NVIDIA SLI: Linux vs. Windows

[phoronix]

Phoronix: NVIDIA SLI: Linux vs. Windows

It is going on two years since support for Scalable Link Interface (SLI) was introduced into NVIDIA's Linux binary display driver. This support had come a year after it was officially launched and supported by the Windows ForceWare display driver. As we had seen at the end of 2005 with two GeForce 6 graphics cards in SLI, its performance was very sluggish, and there were a number of problems to be found with Linux SLI. While we have routinely tested new NVIDIA graphics cards under Linux SLI internally, there hasn't been much to report on as the experience has been very foul. However, things have changed recently and with the recent NVIDIA 100.14.19 display driver release using GeForce 8 hardware -- we finally have some modest numbers to report on in a Linux SLI configuration. Linux SLI is still far from perfect, but in this article we've used two GeForce 8600GT graphics cards in an SLI configuration under both Linux and Windows to compare the single and dual GPU performance under both operating systems.

http://www.phoronix.com/vr.php?view=11147

[Throwing Strikes]

Figures :mad: I noticed performance gap between my OpenSuSe 10.2 and Winblows XP but hey two is always better than one. :rolleyes:

Way to keep em' strait, maybe this will put alittle more pressure on Nvidia to make even better drivers for Linux.

[Thetargos]

The performance delta with G80 hardware between Windows and Linux is immense! I don't remember it being as big with the G70 generation, all the contrary, if anything on some benchmarks, the Linux driver came up faster than the Windows one (not for SLI, though, which has always been orders of magnitude better on Windows).

[Michael]

Digg LInk: http://digg.com/linux_unix/NVIDIA_SLI_Linux_vs_Windows

[Michael]

The performance delta with G80 hardware between Windows and Linux is immense! I don't remember it being as big with the G70 generation, all the contrary, if anything on some benchmarks, the Linux driver came up faster than the Windows one (not for SLI, though, which has always been orders of magnitude better on Windows).

Right, with previous generations the delta was minimal and in some cases the GeForce 6/7 part would be faster on Windows.

[Source Of Radiance]

On my Geforce7 power system. Nexuiz consistently benchmarks faster than on Windows by a hair. Looks like GF8 scales pretty well with SLI, so if the performance delta is reduced even further it'll be almost perfect.

Has Geforce 7 SLI in linux improved since the last time it was tested? I don't have a second 7600gt to test this.

[yossarianuk]

On my machine Linux constantly has faster benchmarks than windows. (geforce 8500)

I have benchmarked (although do not figures to hand) and all these a games are faster (and smoother) in Linux:-

- enemy territory
- nexuiz
- tremulous
- paintball2
- doom3
- quake 4
- call of duty (through wine)

I have an amd64 machine but run 32 as well as 64 bit distros.

The fastest benchmarks have been with gentoo64bit - the most notable difference is that I can max the settings (i.e FSAAX16, etc) and still play with good frame rates (except quake 4..)

The 32 bit distros are always faster than windows too.

One thing I always do is recompile the kernel with the below features - it may also be of note the ID software released a statement saying that to play Quake wars on Linux these features should be enabled.
(http://www.linux.com/?module=comments&func=display&cid=1163938)

- Voluntary preemption
- Timer frequency (1000 HZ)

Does fedora have these features by default ? I know that opensuse doesn't (it uses 250HZ)

[Thetargos]

Fedora since Core 6 switched to 1000Hz kernel by default, however I'm not sure it uses voluntary preemption or low-latency Desktop preemption settings... lemme check:

[gianni@Blackhawk ~]$ grep PREEMPT /boot/config-2.6.22.5-76.fc7
# CONFIG_PREEMPT_NONE is not set
CONFIG_PREEMPT_VOLUNTARY=y
# CONFIG_PREEMPT is not set
CONFIG_PREEMPT_BKL=y


Yup, it does look like Fedora uses voluntary preempt :D

And, yup, it does use 1000Hz tickrate:

[gianni@Blackhawk ~]$ grep HZ /boot/config-2.6.22.5-76.fc7
# CONFIG_HZ_100 is not set
# CONFIG_HZ_250 is not set
# CONFIG_HZ_300 is not set
CONFIG_HZ_1000=y
CONFIG_HZ=1000

[1c3d0g]

I'm wondering why nobody uses tickless kernels...they should be included by most distro's by now, shouldn't they?

[Thetargos]

Th next update for the Fedora kernel (2.6.22.7-85.fc7) uses a tickless kernel for i686. I'm not sure how does that help the overall responsiveness of the system,. though. Will have to test drive one, I guess :D

[Moustacha]

Th next update for the Fedora kernel (2.6.22.7-85.fc7) uses a tickless kernel for i686. I'm not sure how does that help the overall responsiveness of the system,. though. Will have to test drive one, I guess :D

Hmm...You sure?
$ uname -a
Linux sams.computer 2.6.22.7-85.fc7 #1 SMP Fri Sep 21 19:53:05 EDT 2007 i686 athlon i386 GNU/Linux
$ grep HZ /boot/config-2.6.22.7-85.fc7
CONFIG_NO_HZ=y
# CONFIG_HZ_100 is not set
# CONFIG_HZ_250 is not set
# CONFIG_HZ_300 is not set
CONFIG_HZ_1000=y
CONFIG_HZ=1000
CONFIG_MACHZ_WDT=m

[Thetargos]

I'm sure I saw the option just yesterday afternoon when I was creating a custom config (using menuconfig) using the stock configuration as a template (for precisely 2.6.22.7-85.fc7) I'm not sure, then if 2.6.22.9-91.fc7 does so.

Aaahhh... Now this is strange. The CONFIG_NO_HZ option supposedly turns the kernel into tickless (it is not present in the x86_64 kernel, there are still issues with tickless and x86_64), I don't know why is the CONFIG_HZ_1000 still set as 'y', though.

[niniendowarrior]

I haven't read the article, but this sounds a bit like rubbing salt on the wounds. hehehe. :D

[jimhood82]

Am I the only one to notice that they only used 1 game to compare Linux Vs Windows? Is this because someone got too lazy to test others (like the Doom3 they did the initial Linux SLI review with) or because the numbers did not promote Windows? I have just finished setting up my Linux SLI system - and will do some personal testing, and may even post some results... but this seems like very poor reporting to me.

If you are going to test and compare, Do it right. I have seen major performance differences like those just from going from XP to XP SP2! OR even from one driver to the next!

I am sorry that my first post on this forum is not a positive one, but no one else mentioned the obvious... ONE game is not PROVING anything.

[RealNC]

Holy thread resurrection grave digging :D

[jimhood82]

Holy thread resurrection grave digging :D

Sorry bout that; was looking for info on the whole SLI in Linux bit... came across this article... despised it, and replied... didn't think to check the date on it.

[Bill Cosby]

So what about your tests?

Has SLI performance increased on Linux?

I am planning on going for SLI :)