Estimating Turbo Core's operation and aggressiveness

From the previous page, we saw that the improved Llano architecture is good for a ~8% advantage over the plain Stars architecture in Cinebench R10 single-thread test, while in multi-threads test the advantage remain at a quite interesting ~7,5%.

With these numbers in mind, we will try to compare the recorded mobile A8-3500M Cinebench score with a simulated, Turbo-less A8-3500 chip. In this manner, we should be able to measure the (lack of) effectiveness of Turbo core technology:

A8-3500M Llano cinebench benchmark

Look at the graph: starting from a mobile PhenomII P920 (quad core @ 1.6 GHz), I've added ~8% performance to obtain a simulated A8-3500M @ 1.6 GHz with no Turbo core. Comparing the estimated, Turbo-less A8-3500M scores with the real score, we can isolate the frequency gain due to Turbo core intervention. It seems that in single-thread Cinebench R10 test Turbo core technology is good for an ~15% improvement, for a final clock at about 1.73 GHz. This is nothing to sneeze at, but far less then the max 2.4 GHz (60% improvement) published for this APU version.

Lets see if Turbo core come into play in the multi-thread Cinebench R10 test:

A8-3500M Llano cinebench benchmark

It seems that Turbo core can do its magic when four core are loaded also: the result shows a ~9% advantage here.

One question is very important: can we trust these results? As they are based on estimation (starting from a P920, increasing its scores by a certain amount and then comparing the projected results with the real ones), can we trust the graph above? I think that the answer is “yes”. To check this, lets try to estimate Turbo Core intervention using another method: starting by the Turbo-less desktop A8-3850 2.9 GHz scores and scaling back it to the mobile A8-3500M 1.5 GHz base clock.



A8-3850 @ 2.9 GHz

A8-3850 @ 1.5 GHz

A8-3500M Real

Turbo vs base frequency, %

Previous estimated Turbo vs base frequency, %

Cinebench R10 single-thread






Cinebench R10 multi-threads






As you can see, both estimation methods give us very similar results. So, I assume it is safe to consider the above results reasonably accurate.

So, we have analyzed Turbo core's operations both when only one core is used and when all core are actively used: it appear that, using different final frequencies, it can be activated in both case. However, what about an even more heavy test? What Turbo core do in these cases? I will try to answer this question in the following page using a very specific, synthetic test.