Killer Trinity - top A10-5800k - confirmed 5.1GHz on air and 7.3GHz on LN2

[BeepBeep2]

Trinity unofficially supports DDR-2400 DRAM ratio.
Currently it is available on some of the Asus and Gigabyte motherboards atleast.

Does it seem like they have made improvements in the IMC, DRAM frequency wise?

[The Stilt]

Does it seem like they have made improvements in the IMC, DRAM frequency wise?

There are some changes, however the CNB/IMC is very similar to K12h and other K15h chips.

[leo5111]

nice results but at that voltage i wouldnt sell that CPU any life insurance

[Alex-Ro]

Do you have a ssd to test A85X chipset?Hope to see some improvements compared to 990FX and A75....

[imamage]

So the NDA Ended ?!?!

[Miwo]

Gotta leave for work, but I did some really quick SuperPi 1M on my A6-3670k:

PcCI2iminal A10-5800k @ 4GHz , DDR3-1866 @ 9-10-10-33 : 23.993s
A6-3670k @ 3.1GHz , DDR3-1866 @ 9-10-10-33: 23.596s
A6-3670k @ 3.0GHz , DDR3-1866 @ 9-10-10-33: 24.2s~

So in terms of just SuperPi (which is pretty bad and useless benchmark for AMD), Trinity needs a 30%~ higher clock than comparable Llano. I was really hoping for better , but i guess that is expected.

[FlanK3r]

yes, because its about x87 optimizing and Bulldozer was not optimized for this old x87 code. Still, max OC for Llano is average around 3500-3600 MHz stable, Trinity 4500-4700 MHz.

[qcmadness]

Gotta leave for work, but I did some really quick SuperPi 1M on my A6-3670k:

PcCI2iminal A10-5800k @ 4GHz , DDR3-1866 @ 9-10-10-33 : 23.993s
A6-3670k @ 3.1GHz , DDR3-1866 @ 9-10-10-33: 23.596s
A6-3670k @ 3.0GHz , DDR3-1866 @ 9-10-10-33: 24.2s~

So in terms of just SuperPi (which is pretty bad and useless benchmark for AMD), Trinity needs a 30%~ higher clock than comparable Llano. I was really hoping for better , but i guess that is expected.

x87 is relatively weak for Bulldozer derivatives.

[informal]

x87 is irrelevant,especially for client market.

[ZenEffect]

x87 is irrelevant,especially for client market.

not when physx is run from x87

[informal]

Physx and AMD's Radeon GPU in Trinity?

[BeepBeep2]

Software PhysX, CPU Physx, x87 Physx derp

[Miwo]

I thought that Nvidia fixed x87 PhysX since version 3... to run on multicore or at least thats what they claim.
That whole x87 physx thing was like 2 years ago

[The Stilt]

Made some power measurements while the APU was undervoltaged.
The maximum error of margin is around 2.5% according to the component supplier.
Even the actual power measurements might be off by this percentage, the measured values have been constant under the same settings.
Therefore the measured difference in the power consumption is accurate.

Again, please note that the parts I am using DO NOT fully represent a retail product.

Specs:
A10-5800K, a high-leak part (base voltage 1.35000V), 3.8GHz (CPB disabled).
NOTE: Loadline configuration follows the exact specification of AMD.

Load: Prime95 V27.7 Small FFT, power measurements made at the exactly same point of the calculations.

3.8GHz - 1.35000V = 63.688W
3.8GHz - 1.32500V = 60.000W
3.8GHz - 1.30000V = 56.500W
3.8GHz - 1.27500V = 53.625W
3.8GHz - 1.25000V = 50.750W
3.8GHz - 1.22500V = 48.000W

So the CPU power consumption dropped by 15.688W or 25%

Please note that these values are PEAK values, the average is usually lower by 2-3W.
1.225V was the lowest fully stable voltage.

If the loadline would be adjusted a bit it would probably be stable with 1.20000V too.
With standard specifications the loadline slope is very steep.

I would imagine the GPU could be undervoltaged in similar manner.
The default voltage for the GPU is quite high, so I recon there is atleast 10-20% which could be "shaved" off.
Around 10W at the most.

[informal]

Nobody cares about Physx,it's just a gimmick...

[ZenEffect]

Nobody cares about Physx,it's just a gimmick...

in the real world, with real consumers, quite a few people care about physx.... quite a few as in people numbering in the millions.

[dess]

The NB clock is a bit low by default (1800 MHz), isn't it? Could you please lift it as well as core clocks? I think it could significantly affect overall performance.

[The Stilt]

The NB clock is a bit low by default (1800 MHz), isn't it? Could you please lift it as well as core clocks? I think it could significantly affect overall performance.

1800MHz is the highest stock frequency for Trinity.
Actually, under the load the frequency is usually between 1500-1800MHz.
Raising the NCLK helps, however the difference is not that massive.

