Hacker News Reader

Apple acquired PA Semi, which wasn't just an ASIC group, but arguably the best group of processor designers for high-performance, low-power computing. The group previously built StrongArm (a famously high MIPS-per-watt processor in the 90s). It's a much more specialized skillset than normal ASIC dev.

The reason it should work out better for Apple than it did for Sun is

(1) Apple sticks with ARM architecture so they don't have to custom-build the whole toolchain (eg compilers)

(2) Apple ships in insanely high volume so the cost of chip development is pretty low amortized over their volume

(3) Performance-per-watt is really important in mobile, and they're doing it with an amazingly accomplished team in that area.

Apple had StrongArms running on its Newton Message Pad 2000 and 2100 and they rocked!!! :-)

PS: I know it is a bit off topic but I really liked StrongArms back on the day.

I wish people would stop obsessing about PA. That acquisition was a fiasco. All the talent bailed after a tiny lock in period, went to a "startup", and were promptly "acquired" by the googs. The whole thing was a mess that should have ended with a vp or director "pursuing other opportunities".

The Intrinsity talent buy was way better. Those guys actually put out product and have a functional team.

Yup. Apple has been putting together a CPU team for a long long time now, including high-perf simulators, libraries, etc. etc.

However, the point about this catching up to them and biting them is a real possibility. Now, they have a lot of discipline, it seems, and that is critical. The other part is having a very tightly closed platform with control of the entire toolchain.

I have over 100 people in my LinkedIn network who were ever at PA Semi, and 53% of them are still at Apple. Where's the evidence that "all the talent bailed" ? Some of the key architects and leaders moved on, but I think that's not unusual.

I'm not sure that Sun is the best comparison here, as they did both the ISA and the microarchitecture on their chips. If the claims in this article are to be believed, Apple is doing their own uarch, but keeping the armv7 ISA (and gluing some other ARM building blocks together).

When Sun's architecture started to fall behind, there weren't really second sources for compatible chips. Combine that with a greater reliance on ISA optimized assembly than today, and you can see why their customers started to leave.

If Apple did roll their own ARM this time around, they almost certainly built prototypes of the iPhone 5 with a "standard" ARM9 design for contingency reasons. Getting the software to work on either hardware should have only been a matter of setting some optimization flags and swapping out a few header files. Easily worth the risk, in my opinion, as they can always revert back to a uarch designed by ARM if their own designs start to lack performance or become too expensive in the future.

I agree it certainly isn't a 'clean' comparison, but at the time Sun didn't have the option of licensing the 68K architecture from Motorola and extending it in their own silicon. I still remember a talk with both Bill Joy and Andy Bechtolsteim where Bill was over the top enthusiastic about finally having the ability to change the CPU to work better with the software rather than having the software work around what the CPU thought was best. I don't doubt for a minute that he at least would have been just as enthusiastic if the 68K could be augmented in a Sun specified way. There is a huge efficiency in customizing the CPU to do what you want.

This I found interesting:

"If Apple did roll their own ARM this time around, they almost certainly built prototypes of the iPhone 5 with a 'standard' ARM9 design for contingency reasons."

How likely do you suppose it is that they did exactly that and called it the '4S' ? Having worked in places doing new CPUs (Intel in the '80's, Sun in the '90s) you always make a product based on the old stuff in the event the new stuff doesn't hit the schedule. The Sun-3/60 was exactly that machine, it was the 'in case the SparcStation 1 slips' machine.

It seems very unlikely that the A5 used in the 4s was a fallback plan.

First of all, they were already getting a huge speed bump by going from single core A8 to dual-core A9. So existing iPhone users would have been happy. There was also no foreseeable external competitive pressure at the time, everyone knew that there weren't going to be any S4 phones for 6-9 months, and that Tegra 3 would be a joke.

They were also still on a 45nm process, and the A5 with dual-core A9s was already huge (120 mm^2). They would not have wanted to increase that at all, even if they had a working design of their own. What would have been the options? Seems like it would have been either reducing GPU resources, or dropping a core. Hard to see either of those being worth it.

Finally, one could argue that they main plan was their own core@32nm, and the A9@45nm was the fallback if there were process troubles. But these days there seems to be a strong preference to alternate new designs with new process nodes, rather than doing both at the same time. And in fact this conservative option is exactly the one Apple picked next spring: the iPad refreshes were a mix of 32nm parts (simple shrink) and 45 ones (higher clock, more GPU). If Apple weren't willing to risk going to 32nm even for a design iteration as conservative as A5 -> A5x (and the process was clearly working for them with the A5 shrink), the idea that they would have tried for a radical new design + new node 6 months earlier doesn't seem credible.

Could you elaborate on the problems with Tegra 3?

Comparatively weak single-threaded performance, weak GPU performance, constrained memory interface, high power consumption when not running only the companion core.

Tegra 3 isn't terrible, but just like previous Tegra chips it was so hyped up before launch and again turned out to be among the weaker ARM chips of its generation. Exynos, Snapdragon S4, Apple A5 (and now A6) and OMAP 4 beat Tegra 3 on everything except synthetic multi-threaded benchmarks (because it has 4 cores, obviously).

