Stalled kees KogelMier - a tiny powerhouse

[jeshikat]

Wood cable combs would look really nice in a Noctua themed build.

 

[iFreilicht]

espeerimint

Literally three results on google for this word, congratulations!

Spent some time making a wiring harness for the R9 Nano. Made some cable organisers just for giggles.

Looking good, did you laser-cut those?

 

[QinX]

Literally three results on google for this word, congratulations!



Looking good, did you laser-cut those?

Sorry Josh but, that is easier to find then the NFC S4 Mini
The next case from NFC should be called the Espeeri Mini

Yes I did laser-cut those. I might not keep them in the end, but I can get a little OCD when it comes to cable management, sometimes

 

[QinX]

This post will be edit a couple of times to today as I progress, I'm not the type to make 1 big post in one go with this sort of stuff

Dat moment when you're package arrives much sooner

Finished the PCB, I lost my tweezers so apologies for the awfull soldering on the 8pin SMT chip.

First try it boots!

Now for some bad news
I can't load the GPU, it crashes the system.
My conclusion is that the 190W 12V AC/DC adapter isn't enough to power the GPU.

More testing will need to be done by me, but a quick try using the HDPLEX and a 350W AC/DC Adapter on another motherboard the Nano seems to work fine.

Okay I'm trying to narrow down what the cause is.
With the R9 Nano powered by the HDPlex and a 350W AC/DC adapter you would think it get enough juice.

Test system #1:
ASUS Q87T
4790K
16GB RAM
PCIe Riser

Running Heaven at 960x540 and 0xAA I get 5 fps
Running Furmark at 800x600 and 0xAA I get 76 fps
Running Furmark at 1024x768 and 0xAA monitors go dark instantly.

Now moving to Test system #2:
Gigabyte H81TN
4785T
8GB RAM
No Riser

Running Heaven at 960x540 and 0xAA I get 90 fps
Running Furmark at 800x600 and 0xAA I get 140 fps
Running Furmark at 1024x768 (7.8 MPixel) and 0xAA monitors go dark instantly. <--- Strange behaviour
Running Furmark at 1280x720 (9.2 MPixel) and 0xAA I get 119 fps
So the amount of pixels doesn't cause the monitors to go dark instantly lets try at the same vertical resolution 1366x768

Running Furmark at 1366x768 (10.5 MPixel) and 0xAA I get 111 fps
Screen didn't go dark instantly, but as I was typing this it still crashed at some point, Powerdraw at the wall was a mere 100W nothing substantial.

Running Furmark at 1920x1080 (20.7 MPixel) and 0xAA monitors go dark instantly.

This is really weird stuff
I can run Unigine Heaven 1080p 8xAA just fine and powerdraw at the wall peaks at around 200W so what gives.

Last edit for today, going out for Christmas eve.
Furmark crashes the PC at 720p 0xAA. I've got it running at 8xAA and it has been stable for 5 minutes. I feel like it depends on the load that is applied to the GPU. Maybe with 8xAA it has a more stable powerdraw compared to 0xAA but I'm just guessing now.

I should have asked for an oscilloscope for Christmas.

I'm open to any suggestions!

Okay one more
I swapped the Nano for a HD5870 for fun

R9 Nano
Running Furmark at 800x600 and 0xAA I get 140 fps

HD5870
Running Furmark at 800x600 and 0xAA I get 62 fps

Switched the powermeter to show peak power and it hit 262W. that might be part of the reason why it crashes. the R9 Nano does some extreme power management
http://www.tomshardware.com/reviews/amd-radeon-r9-nano,4285-9.html

They measure peaks up to a whopping 437W, the average is 186W but those peaks need to be handled.

 

[dondan]

Do you have a normal ITX board to test it?
Maybe the PCIe Port is the problem.
With no riser you have only 25W on thin itx and with your special powered riser you maybe have 75W but only if it work as expected. Some weeks ago I read an artical about the special powerdraws of the PCIe port on different gpus. If i remember right different gpu generations draw differen power values on 12V and 3,3V.

 

[Phuncz]

You are using FurMark, which is known to cause unrealisticly high loads. It often leads to throttling with more recent cards. I would advise on staying with Heaven benchmark.

If the peaks are well over the sustainable power, a large enough set of condensators might do the trick, but you probably know more about that than I do. The issue could be in the design of the power brick, as it is probably not specced for these kinds of peaks, while ATX PSUs probably do.

