Ultra-Speed DDR3 Revisited: AMD APU Memory Scaling

Test System and Overclocking

After completing the initial review of the Kingston HyperX Predator 2666MHz DDR3 last month, we immediately received feedback from readers looking for benchmarks on an APU-based system, and it was an especially valid request considering the lack of performance scaling from the Intel test platform in that review. Speaking of platforms, right off the bat I will address a choice made in the selection of components for this test platform: the mini-ITX motherboard. Why mini-ITX? With APU’s integration of CPU and GPU, building a system without a discrete graphics card offers the possibility of a full-featured system build in a small form factor (SFF). As previously noted, an APU-based system can be constructed within the same price segment as a modern game console – and we've already seen plans for small PC systems the size of consoles with the upcoming Steam Machines first announced at CES this year. A mini-ITX motherboard would allow any builder to start creating such a system today.

The motherboard chosen for this review bears a special mention, as it was a good performer and an even better value. The ASRock FM2A88X-ITX+ motherboard is a full-featured solution for a SFF APU build, and for its cost ($109.99 on Newegg at time of writing) it has a very impressive feature set. This motherboard features dual-band 802.11n WiFi, Bluetooth 4.0, triple monitor support (with VGA, DVI, and HDMI outputs), gold caps, and even a premium headphone amp from TI! Most importantly for the purposes of this article, however, is this diminutive board’s memory overclocking potential. In a market where the majority of offerings are limited to DDR3-2133 support, this board officially supports memory speeds of DDR3-2400.

Overclocking

Who wants to leave free performance on the table? Naturally, since this review is all about speed - and the Kingston HyperX Predator memory from our first review is made to be highly overclocked - it was clearly appropriate to take the 7850K APU clocks as high as possible! The end result was limited to my own ability and the capabilities of this motherboard, and as these adjustments had to retain stability through all of the benchmarks some compromises inevitably had to be made. Not being an expert at CPU overclocking by any means I was still able to find my way through without too much trouble in the UEFI setup. It was a slow process, however, beginning with the problem of ultra-speed memory support.

The memory speed on the FM2A88X-ITX+ could not be set higher than 1200MHz (DDR3-2400) without adjusting the 100MHz base clock (BCLK), so I had to attempt an adjustment there. Since this HyperX memory can theoretically be run at up to 2666MHz, I tried many voltage and frequency changes to attempt this result. After many failed boot attempts I finally found the ceiling with this motherboard/CPU with a 104MHz BCLK. Since the the highest memory setting (DDR3-2400) on the board is using a multiplier of 12, this 104MHz base clock would have resulted in a frequency of 1248 MHz, which would be high enough for DDR3-2500 (OK, technically DDR3-2496).

In reality, however, the base clock actually operated a little higher at this setting – reading anywhere from 105-107 MHz in Windows using CPUZ.

The 2500MHz DDR3 ran slightly higher according to CPUZ

After finding a stable base clock, I started increasing the multiplier until I reached 4.5 GHz with a multiplier of 42. This CPU overclock was achieved along with the DDR3 overclock with some very high CPU voltages. The next problem became achieving CPU overclock parity with the stock memory speeds used in comparison here, and while some adjustments had to be made to the CPU overclock, the resulting speeds were very close. The BCLK was reduced to the stock 100MHz for the DDR3-1600 and DDR3-2133, and a simple multiplier overclock was employed (x45) to bring the core speed back up to 4.5 GHz for these tests.

The 1600 and 2133 DDR3 tests ran with 4.5 GHz on the CPU as well

Naturally, I couldn’t resist trying to get the IGPU clock higher on the 7850K before beginning the benchmarking process. This is an APU, not just a CPU after all! After raising the voltage to 1.4V for the internal graphics, I ended up settling on the 900MHz mark, which was not only a nice increase, but was easy to duplicate with both CPU overclocks.

Higher clocks on the GPU side are certainly possible as I was able to hit a stable 1GHz without any CPU changes, but such high clocks were not possible once the CPU and RAM were overclocked as well (at least during my experimentation). With 900MHz along with the memory and CPU gains we should see some pretty good performance here!

Memory Speeds

With the system stable it was time to run some benchmarks. Before proceeding we had to choose what memory speeds to compare along with the max overclock, and this is where we chose to duplicate the speeds from our look at this HyperX memory on the Intel system. AMD not only offers official support for 1866MHz on their desktop parts (unlike Intel’s 1600MHz desktop CPU limit), but 2133MHz is officially supported on the newest APU’s as well. For comparison I chose the same stock 1600MHz CL11 and 2133MHz CL10 memory settings from the Intel review, changing all timings and voltage manually to exactly replicate the speeds from the previously sampled DDR3.

Performance Consistency

A quick note on the consistency of the test results: As I mentioned above, though taking two different approaches with the DDR3-1600/2133 and subsequent DDR3-2500 runs, the CPU and GPU frequencies are extremely close, and certainly within limits for a good comparison. IGPU frequencies only fluctuated from 898 MHz to 900 MHz between overclocks, and the CPU held between 4.51 and 4.52 GHz. The baseline DDR-1600 operated with 11-11-11-28 timings @ 1.50 V, and the DDR3-2133 at 10-12-12-31 @ 1.60 V. For the max overclock of DDR3-2500, the Kingston XMP Profile #2 timings for the HyperX RAM were used, providing timings of 11-13-13-32, and required a little more voltage for stability on this system with 1.68 – 1.69 V used during testing.

Now for some benchmarking!