I don't know about the GPU, my impression was that it was competitive (the single-channel memory interface probably doesn't help), but otherwise that's spot on. It really seemed to be a design whose main goal was to have the most cores for marketing purposes.

And of course they launched on a process that was just about to become obsolete, and haven't been able to do a shrink. Not sure why, though. Complications with the companion core? Just NVidia's general inability to get 28nm to work?

(Commercially it didn't help that they were unable to get access to a competitive LTE modem. That meant there was no shot at the US phone market. But that's not particularly a failing of NVidia, everyone but Qualcomm has that problem.)

I believe you're thinking of the Sun 3/80 (aka the "sun3x" architecture)

> I agree it certainly isn't a 'clean' comparison

The degree of compatibility seems to be a major variable, so this puts the comparison into doubt.

> I still remember a talk with both Bill Joy and Andy Bechtolsteim where Bill was over the top enthusiastic about finally having the ability to change the CPU to work better with the software rather than having the software work around what the CPU thought was best.

They should have been able to make Java VM implementers' lives a lot easier on Sun platforms. I guess the question was also, did they necessarily want to?

Man, I would think that Chuck M. would know, but this time, apparently not. The Sun 3/80 shipped simultaneous with the SS1, the 3/60 pre-dated the SS1 by several years.

(Jim "netgod" Thompson)

Thank you for the correction! Had I also recalled my 3/75 which was replaced by the SS1 I would have remembered the 80 vs the 60.

You're mistaken about alternate sources for SPARC. See Fujitsu as the primary example. SPARC architecture in general is licensed similarly to ARM.

That is true but in the early years Sun was not a 'good' licensor since they seems to compete with anyone who tried to build one. Fujitsu was a success story and arguably made better high end SPARC chips than Sun did. ARM Holdings on the other hand doesn't make any hardware at all so they are completely motivated to have their licensees be successful.

If Apple is really doing this, Intel can forget about ever putting Atom or any other of their chips inside the iPad. I know they've been hopelessly optimistic about it, but this is a clear sign it's just never going to happen.

Also, I still believe Apple will start making Macbook Air-like devices, maybe starting in 2014 with the Apple A8 (?) ARM-based chip, when ARM will have moved to the 64 bit ARMv8 architecture.

Well, you never know. If Intel continues to improve their mobile device x86 designs, they might surpass Apple and ARM's abilities some day - they do have a lot of architecture expertise and cutting-edge fab technology.

I still think it would be interesting to see Intel (or AMD) come out with a 64-bit only design for mobile / low-powered servers. I wonder how much space the 32-bit and older SIMD stuff takes on the chip?

Apple has been in the same position as Samsung for a while; both companies take other peoples' CPU and GPU core designs and integrate them into an SoC (A5 and Exynos, respectively). This would put them more in the same field as Qualcomm, which designs its own ARM cores using an ARM architecture license. It looks like this is what Apple is now doing too.

Kind of? There's been in house vlsi & soc at apple forever. Theyve had in house mobile CPU & graphics talent for a couple years at this point. Go check out the stuff Intrinsity was doing as a cash starved startup. Then extrapolate that same team being a well funded semi autonomous design group inside apple.

The Sun model ? Before the rise of the x86 based servers there pretty much evey major server vendor was doing this. IBM power - AIX, HP PA-RISC-AIX, Sun - you know already, Digital Alpha-Digital UNIX. Practically everybody had(and in the case of IBM still do) their own hardware including custom chips. The CPU world was way more diverse in the 90s and early 2000s. Sun was hardly the first or the only one.

Is Samsung making their own cores? The news in the article is that Apple is making their own cores now, they've been making their own silicon for the last few revisions, right?

Yep, the Exynos.

http://en.m.wikipedia.org/wiki/Exynos

The exynos is the SOC which integrates ARM cores (A8, A9, A15). The actual ARM cores within are straight from ARM.

It isn't clear to me that Samsung has publicly enhanced the ARM architecture in the Exynos [1] series but they certainly could and exploit that in their own products. Like Apple they are full licensees to the ARM architecture (which means they can make derivative works of the ARM cpu) and of course if they did they would be able to exploit that in their version of Android if they so chose to. There were a few companies that did that with the 6502 'core' back in the day adding some custom instructions in the space that was reserved or undefined by MOS Semiconductor.

[1] http://en.wikipedia.org/wiki/Exynos

It's really hard to build a custom processor with decent performance these days. There are very few teams that can do it, and it takes way more investment than normal ASIC design. The people who can do it largely work for Intel, AMD, IBM, some at Apple, and probably a few other major shops.

So I suspect it would be hard for Samsung to build a custom ARM processor unless they teamed up or bought a team that already has this specialized skill. That's what Apple did...

Samsung is making them, but they don't design their own.. well, they design the chips but not the arm cores themselves-- AFAIK only qualcomm (and now apple) have done this in the smartphone space.

Custom Silicon != Custom RTL