Mullins Benchmarks: Application Performance, Encoding

So I've already discussed the improvements AMD has made to their tablet SoC offerings on the previous pages; now it's time to check out how it performs, and see if we can verify some of AMD's claims.

The Discovery reference tablet I was provided with is powered by the A10 Micro-6700T (Mullins) SoC clocked as high as 2.2 GHz. It's paired with 2 GB of DDR3L-1333 RAM, which is the maximum speed supported by the SoC, and it features a 1080p display, 64 GB of internal NAND and the latest version of Windows 8.1.

The main competitor to Mullins is Intel's Bay Trail-T, which I've used several times before in various Windows 8.1 tablets. While the top-end Z3770 SoC is available to OEMs, most devices seem to pack the Z3740, which is clocked slightly lower. Bay Trail-T's official "SDP" is 2W, but this is a largely useless piece of information that doesn't truly illustrate the power consumption of the SoC. Mullins has a TDP of 4.5 and an "SDP" of 2.8W, and based on that I'd expect Bay Trail's TDP to be around 4W.

I've also thrown in a few results from our recent look at entry-level desktop hardware to compare Mullins against SoCs that use more than twice the power. The Bay Trail-D-based Celeron J1900 has a TDP of 10W, while Kabini - AMD's last-generation of mainstream APUs, replaced by Beema - has a TDP of 25W. If Mullins can come close to, or match these chips on a performance level, AMD will have achieved pretty fantastic energy efficiency.

Where possible I've also thrown in results from Qualcomm's Snapdragon 801, Apple's A7 and Nvidia's Tegra 4 ARM-based chipsets. These are found on the majority of high-end Android and iOS tablets, as well as a few Windows RT devices, which Mullins will be competing with in some respects.

Over the next few graphs I'll be comparing the following devices, and their specific SoCs which will be the focus of this section:

Device SoC CPU TDP Type
AMD Discovery 'Mullins' A10 Micro-6700T 4x 2.2 GHz 4.5W x86-64
Lenovo Miix 2 8 'Bay Trail-T' Atom Z3740 4x 1.86 GHz ~4W x86-64
Kabini Desktop 'Kabini' Athlon 5350 4x 2.05 GHz 25W x86-64
Bay Trail Desktop 'Bay Trail' Celeron J1900 4x 2.4 GHz 10W x86-64
Microsoft Surface 2 Nvidia Tegra 4 4x 1.9 GHz N/A ARM
Apple iPad Air Apple A7 2x 1.4 GHz N/A ARM
Sony Xperia Z2 Tablet Qualcomm Snapdragon 801 4x 2.3 GHz N/A ARM

Now let's show you some benchmarks.

First up let's take a look at the two browser-based benchmarks: Peacekeeper and Kraken. In the former, Mullins manages to outperform Bay Trail by 48%, while in Kraken AMD's Mullins reference tablet edged out the Z3740-powered device by 18%.

In Cinebench, which is a CPU-heavy benchmark, things start to get interesting. Here we see Mullins beat Bay Trail-T to the tune of 19% in multi-threaded mode, and surprisingly beating Bay Trail-D by a larger margin. Compared to the top-end socketed Kabini part, the Athlon 5350, Mullins was 24% slower in multi-threaded mode.

However, there's a different story to be told when looking at single-threaded performance. With Kabini and Bay Trail-T, we see roughly one quarter the multi-thread performance in single-thread mode, corresponding well to the SoC's CPU core count. Mullins, though, achieves one third the multi-thread performance when just one core is being used, as that core can ramp up to 2.2 GHz instead of just 1.5 GHz (47% higher) when four cores are being utilized.

A similar situation can be observed in WinRAR's compression benchmark. Mullins is 7% faster than Bay Trail-T in multi-threaded mode, but 18% faster in single-threaded mode. In both situations the 4.5W APU is outperformed by more powerful desktop SoCs, although it's worth noting we're not seeing Bay Trail-D double the performance of Mullins despite its TDP being more than twice as high.

Encoding 1080p H.264 videos on your tablet isn't something you'd likely be doing, but it provides an interesting benchmark. Mullins is 36% faster than Bay Trail-T here, and just 6% behind Kabini despite having an 82% lower TDP. Again we're seeing wins for AMD on an energy efficiency level.