[dess]

1800MHz is the highest stock frequency for Trinity.
Actually, under the load the frequency is usually between 1500-1800MHz.

I was comparing it with the NB clocks of the FX'es (2000+MHz at stock and ~2600MHz OC'd). Interesting it's only up to 1800MHz at stock for Trinity, as there is a bandwidth hungry IGP here, as well. Athough, there is no L3 here, but the IMC performance could also be affected, I would think.

Raising the NCLK helps, however the difference is not that massive.

I see, thanks! But... How much not that massive, still?

[The Stilt]

This was the highest I could do on Trinity.
The motherboards were at the alpha stage (at best) when the tests took place.
No power management features or protections could be disabled.
Currently the situation is very different, luckily.

http://valid.canardpc.com/show_oc.php?id=2443149

Cinebench_6.5GHz.jpg

Entertained.jpg

The Cinebench score is badly held back by the low MEMCLK and NCLK.
It was nothing but a maximum CPU frequency test, so all of the other frequencies were latched as much as possible.

With proper NCLK and MEMCLK frequencies the score should be around 5.6pts or higher.
The voltage is also a bit high for those clocks since the VRM controller of the motherboard had a massive cold bug (the voltage setting got stuck). At that point there were no proper motherboards available.

The chip that did 7.32GHz with all cores is the golden one.
Both of the CUs clock identically, which is really rare for an APU.
Usually the first CU is weaker than the second one, due the structure of the APU.

The quality of the parts has increased very rapidly since the beginning.
So I definitely expect to see more

btw. High leak is good now

[Oese]

thanks for the numbers.

what do you people think about that power consumption? I understand right, its directly measured between cpu and vr? Or is it at the outlet or between vr and psu?

how does these numbers compare? i dont have any idea with what to compare... Most interesting would be a comparison with FX, but there is L3 and no GPU...

A very simple calculation is to take these number by x2 to get an idea about FX 8xxx 4 module part... 127w @ 3.8 ghz seems high...

You did any power measurements at like 4.5 to 4.8 ghz to compare?

[The Stilt]

thanks for the numbers.

what do you people think about that power consumption? I understand right, its directly measured between cpu and vr? Or is it at the outlet or between vr and psu?

how does these numbers compare? i dont have any idea with what to compare... Most interesting would be a comparison with FX, but there is L3 and no GPU...

A very simple calculation is to take these number by x2 to get an idea about FX 8xxx 4 module part... 127w @ 3.8 ghz seems high...

You did any power measurements at like 4.5 to 4.8 ghz to compare?

The measurements are DCR indeed.

The structure of Trinity is much more complex, so it cannot be directly compared to anything else (besides Llano).

Also the first compute unit consumes more power than the rest, so you cannot multiply the power consumption by two.
I have not measured the consumption while the "slave" CU is powered off so the exact difference between a master and a slave unit is unknown.

It cannot be measured very accurately either since when the slave CU is turned on, the thermal dissipation increases and so does leakage.

I got "some" data available, however I will chose not to publish it.
I can give you a single number thou, Idd Max: 69A @ 1.652V - 6.3GHz during Cinebench.
Thats with the minimum leakage current, so you can take a guess what the temperature was

[Oese]

Hehe some subzero i might guess -60°C.. Cannot compare this to normal air/water oc but it doesnt sound too much :p

I read some reviews that power improvements, even if measured across plattforms, become visible at around 20% compared 5800k vs 4170 with deactivated GPU. Here's the l3 and probably some 5 watts more power hungry platform, but this still seems to be relatively nice at stock clocks/voltage.

[FlanK3r]

nice Stilt, u have now time for fun ...And Im thinking, not only with Trinity ,-)...Vishera is soon.

[The Stilt]

Did some power measurements with the 24/7 settings I use.
The chip can do 4.7G Prime95 stable but the extra power required for those clocks cannot be justified anymore.
Also the voltage would be higher than the GF 32nm process can safely handle with the resulting current draw.

I swapped the usual Asus F2A85-M PRO to Gigabyte F2A85X-UP4 to illustrate the power consumption.
The VRM controller on Asus is also able to output the same data, however Asus sets it to "knock, knock who´s there" -mode (PW protection) so there is no easy way to read the data on Asus.

The voltage is a bit high since I used 50% loadline setting to prevent overshooting with lower loads (lower current draw).

This chip had the highest leakage current before the 'tests'.
It took almost 20 hours worth of beating under LN2 with voltages topping out at 1.9V+.
Yet the only obvious difference is a slight (~3W) increase in power consumption with identical conditions.

These pictures show quite clearly why Prime95 might_not_be the best stability for daily use.
Prime is only good for the absolute worst case of scenario testing, no other application come even close to the power draw it generates.



Reference Pmax on the same chip at 3.8G: