A small warning regarding 100W APUs, stock cooler, small cases and highspeed RAM

[Kanadian]

Most likely the reason for the high temps, at least in my experience is that you were sitting in the bios doing things. I have had this happen to me in the past where I've been in the bios (I have an MSI board with an A10-5700) and the temps have gone through the roof.
Now when I go into the bios, it's get in quick and make the changes required. Running in windows isn't an issue, it's just in the bios where temps go crazy.

 

[AtenRa]

Forgot...

Gigabyte GA-F2A85XN-WIFI with BIOS F3a.

Since the F3a is a BETA, i would return to F2 and see if it works ok. The F2 supports the A10-6800K so you must be ok, check if you will have the same problem with the ram.

 

[Insert_Nickname]

Ok, that's weird. You should check cpu frequencies and voltage, I'm suspecting some kind of automatic overclocking feature.

Since the F3a is a BETA, i would return to F2 and see if it works ok. The F2 supports the A10-6800K so you must be ok, check if you will have the same problem with the ram.

Working through some combinations currently. The core issue seems to be a series of unfortunate factors, combining cooler location, RAM voltage and height of the DIMM heatsinks, airflow and unfortunate ATX socket placement creating a hot air pocket. The stock cooler can keep temperatures down below 70C with the case lid off.

There is however a very pronounced correlation between RAM voltage and power consumption (using HWinfo as it can read data off the IR3567s used on the board, with Aida64 as backup and a power meter for tertiary information). I'm seeing a massive increase in power consumption with 1.65V vs 1.5V at the same (2133MHz) speed. We're talking from 90-95W max with 1.5V to 120W+ spikes with the RAM at 1.65V. Apparently just enough to overwhelm the stock cooler. The motherboard had not configured the RAM properly, sending a scorching 1.65V through. I should have remembered to check everything manually... (so much for plug'n'set-XMP)

Another thing I found is that it gets completely crazy with 1.35V RAM at 1600MHz vs 1.65V RAM at 2133MHz. Changing brought a 30W(!) reduction in measured power consumption, and both temperatures suddenly became noticeably lower idling at below 30C reported.

I'm surprised no one has stumbled on this before, I have been using some google-fu and found nothing close. If some of you have, would you please post a link?

So the point must be, that if you want to run 100W APUs with highspeed (>1866MHz at 1.5V) you should check if your cooling and airflow are up to it. I'm limited in my case choice to low-profile ones due to placement.

I'm thinking of dumping this kit in my main system (works great) as an "unplanned semi-"upgrade"", and getting some low-profile Crucial 1.35V DDR3-1600. The old kit from the HTPC can be put to good use elsewhere. The performance increase going from 1600MHz 9-9-9-27 to 2133MHz 11-11-11-30 is only around 200MB/s according to Aida64 (~21000MB/s to ~21200MB/s Read and ~9900MB/s to 10500MB/s Write).

Oh, and thanks to everyone for your input... :thumbsup:

 

[William Gaatjes]

How can a 533MHz (1600 - 2133) increase yield such a small gain in bandwidth ?
Is there some bottleneck for the cpu part of the apu or is that just simply the maximum bandwidth for the cpu part of the apu ?

 

[William Gaatjes]

I am happy i did not increase my ram voltage to 1.65V. I have it also on 1.5V.

 

[Insert_Nickname]

How can a 533MHz (1600 - 2133) increase yield such a small gain in bandwidth ?
Is there some bottleneck for the cpu part of the apu or is that just simply the maximum bandwidth for the cpu part of the apu ?

Could be, but honestly I have no idea. IGP performance seems the same, its within the usual 2% difference between benchmark runs. Someone more knowledgeable will have to clarify.

But it seems the memory controller likes low timings more then raw bandwidth.

 

[William Gaatjes]

If i understand correctly, the maximum system bandwidth is 2x the bandwidth of the memory module because of dual channels. For the A10-6800 at 2133MHz that would be around 34GB/s according to a tomshardware article i found. Perhaps the cpu cannot retreive data any faster. It would make sense for me that a memory bandwidth hungry gpu would benefit more from the increased memory transfer then the cpu that might be already saturated. If this is the case, it would show up in games and gpu benchmarks.


http://www.tomshardware.com/reviews/a10-6700-a10-6800k-richland-review,3528.html

Again, the top-end A10-6800K gets 2133 MT/s memory support, which is particularly meaningful for its on-die graphics engine given a lack of shared L3 cache. The theoretical 34 GB/s of DDR3 bandwidth should go a long way to improve frame rates in the games we'll be testing.

 

[William Gaatjes]

Working through some combinations currently. The core issue seems to be a series of unfortunate factors, combining cooler location, RAM voltage and height of the DIMM heatsinks, airflow and unfortunate ATX socket placement creating a hot air pocket. The stock cooler can keep temperatures down below 70C with the case lid off.

