently use MuQSS on my ck-patched copy of 5.5.2. It'd be cool to see it power down all but maybe 2 cores when I'm "idling" but I understand if that sort of thing is restricted to not only the 3000 series, but the Enterprise end of it. Plus, it's a desktop machine anyway, power savings is nice but not a priority -- I imagine I'd save more power by being able to turn my DisplayPort monitors off without losing window placement (one day I'll make those cables that shunt the HPD pin to a resistor or whatever that was) or being able to let them sleep.

 

[Stereodude]

I know I don't get access to Ryzen Master -- I can't remember if that's restricted to the 3000 series in general or just their non-budget line, but I'm very interested to see what I can get out of mine.

Are you sure about that? AMD's documentation says it will install and run under Windows 10 on any AMD Ryzen processor. The Ryzen processors before the 3000 series have a lot less supported features, but it is supposed to run.

It'd be cool to see it power down all but maybe 2 cores when I'm "idling" but I understand if that sort of thing is restricted to not only the 3000 series, but the Enterprise end of it.

Where do you keep getting these understandings from? The 3950X isn't an enterprise CPU. Generally these sorts of features are inherent to the architecture and would extend across all models.

 

[Handruin]

So would this be something I might see on Linux, perhaps when using a different CPU scheduler? I currently use MuQSS on my ck-patched copy of 5.5.2. It'd be cool to see it power down all but maybe 2 cores when I'm "idling" but I understand if that sort of thing is restricted to not only the 3000 series, but the Enterprise end of it. Plus, it's a desktop machine anyway, power savings is nice but not a priority -- I imagine I'd save more power by being able to turn my DisplayPort monitors off without losing window placement (one day I'll make those cables that shunt the HPD pin to a resistor or whatever that was) or being able to let them sleep.

Tracking down the changes made for Ryzen 3 CPUs may take some time but I suspect they are there or being worked on. I'd think you're a bit too bleeding edge with MuQSS to know what benefits you'll get or miss with Ryzen.

It's not just literal power savings that are the goal here. Thermal management is very much a consideration in order to give other single-threaded cores the ability to boost within the packages TDP requirements and maintain that boosted speed on a sticky thread scheduled by the OS. This can be achieved by sleeping cores that are not being used to draw less powered and produce less heat that needs to be managed.

 

[Stereodude]

We bought our cases too soon.

Introducing Define 7 and Define 7 XL

Fractal Design today launched the latest models in the esteemed Define series: 7 and 7 XL.

www.fractal-design.com

 

[Handruin]

I saw it last night and thought the same thing. It's still a good case but maybe could have got them at closeout/discount.

 

[Stereodude]

I did try messing with the PBO numbers the other night a little bit though I didn't bother to change them in the BIOS but with Ryzen Master. I did basically see what was described in the video you linked before. It will run slightly faster with higher power usage with numbers closer to the limits of how much it can consume. I put the Noctua NF-A14 PWM in the case. Spun up to 100% it certainly is loud. I had to set the fan curve so it does not exceed 1000RPM. I guess I could use the ULNA to slow down the maximum speed a bit, but that's not really going to be verydifferent than just shifting the PWM control down.

The included X2 140mm Fractal Design fans aren't bad fans, but their lack of PWM control makes them hard to control from the motherboard. Their response is not very linear to the motherboard controls.

 

[Handruin]

My NF-A14 rarely spin up much past 800-900RPM. I tied their speed to the motherboard temps versus the CPU. Did you couple yours to the CPU temps? They can get noisy at 100% but I don't see myself ever needing that. I think the included FD fans max out at 1000RPM so the Noctua fans can go a bit faster.

I am disappointed that fractal included three fans without PWM given the cost of the case. This isn't a budget case by any means.

 

[Stereodude]

My NF-A14 rarely spin up much past 800-900RPM. I tied their speed to the motherboard temps versus the CPU. Did you couple yours to the CPU temps? They can get noisy at 100% but I don't see myself ever needing that. I think the included FD fans max out at 1000RPM so the Noctua fans can go a bit faster.

