Analysis of the AMD Kaveri Notebook Platform

Before we take a close look at the design of processor and graphics unit, we will give a general overview of the chip. A first and decisive innovation is already obvious here: While Trinity and Richland were still produced in 32-nanometer lithography, Kaveri features the finer 28-nanometer process. Apart from the fact that so-called half node shrink is used (CPUs are usually always produced in full node processes), its technological focus is surprising: While the old 32-nanometer SOI process (High-k + Metal-Gate) still mainly focused on CPUs with high clock rate, packing density got more important in the new 28-nanometer lithography. This is especially important for the complex GPU part.

That is why the die has suddenly space for more than 2.4 billion transistors instead of only about 1.3 billion in Trinity/Richland, although the size of 245 mm² remained virtually the same. Nevertheless, it still only features two CPU modules or four integer cores and a dual-channel DDR controller, which (depending on the model) also supports faster modules up the DDR3-2133. So, the GPU benefits primarily from the higher number of transistors - we will come back to this soon. Other enhancements include support of PCIe 3.0 and TrueAudio technology.

Precise specifications of all Kaveri APUs have already been circulating on the Internet for a while and are now officially confirmed with a single exception: The previously announced A6-7000 dual-core APU is no longer listed in the data sheets. This reduced the line-up to nine models, six of them in the consumer range. (Nearly) all models feature two modules or four CPU cores and an integrated graphics unit. Apart from the usual distinguishing features, clock rate and GPU module, the models also belong to different TDP classes: SV models (standard voltage) belong to the 35-Watt category and ULV models (ultra low voltage) to the 19-Watt category. The more frugal models also lack support of bigger DDR3-1600 modules and the faster PCIe 3.0 standard.

The Pro series in the business range is new. However, its specs are basically identical with the consumer chips of the 19-Watt category. But, AMD guarantees an especially high stability and durability (in terms of availability of the APUs for several years), which the probably cheaper consumers do not offer.

In view of CPU performance and TDP class, the A10-7300 directly competes with the following CPUs/APUs (selection):

• Intel Core i3-4010U (Haswell, 2x 1.7 GHz, no Turbo, Hyperthreading, TDP 15 W)
• Intel Core i3-3217U (Ivy Bridge, 2x 1.8 GHz, no Turbo, Hyperthreading, TDP 17 W)
• Intel Core i3-4020Y (Haswell, 2x 1.5 GHz, no Turbo, Hyperthreading, TDP 11.5 W)

The notebooks and tablets with these processor used in this comparison represent about the average performance level of each platform.

Certainly, a direct comparison with the Richland "predecessor" of the A10-7300 is especially interesting.

• AMD's A8-5550M (4x 2.1 - 3.1 GHz, TDP 35 W)

is still based on the 32 nm process and is only apt for notebooks from a size of 14 inch because of its higher TDP. Alike in the A10-7300, the L2 cache is 4 MB and only RAM types up to DDR-1600 can be connected to the integrated controller. Certainly, the graphics units are also different (see section 3D performance for more information). We have put "predecessor" in quotes since there are no ULV Richland models, but we did not want to go back to even older generations.

In addition, we added the Richland-APU

• AMD A10-5750M (4x 2.5 - 3.5 GHz, TDP 35 W, Richland top model)

 and a device with the widespread

• Intel Core i5-4200U (Haswell, 2x 1,6 - 2,6 GHz, Hyperthreading, TDP 15 W)

to our comparison charts in order to improve the classification.

Apart from an efficiency gain, the most important differences between the A10-7300 and its quasi predecessor A8-5550M are changes in architecture and lithography (28 nm vs. 32 nm), slightly different base and maximum clock rate and particularly the TDP. We could observe that the new model tends to run at lower clocks despite smaller process and that it cannot use its turbo range to the same extent as the predecessor could.

Our CPU benchmarks confirm this observation. Although the A8-5550M performs better than the A10-7300 in all sub tests of all three Cinebench benchmarks, the  lead is small despite the A10-7300's TDP being almost half the Richlands'. The results in other CPU benchmarks emphasize this trend. The candidate can also keep up well with comparable Intel CPUs.

As mentioned in the introductory article, the integration of significantly faster graphics units than Intel's HD GPUs has been a unique feature of AMD APUs for a long time. In order to categorize the 3D performance, we use the pure graphics benchmarks of 3DMarks 2013 and 11 as well as of Cinebench R11.5.

While the Radeon HD 8550G integrated in the A8-5550M performs just under 50% better than the A10-7300 in the relatively undemanding Ice Storm DX9 benchmark, the difference gets smaller with higher requirements (Cloud Gate, DX 10) and is small in 3DMark 11 - although the clock rate of the Radeon R6 is lower and the TDP of the complete A10-7300 is almost half the A8-5550M's. The new GCN architecture appears to be advantageous here. This gets even more apparent in the ComputeMark v2.1 scores since the GPGPU benchmark particularly profits from GCN. In addition, more execution units are active in the R6.  Apart from few exception, the on-chip GPUs from Intel perform significantly worse. Finally it should be noted that AMD's advanced MANTLE API is first supported by mobile APUs with Kaveri. A first benchmark with Thief confirmed that systems with weak CPUs for their graphics units like the Acer Aspire E5-551-T8X3 benefit significantly from it.