Apple change of strategy?

With all of the recent chatter regarding the disappearance of Intrinsity from the technology community, a question opens up about the strategy of the company that has in all likelihood bought them: Apple.

It all started with the announcement of the Hummingbird processor by Samsung - see my January 27th post about the ipad for a link in the comments to the press release surrounding the Samsung/Intrinsity collaboration to create a 1GHz implementation of the ARM Cortex-A8 processor.

When the iPad was announced, and the clock speed was said to be 1GHz, and it runs the iPhone OS, and the timescale was about right for Hummingbird to be in production, it was clear that the great performance of this hardware is in part down to the excellent implementation skills possessed by Intrinsity, since this chip is in all likelihood an exact match for a Hummingbird enabled System on Chip (SoC).

So why would Apple buy Intrinsity? I think the answer lies somewhere between their need for full custom silicon across the board (including at the processor RTL micro-architectural level), and the need to proceed quickly to market with the fastest possible implementation of off-the-shelf components, built in a semi-custom manner using Intrinsity's techniques.

Intrinsity possess a mix of technology and knowhow that allows them to do a really good job of making designs operate faster. They create a custom version of a given off-the-shelf processor design that addresses the slowest paths through the chip, and then adds special fast logic into those paths to speed them up. This makes the logical path between the clocked elements of the processor shorter, allowing the clock to run faster.

Because a strategy that focusses around Intrinsity relies on feeding them with other people's silicon designs (they do not actually design anything - they customise the generic design descriptions for fast operation in a given semiconductor companies process node), this suggests to me that Apple might be continuing with their use of ARM's licensable processor IP for the foreseeable future.

If Apple sticks with optimising off-the-shelf components in ways that other people cannot - now that they own Intrinsity - what does this say about their suspected ambitions towards designing their own in-house ARM processor?

I think the answer is two-fold.

Firstly, Apple are getting good performance out of ARM's "soft" cores by being clever in how they are implemented. This is quite low-hanging fruit, since the whole design and verification task has been done by ARM in creating the CPU in the first place. The "only" thing Intrinsity would have to do is identify the areas of the design that were holding back the top-end frequency, and make optimisations to those for speed.

Secondly, it isn't easy designing a processor from scratch. I believe that Apple retain a large percentage of the CPU design talent that they obtained from PA Semi - it would only have been the top-brass involved in business development that would have left to form Agnilux (now bought by Google). However, PA Semi's last ARM implementation would have been the experience gained from working on StrongARM at DEC and the ARM architecture has certainly moved on to include many new features. For an architect to get up to speed on that, and then work out what the CPU will look like is non-trivial, and bringing a team up to scratch on how to build that - and verify it - is not a two year task. It can take up to 4 years to turn around a CPU from scratch, especially if you've not worked with the architecture either for a long time, or on it's latest incarnation.

It also opens the question of what Apple thinks they can achieve over and above what ARM can achieve in CPU design? ARM are experts at designing their own CPUs, and have a spectrum of nifty high-performance implementations available in the applications CPU space. Give or take wanting a CPU with better performance or lower power, I can only think that Apple have decided to take their time on designing their own processor, and will stick for the foreseeable couple of years using optimised ARM macrocells.

If/once Apple do successfully complete the creation of their own CPU, the Intrinsity guys would then be deployed to analyse it, and make its implementation go faster. However, if you have that many in-house implementation engineers working alongside your designers, it would be possible to feed information to the designers about where the worst paths were in the design, and get them shortened before moving from the design description into the synthesis and layout stages of the project. This to some extent negates the value of having Intrinsity on board, because their skills lie in taking a pre-verified box off-the-shelf that you cannot change in any way, and making logically equivalent cycle accurate implementations of it that are faster and possibly lower power than if they were just synthesised automatically into standard bulk-CMOS. As such, it is a little bit questionable how much of Intrinsity's unique abilities can co-exist with the presence of an in-house design team, and what percentage benefit can be gained by using Domino logic on a design that is as fast as it can be in its RTL implementation.

So I've said it before in a previous post and it is worth repeating here - PA Semi designed CPUs in Apple devices will take many years to appear. Now I can add that they will only be marginally faster than what opther companies are capable of now that Intrinsity co-exist alongside a design team working towards a common cause. Just where I think Apple's own processor designs might wind up appearing in the product line is a matter for another post. But one thing is sure - Apple are intent on crafting their own silicon, and the game is changing still.

iPad - all the hall marks of hummingbird

Apple today announced the iPad - their slick looking tablet computer. It is based around an ARM architecture processor, but in a chip package designed by their semiconductor team from P.A. Semi. Their role has undoubtedly been vast in the creation of the chip (called the Apple A4 chip), taking on various technologies and combining them together into one SoC.

Their semiconductor manufacturing partner is Samsung, and a recent announcement from Samsung regarding a custom implementation of the Cortex-A8 processor using Intrinsity's Fast14 domino logic seems the prime target for the processor in the iPad.