There is however a very pronounced correlation between RAM voltage and power consumption (using HWinfo as it can read data off the IR3567s used on the board, with Aida64 as backup and a power meter for tertiary information). I'm seeing a massive increase in power consumption with 1.65V vs 1.5V at the same (2133MHz) speed. We're talking from 90-95W max with 1.5V to 120W+ spikes with the RAM at 1.65V. Apparently just enough to overwhelm the stock cooler. The motherboard had not configured the RAM properly, sending a scorching 1.65V through. I should have remembered to check everything manually... (so much for plug'n'set-XMP)

Another thing I found is that it gets completely crazy with 1.35V RAM at 1600MHz vs 1.65V RAM at 2133MHz. Changing brought a 30W(!) reduction in measured power consumption, and both temperatures suddenly became noticeably lower idling at below 30C reported.

I'm surprised no one has stumbled on this before, I have been using some google-fu and found nothing close. If some of you have, would you please post a link?

So the point must be, that if you want to run 100W APUs with highspeed (>1866MHz at 1.5V) you should check if your cooling and airflow are up to it. I'm limited in my case choice to low-profile ones due to placement.

I'm thinking of dumping this kit in my main system (works great) as an "unplanned semi-"upgrade"", and getting some low-profile Crucial 1.35V DDR3-1600. The old kit from the HTPC can be put to good use elsewhere. The performance increase going from 1600MHz 9-9-9-27 to 2133MHz 11-11-11-30 is only around 200MB/s according to Aida64 (~21000MB/s to ~21200MB/s Read and ~9900MB/s to 10500MB/s Write).

Oh, and thanks to everyone for your input... :thumbsup:

That is amazing and very interesting. I wonder how much an A10-6700 45W version would consume with 1.35V DDR3.

I also did the aida x64 memory benchmark.
I used Aida x64 engineer version v400.2700.

Read benchmark = 18854 MB/S
Write benchmark = 9429 MB/S
copy benchmark = 16887 MB/S
latency = 63.4ns

My memory timings are 1866MHz,10-10-10-26-1T.

 

[inf64]

That is amazing and very interesting. I wonder how much an A10-6700 45W version would consume with 1.35V DDR3.

I also did the aida x64 memory benchmark.
I used Aida x64 engineer version v400.2700.

Read benchmark = 18854 MB/S
Write benchmark = 9429 MB/S
copy benchmark = 16887 MB/S
latency = 63.4ns

My memory timings are 1866MHz,10-10-10-26-1T.

At what clock is your A10 running?

Here is my cachebench in Aida:

 

[William Gaatjes]

At what clock is your A10 running?

Here is my cachebench in Aida:

That is a good question.
I have cool en quit enabled.
Minimum clock is 1800MHz. Maximum clock is 4300MHz.
I assume my clock speed will go up since i noticed that the "northbridge" clock(NB frequency) also changes from 1500MHz to 1600MHz in cpu-z while testing with prime95.
I assume that the clock is increased to maximum but i am not sure.
I do not know how to check that under AIDA64 al clocks are at maximum.

I should also note that the AIDA64 program notifies during benchmarking that the results from different AIDA64 builds should not be used to compare results.
This because the benchmarking algorithms are constantly improved.
I downloaded the AIDA64 engineer trial version.

 

[inf64]

I use the same built as you (v400.2700). It's odd my memory BW scores are better than yours but I do run NB @ 2.2Ghz and cas latency is 9 (mem clock is 1600Mhz).

 

[William Gaatjes]

I noticed that in AIDA is says dual channel.
Do i need to change my memory settings to unganged or something like that ?
To be honest i have not found it in the bios settings.

 

[inf64]

Nope, unganged is actually better than ganged (for most workloads). If it says dual channel like in the image above I took, then you are set.

 

[William Gaatjes]

I use the same built as you (v400.2700). It's odd my memory BW scores are better than yours but I do run NB @ 2.2Ghz and cas latency is 9 (mem clock is 1600Mhz).

That is indeed odd. And you have a different picture. I did not find any cache testing. Do you use the engineer version from AIDA ?

That NB clock might explain the higher results but that makes me wonder if the A10-6800 from insert_nickname also runs at a higher NB clock than my A10-6700.

 

[inf64]

Go to Tools drop down menu (top right in the main window) and select Cache and Memory benchmark.

 

[William Gaatjes]

Go to Tools drop down menu (top right in the main window) and select Cache and Memory benchmark.

Ah thank you. I forgot to change the language settings to english. Makes it easier to work with.
I have the trial version so unfortunately some fields are blank.

Maybe for me it is no use to run the memory at 1833MHz because the northbridge frequency is too low to have any benefits ?
I will have to see if i can change it.

