Kabini iGPU frequency & memory bandwidth scaling

[coercitiv]

In another thread a fellow forum member argued that AM1 IGP is bandwidth starved to such an extent as any increase in iGPU frequency would be rather useless. (example at hand was jump from Kabini to Beema - 600 to 800Mhz) Following that debate I decided to do a bit of testing. Bellow one can find 3DMark Ice Storm scores for a number of iGPU and memory speeds on Athlon 5350.

GPU 600Mhz / RAM 1600 / CPU 2.05Ghz
Graphics Score 36517

GPU 720Mhz / RAM 1600 / CPU 2.04Ghz (BCLK 120)
Graphics Score 38864

GPU 840Mhz / RAM 1493 / CPU 2.1Ghz (BCLK 140)
Graphics Score 39445

GPU 840Mhz / RAM 1866 / CPU 2.1Ghz
Graphics Score 44409

GPU 840Mhz / RAM 2240 / CPU 2.1Ghz
Graphics Score 46756

 

[monstercameron]

wow that is a massive improvement.

 

[frozentundra123456]

It is some of both, but bandwidth seems to have more effect.

gpu frequency 600-840 = 40% increase, gives 8% improvement

gpu frequency at 840 ram 1493-1866 = 25% increase, gives 12.6% improvement.

So it scales better with increasing ram frequency than with igpu speed, so it is more bandwidth limited than gpu frequency limited, although both have an effect.

 

[coercitiv]

gpu frequency at 840 ram 1493-1866 = 25% increase, gives 12.6% improvement.

And another 20% increase (1866-2240) gives only 5%. Coincidentally Beema clocks are 800Mhz GPU / 1866 RAM.

Granted, the 2240 speed was the only one with modified timings (11-11-11) while all others where at 9-9-9.

 

[Shivansps]

it seems to be giving the exact same result as the A4-4000 get at 720mhz with SC, GCN or not those two igp perform exactly the same.

So a dual channel 1333 will probably kick it intro the 50k with just 720mhz.