After acquiring processor design companies PA Semi and Intrinsity, Apple designs their own processors that go into iPhones, iPads, iPods and Apple TV. On the other hand, handset OEMs like HTC, Nokia and Motorola buy processors from chip design companies like Qualcomm, Nvidia, TI and Marvell. Samsung is sort of a mixture of both models. They design processors in-house, but also try to sell their processors to other handset OEMs, albeit to no avail, and sometimes buy processors from other companies for their handsets. This is in contrast to Apple’s model of containing everything in-house — they design processors just for their handsets and their handsets only use processors designed in-house.
A benefit of Apple’s model of designing processors in-house is tighter integration of software and the underlying hardware. Other handset OEMs have to buy a processor and pair it with Android. Because the development cycles of Android and processors are usually not in sync, handset OEMs face tough decisions. Let’s say a state-of-the-art processor from Nvidia is released now, but the next version of Android is due out 6 months later. Should the handset OEM wait 6 months for the latest and greatest version of Android or just use the current version and pair it with the new processor from Nvidia? If they decide to wait for the new version of Android, would the 6-months old Nvidia processor be good enough to handle the new OS? The Nvidia processor would have been optimized for an older Android OS and might be too slow for the new OS. Moreover, given the rapid improvement of handset processors, there might be much better processors in the market 6 months later. Should the OEM use those processors instead? While Apple can co-design and co-optimize their processor and OS, other handset OEMs need to grab whatever processor and Android OS that is currently available and stitch them together. Either the processor or the OS is a version that is several months old, which is a pretty long time-frame in the current mobile market.
Although Apple designs their own processors, they still need to license the cores from ARM. Simliar to the HW/SW problem, Apple’s processor design depends on when ARM releases new cores. ARM’s highest performing core could have been Cortex A9 when Apple first started designing their A6 processor, but now ARM has released A15, a newer and better core than A9. If Apple chooses to switch to Cortex A15, it would delay their A6 design and might even delay new iPhone releases. If Apple sticks to Cortex A9, their A6 processor might be slower than their competitors’ using A15 cores and iPhones might suffer from a slower processor compared to Android handsets equipped with the latest ARM cores. For this reason, I can imagine Apple designing their own cores similar to Qualcomm designing their own Krait cores. Apple has an army of designers from PA Semi that is capable of doing that.
If this is the case, why don’t all handset OEMs design their own processors? It is because the cost of designing processors is pretty high, especially for companies with no chip design experience. The figure below compares the cost of designing a processor in-house (Apple) versus the cost of buying a processor (HTC). This is a figure created not with real data, but using my assumptions, so there might be some errors. Please comment if you think these are bogus numbers!
Assuming HTC uses a Nvidia Tegra 2 processor, Apple A5 costs $10 to manufacture compared to Tegra 2’s $5 because the A5 uses twice as much die area. After TSMC manufactures the chips, they sell them to Apple and Nvidia with a 50% gross margin shown in green (TSMC’s 50% margin is the highest among foundries). Apple pays $20 to TSMC, Nvidia pays $10 to TSMC. Here’s where it starts to get different for Apple and HTC. HTC needs to buy the processor from Nvidia. Based on Nvidia’s 50% gross margin in yellow (50% margin is for the entire company. I couldn’t find the margin just for the Tegra business unit), HTC needs to pay Nvidia $20. On the other hand, Apple has a non-recurring engineering (NRE) cost, shown in red, that includes all the R&D cost of designing a processor. I chose $1, but I am not sure how much this number would be because it depends on the total volume of the processor. If it costs $100M in R&D to design a processor and Apple sells 100M handsets containing that processor, the NRE cost per unit is $1. However, if Apple fails miserably in the market and sells only 10M handsets, the NRE cost per unit becomes $10. The reason that Apple can design their own processor, a processor twice as large as Tegra 2 with more performance, is because they have a huge volume that can amortize the NRE cost. If their volume nosedives for some reason, they would be left with the burden of feeding an army of processor designers that increases their per-unit cost of handsets, hurting the huge margin on handsets they are enjoying now. An important thing to note is that the total cost is similar even though Apple’s processor is twice as large as Tegra 2. If Apple had chosen to sacrifice performance to reduce die area, they would have had a significant cost advantage over HTC.
Because of the NRE cost, it’s always preferable to use a single design in as many products as possible, but some products need custom designs to maximize performance or minimize power. Another way to reduce NRE cost is to automate as much of the design process as possible and save money on paying design engineers, but this has its own challenges. EDA software has come a long way in automating chip design, but they are not improving fast enough to automate the complex designs in 22nm and beyond.
Although Apple has not officially confirmed, there are news that Apple has just bought Anobit, a Israeli company designing SSD controller chips. SSDs go into iPhones, iPads, iPods and even Macbook Airs, so the volume for SSD controllers are even higher than processors (unless Apple designs processors for Macbook Airs). As a chip designer, I am very happy to see a company like Apple buying semiconductor startups.