 

[William Gaatjes]

I did a test with my memory back at 1600MHz with the same 10-10-10-26-CR1
settings

Read - 19146MB/s
Write - 9491 MB/s
Copy - 16742 MB/s
latency - 67ns

Then i reduced the timings to 9-9-9-24 but with command rate at CR2.

I have the same issue. In the AIDA64 tests, there is no gain in speed when increasing from 1600MHz to 1833MHz.
No, it seems i even gain memory bandwidth at 1600MHz.
Since my multiplier is locked, i cannot adjust it. In the bios the maximum NB (northbridge) frequency is 1600MHz.

Perhaps this is the case, i am just guessing :
When the memory is not running at 1600MHz but higher, it seems the NB logic circuits needs to inject more clock cycles to keep everything synchronized.
That might explain why at a higher frequency the bandwidth seems to reduce. The only other thing i noticed, is that latency is lower at 1866MHz.

I might as well keep the memory running at 1600MHz and see if i can reduce the timings a bit more.

 

[William Gaatjes]

I do wonder what effect this might have on the IGPU performance ?:hmm:

 

[Insert_Nickname]

That NB clock might explain the higher results but that makes me wonder if the A10-6800 from insert_nickname also runs at a higher NB clock than my A10-6700.

That's correct, the 6800K runs the northbridge at 1800MHz standard. I haven't really played around with that though, so I don't how much it affects performance.

I also did the aida x64 memory benchmark.
I used Aida x64 engineer version v400.2700.

Read benchmark = 18854 MB/S
Write benchmark = 9429 MB/S
copy benchmark = 16887 MB/S
latency = 63.4ns

My memory timings are 1866MHz,10-10-10-26-1T.

Very interesting. This kind of ties in with my own test. AMDs memory controller seems to like tight timings, more then -just- raw bandwidth. I'd post a few screen-shots but they are on the HTPC currently...

Any chance you could test again with 11-11-11-30 CR1...?

BTW, I'm using v4.00.2743. Earlier versions don't play nice with my board.

Here is my cachebench in Aida:

Now that is strange. You have a monster latency on the L2 cache, mine is around 9ns. Could raising the NB clock have such a dramatic effect on L2 latency?
----------------------------
If you're only running two DIMMs, try lowering CR to 1. That ought to give a small increase in memory performance. Or isn't that stable?

 

[inf64]

I'm not sure what is going on with L2 latency. I will check it out.

edit- just rerun the test again and all is normal this time. It appears I had some background tasks active which stole some cycles from the CPU

 

[William Gaatjes]

I did the tests with memory settings at 1866MHz, 11-11-11-30,CR1.
I noticed a drop again in benchmark results.

I did the separate benchmarks and the combined benchmarks :
Read - 18782 MB/s
Write - 9358MB/s
Copy - 14279 MB/s
Latency - 66.5 ns

EDIT:
I think the reason why the A10-6700 has a 1600MHz NB frequency is for the 65W TDP labeling. And to give the price difference for the A10-6800 running NB@1800MHz more justification.
So the Richland apu's could gain a lot more in speed if the NB frequency would be increased dramatically.
Since the GPU also have to access the memory through the NB , i would say it makes no difference and is 1866MHz no use for an A10-6700.

 

[William Gaatjes]

I forgot to note that it seems that reducing the command rate has an effect on latency. I have the lowest latency @1866MHz with CR1.

 

[William Gaatjes]

I did a full memtest86 4.3.2. to test the memory at 1600MHz with settings 9-9-9-24 CR1 and it passed. ^_^

EDIT:

Insert_Nickname,
I wonder how much of a memory speed gain you would have if you could set the NB frequency at exact 1866MHz or the memory clock at exact 1800MHz.

 

[Insert_Nickname]

I think the reason why the A10-6700 has a 1600MHz NB frequency is for the 65W TDP labeling. And to give the price difference for the A10-6800 running NB@1800MHz more justification.
So the Richland apu's could gain a lot more in speed if the NB frequency would be increased dramatically.
Since the GPU also have to access the memory through the NB , i would say it makes no difference and is 1866MHz no use for an A10-6700.

It seems to have more effect on write rather then read speeds. I'm pushing slightly over 10800MB/s with 9-9-9-27 CR1 @ 1600MHz and NB @ 1800MHz. The "funny" bit is that I have higher write speeds then with the memory at 2133MHz (10500MB/s) and only slightly lower read speed...

I must say that 1.35V Crucial low-profile 1600MHz with 9-9-9-24 timings is looking very tempting right now. Should be an excellent compromise between performance and power consumption...

 

[coffeejunkee]

Did you check cpu frequencies and vcore? It seems highly unlikely to me that overclocking ram increases power draw by 30W. I don't think the ram slots can even deliver that much.