 

[QinX]

Do you have a normal ITX board to test it?
Maybe the PCIe Port is the problem.
With no riser you have only 25W on thin itx and with your special powered riser you maybe have 75W but only if it work as expected. Some weeks ago I read an artical about the special powerdraws of the PCIe port on different gpus. If i remember right different gpu generations draw differen power values on 12V and 3,3V.

I have a mATX board lying around, I will try and test it tomorrow.

The second system doesn't have the powered riser, it is a regular 4x PCIe slot so max 25W, the rest is provided via the PCIe PEG connectors.
I've just sold my GTX970, but it didn't have these issues. It might be specific to the R9 Nano.


If you have a link to the article you mentioned that would be great.
The Toms Hardware review shows very low power consumption on the 3.3V rail, only 3W peak so I'm fairly confident that isn't causing the issue.

You are using FurMark, which is known to cause unrealisticly high loads. It often leads to throttling with more recent cards. I would advise on staying with Heaven benchmark.

If the peaks are well over the sustainable power, a large enough set of condensators might do the trick, but you probably know more about that than I do. The issue could be in the design of the power brick, as it is probably not specced for these kinds of peaks, while ATX PSUs probably do.

I know that Furmark is unrealistic, but I do think the system needs to be able to at least run it for a couple of minutes. the Toms Hardware review shows that because the way Powertune on the R9 Nano works that the power consumption is generally lower in Furmark then in a game.
I'm still baffled at the fact that some resolution work with Furmark and some don't.

For example:
Furmark 1080p 0xAA = Monitor goes blank
Furmark 1080p 2xAA = Monitor goes blank
Furmark 1080p 4xAA = Seems stable after 5 minutes (PowerLimit -50% - Fanspeed Fixed 50% - GPUTemp 37C - Core Clock ~430MHz - Core Voltage ~0.9145V)
Furmark 1080p 8xAA = Seems stable after 8.5 minutes (PowerLimit -50% - Fanspeed Fixed 50% - GPUTemp 38C - Core Clock ~540MHz - Core Voltage ~0.9325V)

Furmark 1080p 0xAA = Monitor goes blank
Furmark 1080p 2xAA = Monitor goes blank
Furmark 1080p 4xAA = Seems stable after 3 minutes (PowerLimit +50% - Fanspeed Fixed 50% - GPUTemp 70C - Core Clock ~908MHz - Core Voltage ~1.07V)
Furmark 1080p 8xAA = Seems stable after 6.5 minutes (PowerLimit +50% - Fanspeed Fixed 50% - GPUTemp 70C - Core Clock ~1000MHz - Core Voltage ~1.19V)

Heaven 1080P Ultra 0xAA = Seems Stable after 15 minutes (PowerLimit +50% - Fanspeed Fixed 50% - GPUTemp 69C - Core Clock ~1000MHz - Core Voltage ~1.19V)

I'm really starting to feel it has something to do with how the GPU is loaded.
Extremely light loads = fine
Extremely heavy loads = fine
moderate loads = fail

Although running Heaven on any setting doesn't cause it to crash.

I'm just not happy with 90% stable it needs to be 100%, especially because the instability happens in a weird way.

@Josh | NFC it might be a bit early to say but you might be getting some questions over on [H] if the powersupply is this critical when people are using the R9 Nano, I didn't have these issues with the GTX970.

Edit:
Guys!

I'm so so so so sorry.

I pulled a PEBKAC
You know 2 posts up I mentioned Heaven at 960x540 only running @ 5 FPS on the ASUS Q87T.
I've been using a program called Prio for years, it allows you to manage you task manager and save the affinity of processes. I had Heaven set to use only 1 CPU core -_-

 

[Phuncz]

No need to be sorry though Stuff like that happens all the time. By the way, are you running those apps full-screen or windowed ? You should run these windowed if not.

I would still recommend trying other test apps to get more perspective on the matter. 3DMark is a good option.

 

[EdZ]

I'm really starting to feel it has something to do with how the GPU is loaded.
Extremely light loads = fine
Extremely heavy loads = fine
moderate loads = fail

If so, that's not a good precedent for future cards, which will be using even more active power management than Maxwell 2 and Fiji.
It does make sense that laptop PSUs may not handle wildly varying power draw. In almost all cases, they're going to be merely charging a battery, from which the laptop will be drawing a varying load. Maybe adding a transient decoupling capacitor to the final output stage (or beefing up the existing one, or adding one after the output on that lovely adapter board) will keep the load stable enough for laptop PSUs to cope with.