I am disappointed that fractal included three fans without PWM given the cost of the case. This isn't a budget case by any means.

I coupled the exhaust case fan in the back to the VRM temp. The front two fans are tied to other temperature sensors on the motherboard. The front two fans are still the included FD fans.

 

[sedrosken]

Where do you keep getting these understandings from? The 3950X isn't an enterprise CPU. Generally these sorts of features are inherent to the architecture and would extend across all models.

Well, it's certainly a lot more well-heeled than what I've got, and it's more "pro-sumer" level at least. I'd figure they might have saved that specific feature for chips that could actually make more use of it, as in, those meant for use in servers and/or high-end workstations. Much like how Intel gated off VT-d to their K SKU until Haswell, I figured maybe this would be restricted to higher end chips.

Hm. So Ryzen Master would run -- not that it matters a whole ton, since I run Linux on here anyway. Maybe I'll play around with it next time I have a Windows install kicking around on here.

Got everything in, and dialed in my overclock while I slept/worked today. I'm at 3.9GHz/1.3v. I might see if I can bump that voltage just a little bit down, because it was 100% rock-solid all day in both temps and reliability in my Prime95 (mprime) tests -- half the time spent in small FFTs, the other half spent in large FFTs to stress the memory. It's satisfied my requirements -- 3.9GHz allcore on an 85 dollar 6c/12t CPU is more than I ever thought I'd see.

 

[Stereodude]

So I'm actually trying to use the R9 3950X and I have to say it sucks! It's got a lot of CPU grunt, but network performance (wired gigabit) is beyond abysmal. I get like 20-50MB/sec on file copies from my server. The speed is super inconsistent. Pushing data from it to the sever or my E5-2687Wv2 is fast.

Edit: iperf results going each way look fine, no idea why SMB3.1 file transfers are so slow going one way.

Edit2: The Mellanox Connectx-3 10gig card seems to have no such problem. Just the gigabit Intel i211 NIC on the motherboard.

 

[Handruin]

That's definitely odd. I'm guessing some kind of driver problem or maybe a faulty chip?

I'm not having any 1Gb network issues on my motherboard's Intel i211. My NAS is running SMB/Samba via Linux and I map a drive to my desktop. My house is wired with CAT 6A from my office to my HPE switch which then connects to a rack mounted netgear 1GB switch. I don't have anything elaborate software configured and I can keep network transfer pegged near 1Gb/sec. I haven't tried iperf with mine, only file copies.

I just copied a 2GB video file from my desktop on an 7200RPM HDD to my NAS:


Here's the same file copied from my NAS back to my desktop on my SATA 3 SSD different from the original test:


Not really much CPU used and you can see windows reports close to 1000Mb/sec (987) on the send.

 

[Handruin]

I did a little more investigation and I changed my NAS to force a minimum SMB version to get it more similar to yours. I ran a new test copying a 17GB file to my NAS and it stayed at 113 MB/s the entire time via a SMB 3.11 connection.





I also tried a combined send/receive at the same time to my NAS and both directions seemed within reason.

 

[Stereodude]

Yeah, it's weird... My transfers one way look like this:



Moving data the other way looks like your results.

CAT6 cabling with my Mikrotik network gear. The results using 10gig are as I'd expect (and not like this) copying the exact same files from the same sources.

 

[Handruin]

Maybe some weird anomaly with the Mikrotik and the Intel i211? Any chance the MTU is not matching between the switch and your desktop for the 1Gb ports? I could see that causing some performance oddities if you change it to MTU 9000 for the 10Gb cards and it set globally.

 

[Stereodude]

Maybe some weird anomaly with the Mikrotik and the Intel i211?

I'm skeptical of that because the iperf results were 900+mbps each way. I'm pretty sure the last time the Mikrotik units were causing me trouble (before the firmware update) the problem showed in iperf as well, not just SMB. I will have to see how it does transfering to and from another 1gig system instead of 10gig systems.

Any chance the MTU is not matching between the switch and your desktop for the 1Gb ports? I could see that causing some performance oddities if you change it to MTU 9000 for the 10Gb cards and it set globally.

