Sunteți pe pagina 1din 6

Apple processors: Designed differently but here to stay

Don Scansen, Paul Boldt 2/20/2013 1:33 AM EST


What's next in Apple's iPhone- and iPad- powering A-series processor family? Two experts prognosticate. Rumors of an iPhone 5s and iPad 5 are already circulating. Nothing too intense yet but they are there. When these two devices do land they will likely sport an A7 and A7X, respectively. That is, unless Apple pulls the plug on its custom chip design and opts for an off-the-shelf Intel product. This is not our hypothesis but it is out there, even on Wall Street. At the end of November Doug Freedman, an RBC Capital Market analyst, floated the hypothesis that Intel would fab Apple processors if Apple moved the iPad to an Intel processor. It almost suggests the Apple-designed application processors (APs) are somehow a stopgap until something better appears. This seems like the perfect backdrop to look at the A-series family, think about the evolution therein, and weigh the evidence for Apples semiconductor design intentions. Before we get too far ahead of ourselves lets consider the A6 and A6X. The debut of the A6 along with the iPhone 5 was pretty much expected from both the hyperactive Apple rumor mill and the simple logic of refreshing the AP in sync with the iPhone. However, the appearance of the A6X, some 41 days later inside the iPad 4 was very much unexpected. It was a little more than half way into the usual annual iPad refresh cycle and there was only a vague, late warning from the Apple pundits. Die photographs of the A6 and A6X that were first published by Chipworks, with our added annotations, are reproduced below. Probably the single most striking feature of both is the inhouse designed CPU. Going further, it was a custom design where the layout was performed manually, instead of using the common automated place and route approach. In their discussion of the A6 ChipWorks commented: "This is a more expensive and time-consuming method of layout. However it usually results in a faster maximum clock rate, and sometimes results in higher density." They go on to say: "In fact, with the exception of Intel CPUs, it's one of the first custom laid out digital cores weve seen in years!" There isn't much to add except to say that Apple is serious about and making considerable investment in their design capabilities. So, regardless of the reason for the custom design, this clearly supports the hypothesis that Apple is not producing placeholders while it waits for an Intel processor.

Apple's A-series family tree Chipworks presented some quantitative numbers comparing the A6 and A6X, in their November 1st discussion of the A6X. With a die size of 124 mm, the A6X is 28% larger than the 96.7 mm A6. The CPU consumes 15 mm on both parts, while the triple core GPU of the A6 weighs in at 16 mm, and the quad core GPU on the A6X comes in at 35 mm. From the above numbers it is apparent that beyond the additional 19 mm for the GPU the A6X sports approximately 10 mm additional area. This may not seem like that much, but it is actually an area larger than the individual GPU cores. In terms of accounting for this additional area we can again turn to Chipworks who noted that the A6X doubled the SDRAM interface width and included some new interface blocks. They also observed one fewer PLL for the A6X which freed up some mm. So we gain some area here and use a bit less there. The question though is whether there are other differences. We will take a look at this, but in the context of the evolving A-series family. A visual interpretation of Apples chip lineage should help. The Family Tree An A-series family tree is presented below. Starting with the A4 in 2010 and ending with the A6X just a few months ago, there are now 5 members in the family. Basic information for each AP, including the device in which it was introduced, its dimensions (mm) and die size (mm2) are included. The tree has been constructed to schematically illustrate two major evolutions that have occurred within the family over the last three years. First, the original family line bifurcated into two distinct lines of APs. Second, the family went through a process shrink from the original 45 nm to the current 32 nm fabrication process. Each of these events will be considered in more detail.

The bifurcation The X branch was introduced at the 5 generation with the A5X in March 2012. Its die size grew by 35% compared to the A5, i.e. the same generation of the other branch. At least some of this extra area was due to a doubling of GPU real estate. The increase in graphics horsepower compared to its generational counterpart emerged with a four-fold increase in the number of pixels from the iPad2 to the iPad3. Further evidence of the bifurcation of the A-series family came in the digital blocks outside of the CPU, GPU and GPU spine. There are 12 for the A5 and 15 for the A5X. What does this bifurcation look like at the 6 generation of the A-series family? As pointed out above there is again much more real estate devoted to the GPU in the A6X compared to the A6. We also know about differences in the layout, interfaces and PLLs. It is time to look at the digital blocks outside of the CPU and GPU. There is most likely considerable block commonality between the two processors. Again referring

to Chipworks A6X article, Other than the CPU, it appears all the other digital cores have new layouts, and many of the analog and interface cores have been reused from the A6, however there are some new interface blocks. But are there digital blocks on one die and not the other? We saw no comment on this. They did mention that the A6X sported considerable new design and was not just a tweak. It would, however, be interesting to consider whether there is evidence of a design overhaul in the form of different digital blocks, not just layout differences. Looking at the die photographs, the A6 and A6X appear to have 17 and 16 digital blocks outside of the CPU, GPU and GPU spine, respectively. For sure, there are blocks that are the same on both parts. The first thing to notice is that the shapes of the blocks differ between A6 and A6X. Obviously, floorplanning of the entire device is dependent on the number and location of the largest blocks (an additional GPU core in the A6X) leaving smaller blocks to be shoehorned into the available space. With no intention of minimizing the task of re-shaping circuitry to fit major architectural changes, this floorplanning required more than a copy and paste function. There do however appear to be other differences, again beyond the PUs. We noted above there appears to be a difference in the number of digital blocks. Is there more? Given a desire and the concomitant budget, hard-core reverse engineering is possible. Beyond reverse engineering the question of differences can be considered in two ways. One can qualitatively analyze the blocks to consider the amount of cache, for example. With this approach the block labeled B in the A6 die photograph does not appear on the A6X, even considering layout differences. One can also look at this more quantitatively, yet without a large reverse engineering project. Of the five digital blocks (see die photographs) most easily identified between the A6 and A6X, four occupy areas between eight and 10 percent larger or smaller than on the A6. The differences are at the upper end of measurement errors (we are relying on CW published die dimensions and the accuracy of the associated image files posted on web sites), so are believed to represent minor design changes within the blocks. The fifth digital block analyzed varied significantly between the two designs. The A6X version of this digital block was just over three times that of the A6. Getting back to processor cores for a moment, there are four GPUs on the A6X compared to three on the A6 (as noted by Chipworks). The design within each of these also differs significantly. The GPU blocks on the A6 are 58% bigger than their A6 counterparts. Accounting for the GPU cores along with the block that could either be a dedicated decoder or perform a GPU load balancing operation, there is nearly twice the die area directly devoted to graphics processing on the A6X. The shrink The second major evolution to date was the change from a 45 nm process to the current 32 nm process. The A5 was the first public part to be manufactured in the two processes, with the 32 nm A5 appearing in the late winter / early spring 2012. Because of this shrink in fabrication process geometry the raw die and block sizes do not provide a good comparison of the family members. To this end we have normalized die and block sizes to the 32 nm process. The results are shown in the table below.

A number of things jump out from the table. First is the continued increase in 32 nm equivalent die size for both processor streams as seen in the third column. The A6 was introduced during the September 2012 keynote as being 22% smaller than its predecessor. This is true on a silicon basis, but on a circuit basis i.e. corrected for the shrink, it is in fact 36% larger. The same is true for the A6X which is 29 % larger than its 5 level sibling. There are quite a few more transistors on these parts. Looking within the die, the corrected CPU area is 50% larger with the custom designed cores on the 6 level parts. The area devoted to the GPU, on the other hand, has continually increased in the two streams. While the iPad 4 maintained the same 3.1 million pixel count as the iPad 3, there was room to improve the graphics capabilities of the A6X compared to the A5X. Thus, the 39% increase in GPU area on the A6X compared to the A5X was not to handle more pixels, but to provide much better performance. How are Apple design efforts progressing? When we first wrote about the A4 for EETimes in 2010 we found that there were many similarities with the Samsung design. Later in 2011 when the A5 came along, we thought it was a bold step forward, and speculated on possible avenues of customization for Apple because they control both the integrated circuit design and the OS. Probably the most striking thing about the A6/A6X is the Apple designed dual ARM core that each sport. A close second would be the continued increase in circuits compared to earlier generations of the A-series family. So in summary, the A-series family has evolved into a serious bit of circuit design and is here to stay. Period. Some parting thoughts First, it will be interesting to see if the die size plateaus at some point. With the pixel count now stabilized for the last two generations of 10.1 iPad, the added circuit horsepower is driving improved performance not more pixels. Second, it will be interesting to see how the block count evolves going forward, with IP of the anobit or authentec acquisitions possibly showing up in future designs. Finally, the big one: Will there be an Apple designed GPU? It is clear that the GPU is continually consuming more silicon whether you look across the whole family our just within one branch. Based on the move to a custom CPU design this may well be in the back of their head. The question of sufficient engineering resources and/or graphics IP is a topic best left for another day. About the authors: --Paul Boldt is a principal analyst at Ned, Maude, Todd & Rod Inc., an Ottawa-based technology research company. --Don Scansen is a partner at IP Research Group, a technology consulting firm delivering service to intellectual property clients.

Related stories: Apple's A4 dissected, discussed...and tantalizing A5: All Apple, part mystery Apple A6 processor has landed, sort of

S-ar putea să vă placă și