 

[QinX]

No need to be sorry though Stuff like that happens all the time. By the way, are you running those apps full-screen or windowed ? You should run these windowed if not.

I would still recommend trying other test apps to get more perspective on the matter. 3DMark is a good option.

I ran 3DMark11 last night, the full suite and when I went back down to the basement to check, system was crashed :x
I've got the Powerlimit set to +50% so I'm going back to stock 0% and see if that changes anything, I suspect not.

If so, that's not a good precedent for future cards, which will be using even more active power management than Maxwell 2 and Fiji.
It does make sense that laptop PSUs may not handle wildly varying power draw. In almost all cases, they're going to be merely charging a battery, from which the laptop will be drawing a varying load. Maybe adding a transient decoupling capacitor to the final output stage (or beefing up the existing one, or adding one after the output on that lovely adapter board) will keep the load stable enough for laptop PSUs to cope with.

They way these cards manage themselves is insane and a testament to good engineering in some way.
If we compare the first 28nm from AMD, the 7970, with the R9 Nano.
Anandtech Bench
The R9 Nano is almost 2x in performance and the power consumption has stayed the same.
Yes the process node has been further optimized and the engineers had a lot of time to tune their designs. But a 2x performance increase at the same power consumption requires some engineering shenanigans. aggressive power management is the major factor here.

I will be looking to add some capacitors in a next revision.

 

[QinX]

I've run 3DMark 11 a couple of times now and it seems to be stable. Haven't tried a round of gaming yet.

3DMark 11 score: P15114
http://www.3dmark.com/3dm11/10716878

The score looks in line with most reviews so no performance penalty. I'm almost completly positive it is GPU/powersupply related. the powersupply is inadequate for what the GPU needs, but the GPU has a pretty idiotic peak power consumption that is hard to deal with.

The PCB works fine as is, but for the R9 Nano it is not there yet, I'll need to add some capacitors to cope with the peak power.

 

[Phuncz]

It seems the most logical that it is the peak power consumption, looking at the results. I'm really interested in this little project as a part of the case, I hope you'll see this through and that it can became a 100% stable solution, I have faith that you can achieve this.

 

[iFreilicht]

Do you still have the powered riser from H2O-micro? As dondan said, the problem might be that the card is trying to draw more than 25W from the PCIe slot, with the powered riser you could easily test that assumption.

 

[QinX]

It seems the most logical that it is the peak power consumption, looking at the results. I'm really interested in this little project as a part of the case, I hope you'll see this through and that it can became a 100% stable solution, I have faith that you can achieve this.

Thank you for your faith in me
I'm still concerned, I'm currently using the HDPlex and it is a very beefy DC-DC adapter. It has to convert to 12V on the PCB so that might be an issue but it has a fairly big capacitor bank. I'm going to contact HDPlex and see if they have a response.

Do you still have the powered riser from H2O-micro? As dondan said, the problem might be that the card is trying to draw more than 25W from the PCIe slot, with the powered riser you could easily test that assumption.

I am running with the powered riser currently, so that should take care of that.

The R9 Nano is getting all 12V from the HDPLEX and the HP 350W Voodoo AC-DC adapter. Not much out there that with more juice, besides industrial powersupplies like the Meanwell enclosed units, but I don't like to recommend those. I might use the one I have here for testing.
a 1000W 24V 1U beast

 

[QinX]

So it is time to bring in the big guns!

Enter the Meanwell RSP-1000-24!
A 1000W 0-24V programmable powersupply
I'm using a Aquaero 5 with voltage regulated out to set the Meanwell to 19V. I will use this to test the HDPLEX and see if it still crashes.
If it doesn't I'll set the Meanwell to 12V and try my PCB.
I don't have time to hook it up today, one of my monitors is calibrating and I've got more Christmas obligations .

I'll post an again update tomorrow!

 

[Josh | NFC]

Go man go!!!

 

[QinX]

I've been running Heaven for 1 hour and all seems stable even with the Nano Powerlimit set to +50%
Powermeter at the wall reports:
Peak Power: 451W
0.17kWh so 170W average power consumption over 1 hour.

