AMD Trinity Memory Scaling

[Musho]

I've found this interesting page on Trinity's IGP scaling with memory speeds: http://www.tomshardware.com/reviews/...0k,3224-5.html

Does anyone happen to know more articles/tests where they explore memory scaling? I'd like to see more games benchmarked and higher memory speeds reached as well, together with an overclocked IGP if possible. What kind of Memory/GPU/CPU 24/7 stable speeds can you expect on a decent aftermarket air cooler for maximum IGP performance?

[donitsi]

Here is some games with memory scaling from 1333mhz to 2400mhz

http://www.inpai.com.cn/doc/hard/182755.htm

[slaveondope]

Here is some games with memory scaling from 1333mhz to 2400mhz

http://www.inpai.com.cn/doc/hard/182755.htm

20-50% improvement in fps

[Musho]

Here is some games with memory scaling from 1333mhz to 2400mhz

http://www.inpai.com.cn/doc/hard/182755.htm

Wow, thanks for the link! Shows some very impressive scaling! This was done at stock GPU speeds, I assume? I wonder how much further that IGP can be pushed with higher GPU speeds! Do you happen to know if that 2400mhz memory was 24/7 stable?

[The Stilt]

With a correct configuration Trinity can do DDR-2400-2500 without much of effort (24/7).
The correct configuration: = Samsung D-die modules, one module per channel (i.e. two modules in total).

With Hynix it is also doable, but much harder.

Most of the boards currently configure the CNB in the wrong way.
It can be reconfigured manually, however the mileage may still vary.

E.G. Fused NBPState0 voltage 1.16250V, fused SCLKPState2 (3D) voltage 1,27500V.
The 'DDR-2400' fuse voltage is usually 0,0125V higher than the NBPState0 voltage, on production chips.
In this case the minimum rated voltage for DDR-2400 MEMCLK frequency would be 1,17500V.

However, since the Trinity APUs were never designed for such MEMCLK frequencies (DDR-1866 is the highest officially supported frequency), the other two NBPStates are configured to frequencies which will violate the NCLK:MEMCLK ratio (MEMCLK/2)*1,25 = Minimum NCLK frequency).

On A10-5800K parts the NBPStates are configured: PState0 = 1800MHz, PState1 = 1500MHz, PState2 = 1400MHz, PState3 = 800MHz. Since 1200*1,25 = 1500 the NBPState2 & NBPState3 will violate the frequency rule with default configuration. Also the voltage set by these PStates is too low for DDR-2400.

AND because Trinity is highly optimized for power efficiency, there is an additional -25mV loadline offset set for the UNB voltage.
With the voltage levels used in the given example the resulting UNB voltage would be 1,13750V (2D) and 1,25000V (during 3D).
In 2D the UNB would only have 1,13750V instead of the 1,17500V specified by AMD.

Currently Asus is the only one who configures the UNB voltage correctly based on the selected MEMCLK frequency
To make it happen the VRM controller must be set it "Manual" -mode and the voltage option should be left to "Auto".

Because Trinity has the CNB/GNB and IMC in the same package (UNB) all of the frequencies are related when overclocking.
E.G. High MEMCLK frequency puts more strain on CNB and GNB, higher CNB puts more strain on the IMC and GNB, and so on.

It is very hard to find the optimal frequencies for all of the three domains. They cannot be tested individually because a previously stable CNB frequency might come unstable when GNB or memory is overclocked, etc.

[FlanK3r]

so, with 2400 MHz RAM is min unclk 1500 MHz? FX had MEMCLK/2)x 2 = minimum unclock CPU/NB

[cezar]

Here is some games with memory scaling from 1333mhz to 2400mhz

http://www.inpai.com.cn/doc/hard/182755.htm

Actually it's taken from Xbit Labs (as it's said in the article):
http://www.xbitlabs.com/articles/gra...raphics_8.html

[The Stilt]

It seems DDR-2600+ is very much doable for benching, even with an 'average' IMC and air cooling.
The timings need to be heavily recalibrated thou... and none of the most popular motherboards have these adjustments available in bios.
I have requested these to be added, however none of the vendors have reacted in any way (as usual).

