First of all, we will start with the GPU of A8, and it is obviously unrealistic to simply get the specific structure of this part through software analysis. First, Apple said that the new GPU has a 50% performance improvement over the previous generation, so our initial guess was that the built-in GPU model of A8 was the PowerVR GX6650 of Imaging.
In the disassembly analysis, we found that the number of GPU cores of A8 was only 4, which directly ruled out the possibility of GX6650 that we had guessed before. According to the four core GPU units, The GPU used in A8 should be GX6450, which is the new generation of G6430 used in A7. GX6450 has added some performance optimizations and functional upgrades - including ASTC (adaptive scalable texture compression, see details ）Yes, this has been confirmed in the Apple store file, so in terms of GPU, A8 is just a daily upgrade of Apple.
The A8 adopts the latest 20nm process, with a total area of 19.1 square millimeters, which is smaller than the A7's 22.1 square millimeters. This also saves Apple some valuable space on the motherboard, which is occupied by the GX6450, which is larger than the previous generation, and the additional SRAM to enable the CPU to have a larger cache. At the same time, when we look at the A8 processor from the geometric perspective, we can easily find that the area occupied by a shared texture processing unit corresponding to each two GPU cores is so large compared with the area occupied by each GPU core. It is not difficult to understand why Imagination is unwilling to equip each GPU core with a separate texture processing unit.
At the same time, opposite the GPU is the CPU core of A8. Different from the GPU, the area of the CPU core has been significantly reduced: from 17.1 square millimeters of A7 to 12.2 square millimeters, while the CPU core of A8 is the same as that of A7. It is not easy to clearly separate the two cores, but it can be determined that, The CPU structure of A8 is obviously an evolutionary product of Cyclone. Apple made several improvements to it this year, making the new architecture more potential than the original architecture. Chipworks analysis said that the L2 cache in A8 was inserted into each CPU core. After all, it can be seen from the picture that A8 processors do not have the obvious L2 cache structure like A7.
The last landmark module of A8 is the SRAM cache memory. On A7, this module has a capacity of 4MB and provides services for both CPU and GPU. On A8, it is almost the same. When the CPU adopts the 20nm process and significantly reduces the area, the SRAM module of the same size as A7 is very eye-catching. In Chipworks' preliminary estimation of the module area, The SRAM capacity on A8 should be larger than 4MB. After actual analysis, Chipworks determined that the area of each bit in the SRAM of A8 was reduced from 0.12 square microns of A7 to 0.08 square microns, only 33%, which explains why the SRAM module in A8 occupies more space than other parts, but at least after reducing the bit area, It is worth affirming that the new SRAM can provide more storage space in the same volume.
A8 is one of the first SoCs in the market to use the 20nm process. The smaller CPU area allows Apple to further reduce the volume by about 15% while injecting more functions into A8. Chipworks measurement results show that the area of A8 is 89 square millimeters, which is smaller than the 104 square millimeters of A7 using Samsung's 28 nanometer process. The reason why the reduction does not reach 50% in theory is that Apple has further improved its performance.
Via &