The RSP-1000-24 is fairly efficient, its typical rating is 88% for the 24V version.

I'm going to run Furmark next at the previously impossible resolutions and see if the system crashes.
AND it works, I'll leave it running for a little bit to get some data.

I think we are safe to say that running a R9 Nano of a 350W AC-DC Adapter is not enough to guarantee a 100% stable system.

The RSP-1000 has fans that are controlled by powerdraw.
What is funny is that Furmark forces the Nano to change the clockspeed between 700MHz and 880Mhz and you can hear the RSP-1000 rev up and down as the clock and with it the powerdraw changes.

Here are the Furmark runs/data in the same format as a couple of posts back

Furmark 1080p 0xAA = Seems stable after 5 minutes (PowerLimit +50% - Fanspeed Fixed 100% - GPUTemp 63C - Core Clock ~840MHz - Core Voltage ~1.00V)
Furmark 1080p 2xAA = Seems stable after 5 minutes (PowerLimit +50% - Fanspeed Fixed 100% - GPUTemp 63C - Core Clock ~800MHz - Core Voltage ~0.96V)
Furmark 1080p 4xAA = Seems stable after 5 minutes (PowerLimit +50% - Fanspeed Fixed 100% - GPUTemp 67C - Core Clock ~967MHz - Core Voltage ~1.14V)
Furmark 1080p 8xAA = Seems stable after 5 minutes (PowerLimit +50% - Fanspeed Fixed 100% - GPUTemp 64C - Core Clock ~1000MHz - Core Voltage ~1.19V)

I might have some more info:
Why did Furmark crash at 0xAA and 2xAA before but not at 4xAA and 8xAA?

0xAA Peak Power was 457W
2xAA Peak Power was 450W
4xAA Peak Power was 410W
8xAA Peak Power was 315W

Peak Power is higher when using 0xAA and 2xAA. 4xAA drops about 40W and 8xAA a whopping 100W!

Remember the 1024x786 crashing?
Now 614W Peak Power!
Rerunning the test and the highest I've seen now is 542W. It is a very short peak.

Running at a regular Power Limit of 0%:
Furmark 1080p 0xAA (PowerLimit 0% - Fanspeed Auto 40% - GPUTemp 68C - Core Clock 727MHz - Core Voltage ~0.93V)
Furmark 1080p 2xAA (PowerLimit 0% - Fanspeed Auto 40% - GPUTemp 70C - Core Clock ~687MHz - Core Voltage ~0.92V)
Furmark 1080p 4xAA (PowerLimit 0% - Fanspeed Auto 38% - GPUTemp 72C - Core Clock ~796MHz - Core Voltage ~0.98V)
Furmark 1080p 8xAA (PowerLimit 0% - Fanspeed Auto 100% - GPUTemp 64C - Core Clock ~1000MHz - Core Voltage ~1.19V)

0xAA Peak Power was 399W
2xAA Peak Power was 428W
4xAA Peak Power was 375W
8xAA Peak Power was 321W

at 0xAA and 2xAA I consistently see the power meter hit 350W
at 4xAA it is lowered to hit 320W consistently
at 8xAA I see it rarely goes over 270W.

Even more facinating is that the power meter goes from 60W all the way to 350W in seconds, the powermanagement is real guys.
Only at 8xAA does the power meter stay very stable between 200W and 270W, almost never dipping below 200W or going over 270W. This is most likely because it is no lowering clockspeed because of power but temperature reasons.

 

[Phuncz]

Interesting results. Are you going to test more or try to fix it so it works off of the 350W DC adapter ?

 

[QinX]

As I go along I will test more, but I've got most of the information I want. The only measurements I would like to take would require an oscilloscope to get a much more detailed look.

Tomorrow I will try with my PCB just to see if the PCB is 100% okay

I'm doing research how to get the system to work on the 350W AC-DC adapter, but my goal is to get it to work with a 12V 250W AC-DC adapter, because my PCB needs 12V. It could be the R9 Nano is just not suited for uSFF cases, which defeats the entire point of this card.

Remember Project Quantum? I'm beginning to wonder what kind of powersupply they had for it.

 

[Phuncz]

A hefty ATX PSU it seems:

https://smallformfactor.net/forum/threads/amd-project-quantum.99/#post-1338