KSP and CPU instruction sets?

[iSilver 7]

Does KSP utilize AVX, AVX2, and FMA3/FMA? Or still on SSE2/SSE4 codes? I'm thinking of getting a new cpu optimized for some physics based games like KSP and BeamNG.

[sal_vager 5,842]

Hi @iSilver, KSP is limited to what the Unity3D engine supports.

Quote

For running Unity games

Generally content developed with Unity can run pretty much everywhere. How well it runs is dependent on the complexity of your project. More detailed requirements:

• Desktop:
  • OS: Windows XP SP2+, Mac OS X 10.8+, Ubuntu 12.04+, SteamOS+.
  • Graphics card: DX9 (shader model 3.0) or DX11 with feature level 9.3 capabilities.
  • CPU: SSE2 instruction set support.

Hope this helps.

[iSilver 7]

6 hours ago, sal_vager said:

Hi @iSilver, KSP is limited to what the Unity3D engine supports.

Hope this helps.

Thank you @sal_vager

Turns out I don't need those latest extensions yet, keeping my 3570k for now.

Much appreciated.

[iSilver 7]

I've just found out that Unity 5.x Physx 3 actually runs with two code paths (SSE2 for older/value cpus and AVX for newer/mid-to-high cpus).

[sal_vager 5,842]

46 minutes ago, iSilver said:

I've just found out that Unity 5.x Physx 3 actually runs with two code paths (SSE2 for older/value cpus and AVX for newer/mid-to-high cpus).

Interesting! Would you happen to have a link?

[iSilver 7]

http://physxinfo.com/wiki/PhysX_SDK_3.x Deformables section.

[sal_vager 5,842]

Ahh I see, I don't know if the version of PhysX in Unity does all of that, one difference with the Unity version that I know of is that it only provides software physics at this time, except for a handful of optional things such as hair rendering, but SSE2 can be taken as a minimum.

[iSilver 7]

I ran several fps tests with a 175-part rocket+lander and no noticeable difference between AVX (normal) and disabled AVX (xsavedisable 1: disabled AVX in Windows 10). Both result in consistent fps range of 60 to 70 (unlocked framerate & no vsync). My 3570k doesn't support AVX2 though.

[sal_vager 5,842]

1 hour ago, iSilver said:

I ran several fps tests with a 175-part rocket+lander and no noticeable difference between AVX (normal) and disabled AVX (xsavedisable 1: disabled AVX in Windows 10). Both result in consistent fps range of 60 to 70 (unlocked framerate & no vsync). My 3570k doesn't support AVX2 though.

Makes sense, neither the Unity manual or the Unity 5 blog post mention AVX at all, so it's likely unavailable.

[Padishar 1,536]

19 hours ago, iSilver said:

http://physxinfo.com/wiki/PhysX_SDK_3.x Deformables section.

Given the only mention is in the deformables section, which KSP doesn't use, I wouldn't expect anything in KSP/Unity to be any faster.

14 hours ago, iSilver said:

and disabled AVX (xsavedisable 1: disabled AVX in Windows 10)

The MS documentation for this option indicates that it disables the use of the XSAVE instruction in the Windows kernel.  Nothing to do with AVX as far as I can see, it concerns the saving of the processor state so is only relevant to CPU context switches.

The bit of software on your machine most likely to use AVX is your graphic card driver (or video encode/decode stuff but that's irrelevant to KSP) but even this is unlikely to make any significant difference as it just won't get used enough to be significant.  If only 1% of the time is spent doing things using AVX instead then you will never see more than a 1% improvement (and probably more like 0.25%).

16 hours ago, sal_vager said:

but SSE2 can be taken as a minimum.

Indeed, SSE2 is an absolute requirement for Unity 5 (and probably 4 too) games.  SSE2 was introduced in the first Pentium 4 CPUs so I really wouldn't want to play KSP on a machine without it, even if it were possible...

[iSilver 7]

7 hours ago, Padishar said:

The MS documentation for this option indicates that it disables the use of the XSAVE instruction in the Windows kernel.  Nothing to do with AVX as far as I can see, it concerns the saving of the processor state so is only relevant to CPU context switches.

It does disable the cpu extended states (AVX+, FMA, F16C) by disabling the use of xsave instruction.

Here, with AVX enabled and disabled AIDA64 FPU mandel (double-precision fp) benchmark results:

AVX on (normal windows):

 

AVX off (xsavedisable 1):

 

CPUID bit readings:

Just run any AVX/AVX2/FMA optimized benchmark with both settings.

[Padishar 1,536]

32 minutes ago, iSilver said:

It does disable the cpu extended states

I stand corrected, thanks for that.  It would be nice if MSDN said something about it actually disabling the AVX instructions in the CPU but it isn't the first bit of MS documentation that's wrong or misleading...