With Apple's investment in Imagination Technologies, Apple will have one of the latest PowerVR SGX cores sitting nicely alongside the Cortex-A8. This pairing - alongside possible video engines, and other features like high performance memory controllers etc., make the SoC design undertaken by Apple thoughtful, but not mind-blowing. Their use of a domino logic implementation  for the A8 macrocell is the key to the speed uplift, and the low power usage even at just 1V. But there is no mystical P.A. Semi designed custom ARM CPU implementation in there right now. They've merely added the necessary tweaks to get what they already use in the iPhone running as fast as it can possibly go without burning a hole in your purse.

Where Apple seem to be best at differentiating themselves is in the software, though. That thing looks slick, and I can't wait to get a play around with it! The browser and maps look flawless, indicating a very good level of performance in the adapted iphone OS. And their graphics drivers will have to be a step up from what is required for iPhone, given the extra clarity in the display.

I didn't like the name, and I didn't originally like the concept, but Apple stuff is always so beautiful, and it always just works. I now look forward to its arrival here in the UK.

The fact that it doesn't have Intel anywhere near the thing also makes me glad.

There will be no PA Semi designed ARM core in apple tablet....

....unless they are planning to release the tablet in 2012 (do read on for why)....

I've been reading the hype regarding Apple's tablet computer thingummy, and there seems to be a zeitgeist of opinion that Apple's latest gadget will contain an ARM compatible CPU designed by the team they gained after acquiring PA Semi in April, 2008.

This article claims that a device will appear at the beginning of 2010, and that during its development (over a 4 year period) Apple had considered use of the Atom processor, rejecting it on grounds of power consumption. To bridge this gap, it is claimed, Apple bought PA Semi at the end of April 2008, in order for Apple to have an in-house capability for System On Chip (SOC), as well as to allow it to differentiate it's technology from everyone elses.

At the moment, iPhone teardowns reveal exact component specs, allowing the revelation of every piece of hardware (and, subsequently, the processor). This is how we all know that the iPhone 3GS contains an ARM Cortex-A8 (along with the Apple Job adverts specifying knowledge of the ARM NEON instruction set - Cortex-A8 is currently the only ARM silicon that has a NEON pipeline attached). Apple could - by designing its own chips - keep the content completely secret - something you cannot do when you buy an off-the-shelf component.

Right now, talk of this tablet computer has people speculating that it will contain PA Semi's first CPU design for Apple.

I think this is unlikely.

Apple acquired PA Semi in April 2008. It takes nearly 2.5 years to bring a new microprocessor - from initial spec, to releasable quality - into a form where you could conceivably manufacture it, having validated it enough.

Various stories have emerged over the last year or so speculating that Apple will be using the PA Semi team to design an ARM CPU. Who knows if this is true, or if any of the rumours about this tablet are true?

PA Semi's last exposure to ARM CPUs was when some of the team worked on the StrongARM processor at DEC. This was an ARMV5 architecture processor, a very different beast to todays ARMV7 processors. Time has moved on, and there is a whole different architecture to learn, with many new features. Just to learn this new architecture would take an architect 6-8 months to master, during which time they would model some of the micro-architectural ideas they have floating in their head. This could happen in parallel with some Register Transfer Level (RTL) trialling, but only in a behavioural sense. By the time PA Semi has done all of this, it brings us to the end of 2008, with nothing more than an understanding of the architecture, and various models of the micro-architecture of the CPU.

Even with an installed product design team working on concepts and prototypes for the actual look and feel of the hardware (remember, they'd been working on this for nearly 2.5 years, if the story is true), the actual silicon inside is a very different beast. Assuming that work began on the actual design of the CPU, it would still take nearly two years to get the core to a state where ARM would sign off on their architectural compliance (if, indeed, the rumours of an ARM compatible CPU design are true). This takes us to the end of 2010 for Apple to have bespoke, manufacturable Silicon.

So no, the timescales are all wrong. And remember, the task above just relates to designing the CPU - it doesn't include the time taken to design a SOC around that CPU. They may have access to various designs already, but if they are going to the trouble to work on a bespoke CPU, then they won't at the same time go and take some other SOC - htey'll build a new one.

Their recent share interest in Imagination Technologies graphics processor design house indicates that Apple will be using their technology in their SOC, but without an in-house design for a CPU, where will they get that from?

To design a SOC - into which they would drop an ARM CPU - they would need an ARM CPU license. When they just purchase an off-the-shelf chip package - like the ones they get from Samsung - they do not need this, because Samsung are the ones who have taken out the ARM CPU license and manufactured a SOC with it.

I think the likely first set of tasks that the PA Semi team will be doing is to create a SOC using a pre-existing set of designs that they would license, so as to keep the flow of products emerging from their portfolio.

There are interesting times ahead, but I don't believe that Apple have a custom ARM CPU ready to go for an early 2010 release. If this even has a shred of truth to it, expect an early 2011 release for anything like that.