Just as products based on ARM's much anticipated Cortex A9 are finally poised to hit the market, the company has announced yet another, even higher-end core design: the A15. Codenamed "Eagle," the A15 architecture is ostensibly aimed at netbooks and tablets, but a look at the spec sheet leaves no doubt that ARM is absolutely gunning for the server market that Intel and AMD currently dominate. Indeed, even going by what little ARM has revealed about the A15, it's very hard to imagine this thing in a smartphone when it launches at 32nm in 2012 or 2013. This is a laptop and server part, and ARM will use it to take the fight to x86.
The overall position of the A15 in ARM's lineup is as the next logical step up the ladder from the A9, so that in order of core size, complexity, performance, features, capability, and power consumption, the Cortex family goes by the numbers: A8 at the bottom, then A9, then A15 at the top. The A8 is an in-order part with a very simple architecture that's comparable to Intel's Atom in some key anatomical respects, but is much lower-power and more efficient than the latter. The Cortex A9 is an out-of-order part that brings the ARM line into Atom's performance territory, and also closer to Atom in power draw (though the A9 is allegedly still much more efficient than Atom).
The A15 takes things to the next level by pairing the out-of-order nature of the A9 with the expansive feature list (virtualization support, double-precision floating point, ECC cache) that characterizes the Atom line, and that costs Atom in terms of power. In this respect, the A15 looks somewhat like AMD's recently unveiled Bobcat core, to the point that I would be very curious to see a comparison of the transistor counts of the two cores.
ARM has yet to release a real block diagram of A15, so there are many crucial aspects of the design that are very unclear. Specifically, the capability and robustness of the A15's branch prediction hardware, the depth of its pipeline, and the size and configuration of its instruction window will all have a direct impact on both raw performance and the all-important performance/watt metric. If ARM has beefed up the size of the instruction window and/or increased the design's branch prediction resources (possibly to mitigate the effects of a longer pipeline), then that will be performance positive but power negative.
Not a successor to A9, but a big brother
The fact that A15 is such a significant step up the performance/complexity/power ladder from A9 makes it clear that A15 isn't really a successor to A9, any more than A9 is a successor to A8. Rather, the three designs will coexist in ARM's lineup for many years, enabling the Cortex family to span a range of power/performance/feature points. So the difference between the A8, A9, and A15 parts, and, say, Bobcat and Bulldozer, is that ARM, being a much smaller company than AMD with very limited resources and a different historical user base, has trickled the three designs out slowly over the course of a few years, vs. putting them out them all at once.
The leisurely pace of ARM's roadmap will come as a surprise to Intel and AMD watchers, but as with all embedded designs, ARM architectures historically have very long lifecycles. (This is true of MIPS, as well.) Older ARM cores that aren't part of the Cortex family are still in production in a range of devices, from feature phones to appliances—the same will be true of the Cortex family in a decade.
Forget netbooks: this is a laptop and server part
Let's run down some of the important distinguishing features of A15:
- Clockspeed of up to 2.5 GHz
- 1, 2, 4 or 8 cores
- Support for 1TB main memory
- ECC L1 and L2
- Fully cache coherent bus protocol (for multisocket systems)
- Support for virtualization
When you add the above to some of the other features, like double-precision floating-point, vector extensions, cache snooping, and so on, it's very clear that this is at the very least a laptop part—not phones, or netbooks, but laptops. And not just laptops, but servers. And then again not just servers, but high-density multisocket servers.
So with the A15, ARM finally makes the jump all the way into niches currently occupied by Intel and AMD. With respect to netbooks—or "smartbooks" as they're called when they have an ARM part in them—the line between a netbook and a laptop isn't super clear anymore. But wherever the line is, anything based on A15 will be on the laptop side of it. And in the datacenter, the part will also strengthen ARM's burgeoning presence in the nascent market for high-density, low-power cloud servers. Over the next two years, the A9 will establish ARM's presence in this new niche, so that when the A15 drops the company will (ideally) use it to build on its existing presence.
All told, the A15 looks like a promising complement to the A9 and A8, but it's too early to say much more than that. The part is well over a year from the market, if not two or three. And when it debuts at 32nm Intel will be at 22nm, and with no telling what kind of successor to today's Atom. AMD will still be working with the basic Bobcat technology, and will probably be at 32nm. This means that all we can do is make vague guesses as we wait for more A15 details to trickle out. But despite the paucity of detail about A15's real prospects, one thing is for certain: the processor scene just got a lot more interesting.
Source: ars technica
