Well Known Linux Kernel Developer Recommends Against Buying Skylake Systems

[s_j_newbury]

I don't know where people get the idea that low-end Xeons are faster than POWER8 chips. I know Anandtech did a benchmark comparison between POWER8 and mid-end Xeons which favoured Xeons, especially for "bang-for-the-buck", but even then it wasn't so clear-cut for ultimate performance, and the benchmark was seriously flawed in using the "new" Ubuntu 14 LTS for POWER8 LE distribution, any sensible deployment of POWER8 would be on BE mode. It also compared and conflated per-core performance with overall system performance, while interesting, is completely not-the-point for a 8-way SMT system like POWER8. I'm also of the opinion; it's necessary to build a distribution with a current POWER8 toolchain release, even if that means a source-based distribution if there isn't yet an optimised tier-one Linux distribution given the relative newness of the platform.

All that said, "Power Management" is much better with Intel, IBM builds their hardware to run optimally 100% maximum utilisation 24/7, for years, with hot-replacement of failing parts, so power management is just a nice thing to have, while Intel design philosophy seems to work to a power/thermal budget where peak performance can not be maintained long-term without seriously curtailing system longevity, so the hardware actively manages performance states. If you want to design a laptop/mobile workstation, right now, unless IBM creates a Mobile POWER8 with Intel-like PM capabilities there's sadly little choice. That probably also applies to any server/workstation deployment that spends significant time idling or running at something less than 100% capacity. I guess we'll see how well third-parties do squeezing idle efficiency from the platform, it's not something I'd expect to see from IBM.

[e8hffff]

People need to realise that Intel is in the pockets of Microsoft. Don't doubt it they have Code-Saboteurs in Linux community and also deliberate lack of support.

We all need to lean toward ARM and maybe NVidia/AMD for our future computing.

[AdamOne]

Always support the cooler competitor. Intel is a smaller scale, wannabe Micro-take-a-hike-ro-soft piece of sh&% douchebag hardware maker.

AMD, eventtough they've been a little slow to support open source, community development, deepdown has a way cooler disposition towards open source.

NVIDIA are pretentious douchebags too. #AMDforthemin #EffIntel #NVIDIAEffYOU

[mjg59]

I'm not sure how to interpret these. Can the blog author verify if this laptop is affected?

The only Package C state you're getting into is PC2, so yes, you're affected.

[H.F.P]

I wasn't planning anyway since these days they keep throwing at us binary blobs so they can put their crap in there.
I think I will go with a previous version or change sides.

Are you paid to repeat this over and over again? It's pretty known (for informed people) that Special Interest Groups (SIGs) are causing blobs. Any HW vendor who wants to support a
(SIG-)blessed (a.k.a. proprietary) codec needs to adhere to the rules. The latter of which are causing us these beloved binary blobs.

Does this sound reasonable for you?
Your reasoning right now looks like attributing "blobs" to Intel like the last idea they had to become evil. This is not only (historically) weird but shooting ("throwing") against an unrelated wall.

[pal666]

Awesome guy. Explains that Skylake has a feature that Linux doesn't properly implement and then concludes that Skylake is bad. F*ck logic.

main linux kernel developer and skylake seller is the same company. so you fucked either your logic, your knowledge, or both

[chadotter]

I'm not using the system now, but I previously tested Ubuntu 16.04 Beta on a Latitude 7275 with a Core m5 processor and was able to get to the lower package C-states.

[eggbert]

I have a Dell Inspiron 7559. This has a skylake core i7 in it. When I look in powertop idle stats, under package I see C10-SKL is around 89%, so apparently, it seems to be working here (unless I misunderstanding what I am looking at, which is entirely probable).

One thing however with this laptop under any kernel I've tested is that the fans never stop, and it runs pretty warm. Whereas on Windoze, the fans hardly ever come on and it stays nice and cool.

I regret getting a skylake laptop.

[edwaleni]

I don't know where people get the idea that low-end Xeons are faster than POWER8 chips. I know Anandtech did a benchmark comparison between POWER8 and mid-end Xeons which favoured Xeons, especially for "bang-for-the-buck", but even then it wasn't so clear-cut for ultimate performance, and the benchmark was seriously flawed in using the "new" Ubuntu 14 LTS for POWER8 LE distribution, any sensible deployment of POWER8 would be on BE mode. It also compared and conflated per-core performance with overall system performance, while interesting, is completely not-the-point for a 8-way SMT system like POWER8. I'm also of the opinion; it's necessary to build a distribution with a current POWER8 toolchain release, even if that means a source-based distribution if there isn't yet an optimised tier-one Linux distribution given the relative newness of the platform.

All that said, "Power Management" is much better with Intel, IBM builds their hardware to run optimally 100% maximum utilisation 24/7, for years, with hot-replacement of failing parts, so power management is just a nice thing to have, while Intel design philosophy seems to work to a power/thermal budget where peak performance can not be maintained long-term without seriously curtailing system longevity, so the hardware actively manages performance states. If you want to design a laptop/mobile workstation, right now, unless IBM creates a Mobile POWER8 with Intel-like PM capabilities there's sadly little choice. That probably also applies to any server/workstation deployment that spends significant time idling or running at something less than 100% capacity. I guess we'll see how well third-parties do squeezing idle efficiency from the platform, it's not something I'd expect to see from IBM.

We tried to get RHEL on POWER8 to help us get off AIX long term, but the POWER snobs wouldn't have it. With 8 threads per core on POWER and less than 2 with Xeon, it was going to cost us a bundle to get enough cores to transfer the workload. The additional cost of the software licenses was worse. All those additional Xeon cores to license was adding $5m to the tab. Those POWER7's have been running nearly 100% for 4 straight years and are running at a huge queue depth now. So I agree with the poster, no power management needed to protect their useful life.

Now if I could just get an openPOWER box in my hands and test. I would like to see the workload per watt up close.

[Kano]

Your server workloads are not relevant for desktop users. Also for current Flash or HTML 5 with DRM you need a x86-64 CPU. Try watching Netflix with Power8, have fun. I don't have got Skylake myself but basically i am more interested in Kabylake (due to HEVC Main 10 hardware support). It should be very similar so it would be nice if that primary mobile bug would be fixed. Also it would be interesting if disabling p-state changes something. Skylake has got a new technology handling p-states - before the OS needed to set lower P2+ states (P1 is default, P0 turbo) and now the CPU itself can switch, this change is much faster but may not be faster for all workloads but "should" save power. But maybe the p-state driver does not correctly identify mobile Skylake, that's where I would look. As you see in

https://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/drivers/cpufreq/intel_pstate.c?id=7791e4aa59ad724e0b4c8b4dea547a5735108972

This feature should show HWP in dmesg...