Everything is at the default. I'll have to look a little closer into the details. I was basically out all day and only just got home a little while ago and won't get any time on this until tomorrow.

 

[Stereodude]

I was able to benchmark my new R9 3950X system against my E5-2687W v2 in x265 using the settings I've been playing around with for high image quality encodes.

Summary:
E5-2687W v2 - 0.417 fps (4 simultaneous encodes combined)
R9 39050X - 1.391 fps (4 simultaneous encodes combined) [CPU not saturated]
R9 39050X - 1.770 fps (8 simultaneous encodes combined)

So, better than a 4x speed improvement. That's a pretty serious upgrade.

 

[Stereodude]

Maybe some weird anomaly with the Mikrotik and the Intel i211?

It looks like you may have been correct after all. 1gig to 1gig does this whether pushing or pulling.



Updating all the Mikrotik gear to SwOS 2.10 fixed this problem with all my other PCs and I haven't seen any other reports online regarding this issue for versions 2.9 or later. I'm puzzled why it's still happening with the R9 3950X's i211 adapter. Maybe there's some incorrect NIC setting that is causing the issue and it's not really the same problem.

Did you load any drivers for the NIC or are you using what Windows 10 installed for it? I was using the drivers included in Windows 10 1909 for it. I'm installing the latest Intel drivers now and will report back and differences.

Edit: Intel drivers = no improvement vs. included Windows drivers.

 

[Stereodude]

I did more iperf testing and used NetIO-GUI. Neither showed any issue maxing out the connection either way. I never tried NetIO-GUI before, but other people used it to show the Mikrotik issue in the past. I guess I'm stumped. I don't really want to try a clean install of Windows and I plan to use the R9 39050X with 10gig fiber, not 1gig. It's only on 1gig now because I don't have fiber run to where it's currently at (while testing).

 

[Stereodude]

In my x264 compression test the new Ryzen 9 3950X (stock) decimates my old i7-4770k (oc'd). 54.26 fps vs 10.39 fps. The E5-2687W v2 does 17.73 fps.

 

[sedrosken]

In everything I've thrown at it my R5-1600AF handily outpaces my 3570K, even with the clock deficit. Even single-threaded tasks usually run noticeably better -- not world-shatteringly so unless it makes use of AVX2, but still, the gains are there. I was a little surprised actually.

I got 3200 CL16 out of my RAM at 1.4v. I'm hesitant to push it any further because while I know my board has pretty much the best voltage regulation in its price bracket, it's still a pretty budget-level board and I need this thing 100% stable, 100% of the time. Messing around and dialing in overclocks is one thing, but once stress-testing is done and it's in production use, it needs to hold up under pressure. Same logic behind putting my CPU at 3.9GHz/1.2875v. I can get 4.0 or more out of it if I want to run right up against the VRMs' limitations, but the extra 1-200MHz isn't worth it to me.

 

[Stereodude]

I will give AMD some credit. They seem to have done a decent job working with Microsoft on how threads are distributed to the Ryzen 9 3xxx cores by the OS (Windows 10 1909). It automatically spreads out threads (at least in x265) to different physical cores first before using the "hypertheaded" logical cores. This is good because running 8 threads on 8 different physical cores is ~85-90% faster than running 8 threads on 4 different physical cores/8 logical cores. My E5-2687W v2 does not spread threads in the same way.

 

[snowhiker]

We bought our cases too soon.

Introducing Define 7 and Define 7 XL

Fractal Design today launched the latest models in the esteemed Define series: 7 and 7 XL.

www.fractal-design.com

Linus installs 320 TB into a Define 7 XL case.

Works out to just about a $1000 (NO drives/video card) for a huge capacity NAS setup.

 

[Handruin]

I watched the video and it's a bit ridiculous where he mounted some of the drives. Those drives are about $500 each. If I was going to spend that much on drives I'd find a better way to mount them so they all get proper cooling.

Also his plan with zfs under proxmox was not ideal for performance. The more optimal config would have been 2 x 10 disk raid z2 vdevs with ashift 12. That would balance and align the writes across 10 disks.