Big and little POWER shouldn't just be endian

Date: 2021-11-10T14:44:00-08:00

Location: www.talospace.com

While the majority of OpenPOWER installations by this point are probably running little-endian, every single POWER chip runs big — big power usage, that is. While POWER9 is still performance-competitive with

x86_64

and this situation continues to improve as more software gets better optimized, and there have been huge gains since POWER4/the PowerPC 970 in particular, POWER chips still run relatively hot and relatively hungry. Anandtech tried to normalize this for POWER8 systems by estimating transactions per watt; power measurements can be very imprecise and depend on more than just the system architecture, but even with that consideration the tested Tyan POWER8 in particular was outclassed by nearly a factor of three by a Xeon E5-2699. Possibly in response POWER9 is more aggressive with power savings than POWER8 and makes a lot of microarchitectural improvements, using 25% less juice for 50% more zip (so roughly a doubling of performance per watt), and Power10 supposedly improves on POWER9's performance per watt even more by at least 2.6 times according to IBM's figures.

But IBM's playbook for improving perf per watt hasn't really changed. Either you're boosting performance by juicing the microarch, jimmying IPC with more instructions and more cores, or both, or you're trying to diminish power usage with heavier clock speed throttling or turning off cores. While shooting the die budget at lower-wattage pack-in accelerators is a clever hybrid approach, their application-specific nature also means they're rather less useful in typical situations than their marketing would allege (look at how little currently uses the gzip accelerator in every POWER9, for example). You can do a lot with strategies like these — AMD certainly does — but sooner or later you'll hit a wall somewhere, either against the particular limitations of the design you're working with or against the intrinsic physical limitations of making a hippo do gymnastics while eating fewer calories.

Apple Silicon has a lot of concerning issues with it from a free computing perspective, but its performance is impressive, and its performance per watt is jaw-dropping. A lot of this is the secret sauce in their microarch which ironically came from P.A. Semi, originally a Power ISA licensee, and some may be due to details of the on-board GPU. But a good portion is also due to the big core-little core approach largely pioneered with the ARM big.LITTLE Cortex A7 and used to great effect in the M1 series. After all, if you want to get the best of both worlds, make some of the cores use less power and give those cores tasks that require less oomph (efficiency or E-cores), reserving the heavy tasks for the big ones (power or P-cores). Intel thinks so too: Lakefield and Alder Lake both attempt the same sort of heterogenous CPU topology for x86_64, and it would be inconceivable to believe AMD isn't looking to make the same jump for their next iteration.

The chief issue with going that route is making sure that the cores are getting work commensurate with their capabilities. This is easy for Apple since they control the whole banana: macOS Quality of Service is all about doing just that (you'd think they would do something based on nice levels as well, but I guess all the sweet talk about being desktop Un*x went out the window somewhere around Mavericks). Linux added initial support for big.LITTLE with kernel 3.10 but it took years for other improvements to the Linux scheduler to make it meaningful. Intel made things worse for themselves in Lakefield and Alder Lake by using lower power Atom-based E-cores that didn't support AVX-512 (and the Tremont E-cores in Lakefield didn't even support AVX2, meaning such tasks couldn't be run by them at all). Rather than hinting Windows 11 or the internal hardware not to send AVX-512 code to the Gracemont E-cores, Alder Lake just doesn't support AVX-512, full stop — on any core. Kernel 5.13 supports Alder Lake, but kernel 5.15 has dawned and there is no specific Intel Thread Manager Support so far, though there is scheduler support for AArch64 E-cores that can't run 32-bit code. And Alder Lake is turning out to be very power-hungry, which calls some of the design into question, in addition to various compatibility issues when software unwittingly puts tasks on the E-cores that don't work as expected.

Still, the time is coming where Power ISA should start thinking about a big-little CPU, maybe even for Power11. We already have big cores (if IBM will ever get their heads out of their rear ends and release the firmware source), but we also have an already extant little OpenPOWER core: Microwatt. While Microwatt doesn't support everything that POWER9 or Power10's large cores do, it's still intended to be a fully compliant OpenPOWER core, and since the Linux kernel is already starting to cater to heterogenous designs a set of POWER8-compliant Microwatt E-cores could still execute on the same die along with a set of Power11 full fat P-cores. Add logic on-chip to move threads to the P-cores if they hit an instruction the E-cores don't support and you're already most of the way there with relatively minor changes to the Linux kernel.

What IBM — or any future OpenPOWER chip builder, though so far no one else is in the performance category — needs to avoid is what seems to be dooming Alder Lake: they've managed to hit the bad luck jackpot with a chip that not only uses more power but has more compatibility problems. Software updates will fix this issue somewhat but a little more forethought might have staved it off, and the apparent greater wattage draw should have been noticed long before it left the lab. But IBM has already shown wattage improvements over the last two generations and if the P- and E-core functionalities are made appropriately comparable, a big-little Power11 — with open firmware please! — could be a very compelling next upgrade for the next generation of Power-based workstations and servers. Apple has clearly demonstrated that highly efficient and powerful computing experiences are possible when hardware and software align. There's no reason OpenPOWER and Linux or *BSD can't do the same on open platforms.