A look beyond x86: OpenPOWER and AArch64
lvalsan.web.cern.ch
lvalsan.web.cern.ch
- Power consumption of Intel systems at off, idle and full power states are less then of ARM64 and POWER8 ones
- ARM 64-bit and Power 8 are currently providing a significantly lower performance / Watt than Intel
Damn.
For memory-bandwidth-limited problems (such as you often see with HPC), power8 looks better.
Xeons consume less power even when off...? Am I not parsing this sentence correctly?
Xeon: 0W
OpenPOWER: 11W
HP Moonshot (Avoton): 121W
HP Moonshot (ARM): 218W
This is mostly due to chassis specifics, though.
Then again, servers are not usually turned off so perhaps they just kept all the CPUs powered up but in a halted state...
The purpose of the off power measurements is to see how much power the chassis infrastructure is drawing.
The power consumption of the Intel Xeon E3-1285L v3 (Haswell) chassis with the system off is very low, below the threshold of the Racktivity smart metered PDU used to perform the power measurements. The PDU reports 0, but the real power consumption is in the order of a few Watts.
The tested Power 8 processor is inside a Tyan reference system GN70-BP010. Being a reference system, not all components and firmwares have been fully optimised. In the mean time Tyan has launched the TN71-BP012 platform which is not a reference system and should be much more optimised and ready for production use. I would be curious to see what are the improvements, but I don't have any at hand for now.
So Intel wins on Fabs Tech. And this will not change for another 5 years.
This has enormous power implications, and given that they're pretty close in power consumption with a huge difference in the fab process, Aarch64 is a strong contender as an architecture.
Same for ARM, but I guess the sweet spot there is scale out.
Intel has been putting out amazing chips since Haswell.
We need the others to do well to keep them honest and innovative.
TLDR; These things are for a specific segment of HPC, albeit the one that doesn't care too much about Linpack performance.
[1] http://en.wikipedia.org/wiki/POWER8 [2] http://www.pugetsystems.com/blog/2014/09/08/Memory-Performan...
They've done this in the past.. PowerPC (6xx), PReP, CHRP, PowerLinux, now OpenPOWER. It looks too similar this time where they cannot get things into gear lowering margins to gain volume. The Tyan box is basically a reference platform.. over priced and poorly designed (2U, 3.5" disks, several thousand dollars?!). Crossing my fingers something good will happen, but placing little stock in it.
But, what I really came here to say, is that POWER8 has nearly twice the single thread memory bandwidth available. From some of my tests a single thread can access ~40GB/sec of bandwidth vs the intel at ~20GB/sec. But this doesn't translate into clear cut wins even on codes which at first pass appear to be bandwidth limited. Mostly because the processor cores themselves are about 1/2 the speed in tight loopy code (think nbench) running out of cache.
So, in our in house benchmarks the POWER ran anywhere from 75% as fast to roughly as fast. Only on a few tests did it actually manage to best the E5 xeon we were testing it against, and then never by more than 150%.
Plus, the dynamic thread controls are cool, but beyond 4 threads/core I could never actually improve performance, and beyond 2x it was major diminishing returns.
I wonder where the bottleneck is there, since on paper the POWER8 specs beat Haswell: it's a 10-issue core, 224-entry reorder buffer vs. a 4-issue core with 192-entry reorder buffer for Haswell. But if it takes several POWER instructions to do the work of one x86 instruction, or icache bandwidth is insufficient, the advantage goes the other way.
It'd be interesting to hear which company it is -- I don't really know who uses POWER these days... email in profile in case you're more comfortable sharing that privately. I of course understand if you feel you can't do that.
The gap should close as the process technologies get closer.