Unlocked AM1 APUs... why not?

[Shivansps]

Its not mine...

Ive tested myselft, its easy to get to 140 fsb whiout extra devices attached to PCI-E, and it also works with dgpus.

The point is, there is a wall at about 2300mhz where vcore increase is needed, and there is another big wall at about 2800mhz where big vcore jumps are needed, so a unlocked AM1 is just no good considering you can get to 2200mhz in a 5150 and 2700mhz in a 5350, everything after that just needs crazy vcore.

 

[Abwx]

You can overclock a lot on AM1 if the board supports OC and CPU vcore increase, just disable AHCI and you can go very high.

But that destryo the "25W" thing and im not sure how long VRMs will hold.

Actualy at those settings it should be at about 25W in Cinebench and at 30W in Prime 95, voltage scaling with frequency seems quite good despite the slower transistors than Kaveri s or Carrizo s.

EDIT: 1.632V, HOLY D:

That s not that dramatic as GF slower (and high Vth, aka voltage threshold) transistors (Low Power) can withstand higher voltages than regular 28nm HPP ones.

 

[superstition]

That s not that dramatic as GF slower (and high Vth, aka voltage threshold) transistors (Low Power) can withstand higher voltages than regular 28nm HPP ones.

Higher meaning what? And for how long under what conditions?

 

[Abwx]

Higher meaning what? And for how long under what conditions?

I dont have explicit numbers for this process, IIRC GF s 32nm SOI max safe voltage is in the 1.6-1.65V range, a low power 28nm process like the one used for Kabini should withstand this voltage permanently as well.

That said i wouldnt advise doing such extreme overclocking on a permanent basis, surely that it can be fun but it wont provide that much better perfs, most is gained at about 2.66GHz at wich point the increased RAM and NB frequency result in Integer IPC increasing notably, particularly on 7zip for wich the MT IPC improvement is 30% in compression and 10% in decompression.

 

[The Stilt]

That said i wouldnt advise doing such extreme overclocking on a permanent basis, surely that it can be fun but it wont provide that much better perfs, most is gained at about 2.66GHz at wich point the increased RAM and NB frequency result in Integer IPC increasing notably, particularly on 7zip for wich the MT IPC improvement is 30% in compression and 10% in decompression.

1.6 - 1.65V isn´t safe, 1.61V is the rated brake down voltage for AMD 32nm SHP SOI parts (excl. Llano). Llano (and first Bulldozer silicon) had huge issues with the process, so they will die with voltages below the normal rated BDV.

Right, so increasing memory bandwidth and latency increases the core IPC

7Zip shows gains because LZMA is bandwidth and latency critical. It scales almost linearily with the memory bandwidth and decreasing latency. The same effect can be replicated with a PI program which uses either arithmetic-geometric mean / Gauss-Legendre algorithm for computing the decimals. Extremely fast way to calculate PI, but also extremely memory and latency intensive, which makes it slow.

Go and try with X264 or some other application which isn´t bandwidth and latency critical. The difference will fall within the error margin.

 

[Abwx]

Right, so increasing memory bandwidth and latency increases the core IPC

7Zip shows gains because LZMA is bandwidth and latency critical. It scales almost linearily with the memory bandwidth and decreasing latency. The same effect can be replicated with a PI program which uses either arithmetic-geometric mean / Gauss-Legendre algorithm for computing the decimals. Extremely fast way to calculate PI, but also extremely memory and latency intensive, which makes it slow.

Go and try with X264 or some other application which isn´t bandwidth and latency critical. The difference will fall within the error margin.

You should better read what i wrote, i said the MT IPC, if you have not noticed Kabini lose quite some IPC when the 4 cores are used and that the soft is RAM bandwith sensitive, so keep your condescendency, and maths explanations that are useless in this case, for yourself..

Some numbers, do the maths, it will be usefull this time :

http://forum.hardware.fr/hfr/Hardware/Processeur/unique-basses-temash-sujet_887276_50.htm

 

[MrTeal]

Did you make that chart? If so, scale the Kraken or move it to a secondary axis and then switch back to a linear scale. The numbers are quite clear but the chart is useless as is.

 

[The Stilt]

Still, performance improvements achieved by removing bottle necks has nothing to do with the "IPC increasing". The biggest single bottle neck in Jaguar and Puma is the L2 cache. For a design which operates at < 2.4GHz, > 25 cycle L2 cache latency is completely unacceptable.

 

[Abwx]

Did you make that chart? If so, scale the Kraken or move it to a secondary axis and then switch back to a linear scale. The numbers are quite clear but the chart is useless as is.

The chart is a HFR s member s one, i guess that he s not a marketing guy..

Still, performance improvements achieved by removing bottle necks has nothing to do with the "IPC increasing". The biggest single bottle neck in Jaguar and Puma is the L2 cache. For a design which operates at < 2.4GHz, > 25 cycle L2 cache latency is completely unacceptable.

Each bottleneck add its losses and the end result is the effective IPC, that it can be improved more or less easily is another matter.

Surely that they intended to fix the issues you re pointing in future iterations, but as it is Kabini is more than good enough in respect of the competition, too bad that an updated chip with dual channel and new UVD wont be released so far as the projected replacement was really very promising.

 

[FlanK3r]

I got in stable OC max 131MHz BCLK
http://pctuning.tyden.cz/hardware/p...1-athlon-x4-5150-a-deska-asus-am1i-a?start=13

Validation and superpi 1M was possible up to 148 MHz

 

[SPBHM]

they should probably unlock at least the top X4 model and some X2 model,
but the motherboards for the AM1 platform are so weak that this is probably not a good idea, and with the best AM1 boards you can already do a lot of bclk OC.