A10 5800K IGP (7660d) overclocking issues

The thing is I haven't overclocked the CPU part at all. In fact, I've undervolted it and disabled turbo, reducing heat & power consumption.

What I have overclocked is the GPU part, and my problem is that after a point (950mhz, keep in mind default is 800mhz) the OC isn't improving results, and I do not understand why. I can't think of a reason for it to behave that way.

I understand luck of the draw and all with chips. My issue is that I haven't even tapped into its full potential yet.

Lets say you have a CPU. Its running happily at 2 Ghz. And then you increase it to 2.1 Ghz, and its perfectly stable. That is a 5% increase in clock speed, correct? which theoretically will yield a 5% increase in performance, and reality is usually very close to that with CPUs.

Now here is me, with my APU. It's IGP was running happily at 950mhz. Increased it to 1000mhz, and it runs perfectly stable (at stock voltage too, meaning I should be able to increase it further). That is a (roughly) 5.2% increase in clockspeed. Realistically I won't see the full 5.2% increase in performance, but there should be a measurable increase.

- and here is the problem. My 5.2% increase in clockspeed... yielded a .06% decrease in performance. Even taking into consideration limited bandwidth from DDR3 and luck of the draw with chips, something is very wrong in that picture.

Yeah. I don't understand why the IGP has gotten so little attention. Most reviews hardly spend any time on it, yet I was under the impression the IGP was the one thing that made APUs worthwhile.

Im wondering if I've hit a wall due to memory bandwidth. My memory is at 2400mhz and I thought that would be enough to give 1000mhz room to breathe but maybe its not.

Later today I'll do some tests. I'll drop my memory to something stupid low like 1066 or 1333 and compare 800 vs 900 vs 1000 on the IGP. Think that would be enough to determine if it's memory bandwidth related?

Since there really haven't been any in depth guides on Trinity/Richland IGP overclocking it would be good to gather the information into one place. At the least, I'd input what I've learned.

I ran three 3DMark11 benches real quick in an attempt to find out if memory bandwidth alone is what is holding back the IGP in terms of gains from clock speed increases.

I set my memory to 1333 with default timings set by my motherboard. Since my goal was to compare my results at 2400mhz to a setting that purposely had significantly less bandwidth, I figured it was good enough. I set the northbridge to 2000mhz.

http://www.3dmark.com/3dm11/7447278
http://www.3dmark.com/3dm11/7447143
http://www.3dmark.com/3dm11/7447357

The IGP benches are at 800, 900, and 1000mhz respectively.
800 -> 900: +4.8% (graphics score)
900 -> 1000: +2.3%

It can be seen that the returns are rapidly diminishing, but I actually expected the gains to be smaller with such reduced bandwidth.

I only ran these tests once, but I feel the gains demonstrated are big enough to be outside the margin of error. Unfortunately I'm still left with my question unanswered, as I still have not figured out why I have not seen a performance improvement going from 950mhz to 1000mhz when at 2400mhz memory (at that speed, the difference I saw was basically 0% taking margin of error into consideration).

This is currently my record using clocks & voltages that I would use 24/7 (3.8Ghz, IGP 950, NB 2300). It's worth noting that many comparable setups with the IGP at 1000-1100mhz are only equaling, not besting, my score at 950mhz, which I feel demonstrates the severely diminishing returns after 900mhz.
http://www.3dmark.com/3dm11/7448474

It is already well known, but the benefit of memory is huge. Increasing memory frequency 25% (which is roughly 1866->2400) seems to improve performance far more than a 25% increase in IGP clock speed (800->1000mhz).

Currently I am experimenting with the baseclock, and trying to push my memory farther. So far I've been unable to complete a 3DMark run at 2500mhz memory, but I did have a successful run at 2475ish. When I get more time I'll need to find another program to test graphics, as I feel using 3DMark alone isn't an adequate test.

I'm not sure if people are aware of this or not (perhaps I was just lacking common sense) but aside from affecting CPU, RAM, and NB, the BCLK also affects the IGP clock. With an increased BCLK, the IGP clockspeed reported by CPU-Z is incorrect (a number of other programs reported incorrect speeds as well, I can't remember which ones).

GPU-Z does correctly display the IGP clockspeed. If you are trying for a high BCLK, make sure you reduce clocks appropriately.

It is supposed to dynamically determine the amount of memory it uses depending on how much it needs, all the way up to 2GB I believe. There is an option in my BIOS to manually set it, and I set it at 1GB and didn't notice a difference.

Also, I noticed an interesting entry in the 3DMark scores that matched mine (P1864). It had...
Stock CPU (4.2GHz)
Stock GPU clock (800mhz)
1333mhz memory
6670 setup in dual graphics.

Basically having a 6670 dual graphics setup with 1333mhz memory bottlenecks it enough to hold it back to IGP + 2400mhz memory levels.

manually setting the memory will not improve the performance in any game.

You seem to be missing the point. I am overclocking the integrated graphics, not the CPU, so I have zero interest in overclocking an FX processor.

Kaveri is not FM2 compatible. It requires a new socket, FM2+. I actually waited to get my motherboard so I have a FM2+ board.

FM2 processors (trinity/richland) work on FM2+ boards, but FM2+ processors (Kaveri) will not work on FM2 boards.