Fortunately I am making a small tool (TCI K² Loader) which will latch these timings so that you can boot into Windows without crashing
No user interaction required, the system only needs to be able to post and you will need to have a spare USB-stick.

Testing the actual stability of overclocked memory on Trinity is the funny part (not really).
With memory testing tools (such as Memtest86/+, HCI Memtest, R.S.T, Linpack, Prime95 Blend, etc) you will only be able to tell if the actual DRAM modules are failing. This tells nothing about the actual stability (CNB/IMC). These tests will not catch the CNB/IMC related errors unless it is absolutely melting down.

Since the GNB and CNB/IMC are a single unit (UNB), stressing GNB will also stress the CNB/IMC.
Just like on Llano the ultimate stability test for overclocked memory (and / or CNB) is 3DMark06 GT-1 (Return to Proxycon).
DRAM module related unstability is immediately visible as artifacts and any unstability in IMC/CNB will halt or crash the program (/system) usually within the first 20 seconds of the test. If the system passes 10 loops of GT-1 it will never crash in any other 3D app either.

[FlanK3r]

Stilt: how do you mean? Example 2650 MHz and 12-14-14? Or subtimings and drivings too?

[The Stilt]

Stilt: how do you mean? Example 2650 MHz and 12-14-14? Or subtimings and drivings too?

DRAM timings doesn´t matter really, it is up to the DRAM (e.g. 10-12-11-xx is doable if your DRAM is up to the job).

The timings in question are mostly related to IMC. In terms of the overclocking capabilities of the DRAM modules themselves, they do not have any major effect.

These timings have a MASSIVE effect to the bandwidth, when they are fully latched the bandwidth is around 8-9GB/s lower
I am currently trying to balance the overclocking capability and bandwidth.

I ran Prime95 at DDR-2666 at one point, but the bandwidth was awful since the timings were latched too much.

[Alex-Ro]

Oh,this is something like the tertiary timings on ivy bridge,read to read same rank,write to read,write to write etc....?

[The Stilt]

Getting the MEMCLK up to DDR-2600+ even on air cooling is easy...
However finding the fastest stable IMC timings at high NCLK (taking the high MEMCLK in account) is... something else.

DDR-2400 9-11-11-xx with 2GHz NCLK can outperform DDR-2600 10-12-11-xx unless the NCLK and IMC timings are optimal.
The NCLK needs to be raised to 2.2GHz+ before it can fully saturate the DRAM at DDR-2600.
Naturally a higher NCLK frequency puts more stress on the IMC too

Had to use basically all of the cards from my sleeves to make this 3D stable.
Cooling has a massive effect on IMC overclockability on all K12h & K15h chips.
Need to try with something colder next time, this was the max at air cooling:

DDR-2600 can keep the GPU fully saturated (below 100%) up to 980-1050MHz depending on the test

[The Stilt]

Lower density modules put much less strain on the IMC...
Good old PSC. VapoLS cooled, max MEMCLK only.



1400MHz fails in the middle of GT-1 once the evaporator raises above -45c.

Soon...

[BeepBeep2]

Was getting a little worried there haha

I don't use that slow "lets do 2800 CL11" Sammy/Hynix XD

[Yeroon]

Um, guys how do if find this impressive samsung memory? I just swapped the watercooling kit off the llano (testing for temp difference to make sure my modded kit works as intended) and since my air cooler (scythe big shurk) isn't fond of ram sinks, I pulled them off (again my swapped back to the llano gskill 1866 ripjaws. Got it all together (pain in the ass, the mobo has to come out to put that cooler on with the little cheap case) get started folding and start to clean up, when I cam across this.
ram module.jpg
So, I'm pretty sure I'm in the market for a new pair of memory sticks. I can either get new mem for llano in the 2x2gb semi-cheap ram, or get some new sticks for trinity, and put the 2133 mushkin hynix ram back into llano.

Thoughts?

Yeroon

Edit, bios shows both as 4096, sisandra mem benchmark puts it as 2x2gb. Not sure i need to spend the money on new ram yet...