To nvidia or not to nvidia...

Very cool, Richard Cranium. It's good to see it can run at all and I am curious to see if nouveau actually does better as I suspect on older design and/or lower performance chipsets. I'd like to know if you installed Mesa or GLX libraries since I made no changes to those other than those the nvidia installer effected. Incidentally I bought one of those 2GB 640s for my son and was rather knocked out with it's bang for the buck. Terrific value. I ended up buying another for a secondary boxen of mine.

NOTE: - In my last post I just now noticed the platforms. The SlackLive/Nouveau was 64bit but the hdd install running the nvidia driver was 32bit. I doubt though that this accounts for much with such a huge difference but in the interest of full disclosure I had to point that out..

Nvidia only really supports the latest drivers well, their legacy kernels can be subpar at times. On the other hand nouveau explicitly supports older cards that nvidia no longer will, but without documentation there are a few old issues that have been still unsolved. A good example is that nouveau does not do well with multiple GL contexts at once and will crash or freeze the program and sometimes the entire system.

Originally Posted by Richard Cranium When I run Unigine Heaven Benchmark 4.0 with the nouveau driver, the benchmark results are: FPS: 1.9 Score: 49 Min FPS: 1.4 Max FPS: 3.3 System: Platform: Linux 4.4.132 x86_64 CPU model: AMD Phenom(tm) II X6 1100T Processor (3409MHz) x6 GPU model: Unknown GPU (256MB) x1 The card is actually an ASUS GT-640 2GD3 Settings: Render: OpenGL Mode: 1920x1080 2xAA fullscreen Preset: Custom Quality: High Stereo: Horizontal Tessellation: Normal Now I'm curious. I'll install the nvidia drivers and see what the delta turns out to be.

And the answer is...

Unigine Heaven Benchmark 4.0

FPS:	9.6
Score:	242
Min FPS:4.2
Max FPS:17.9


System

Platform:	Linux 4.4.132 x86_64
CPU model:	AMD Phenom(tm) II X6 1100T Processor (3409MHz) x6
GPU model:	GeForce GT 640 PCI Express 390.59 (2048MB) x1

Settings

Render:	OpenGL
Mode:	1920x1080 2xAA fullscreen
Preset	Custom
Quality	High
Stereo	Horizontal
Tessellation:	Normal

The benchmark was able to fully identify the card, for one thing.

Does nouveau reclocking work for your card? It requires the debugfs to be built into the kernel to test.

For example:

# cat /sys/kernel/debug/dri/0/pstate
07: core 324 MHz memory 648 MHz AC DC *
0a: core 324-692 MHz memory 1620 MHz
0d: core 549-1137 MHz memory 7000 MHz
0f: core 549-1137 MHz memory 7000 MHz
DC: core 324 MHz memory 648 MHz

# echo 0f > /sys/kernel/debug/dri/0/pstate

# cat /sys/kernel/debug/dri/0/pstate
07: core 324 MHz memory 648 MHz
0a: core 324-692 MHz memory 1620 MHz
0d: core 549-1137 MHz memory 7000 MHz
0f: core 549-1137 MHz memory 7000 MHz AC DC *
DC: core 1072 MHz memory 6999 MHz

The first 4 lines represent the different reclocking states and the last line shows the current state. This could help the performance with nouveau a lot. However I think the kernel in 14.2 might be too old, I recall there was a lot of improvements to reclocking around 4.10 (?). It should be pretty trivial to build the kernel from current with Pat's config for 14.2 to test this however.

This helper function could also be added to /root/.bashrc to make it easier to control.

pstate='/sys/kernel/debug/dri/0/pstate'

pstate () {
  if [ "${1}" ] ; then
    case "$1" in
      -1) echo 07 > "$pstate" ;;
      -2) echo 0a > "$pstate" ;;
      -3) echo 0d > "$pstate" ;; 
      -4) echo 0f > "$pstate" ;;
      -l) cat "$pstate" ;;
      -h) echo 'pstate [-1][-2][-3][-4][-l][-h]' ;;
       *) echo 'unrecognized option, use -h for help' ;;
    esac
  else
    echo 'use -h for help or -l to see the pstate file'
  fi
}

# pstate -l
07: core 324 MHz memory 648 MHz AC DC *
0a: core 324-692 MHz memory 1620 MHz
0d: core 549-1137 MHz memory 7000 MHz
0f: core 549-1137 MHz memory 7000 MHz
DC: core 324 MHz memory 648 MHz

# pstate -3

# pstate -l
07: core 324 MHz memory 648 MHz
0a: core 324-692 MHz memory 1620 MHz
0d: core 549-1137 MHz memory 7000 MHz AC DC *
0f: core 549-1137 MHz memory 7000 MHz
DC: core 1072 MHz memory 6999 MHz

Note the values can differ between cards so its important to not simply copy/paste this function and make sure the values are correct first.

Thank you, Richard Cranium, for your efforts to discover and provide some actual numbers in the area of accelerated 3D performance on a "middling" performance chipset.

Originally Posted by Richard Cranium And the answer is...

Very interesting test and results. What a disparity.

Originally Posted by enorbet Thank you, Richard Cranium, for your efforts to discover and provide some actual numbers in the area of accelerated 3D performance on a "middling" performance chipset.

Thanks, I think.

Originally Posted by orbea Does nouveau reclocking work for your card? It requires the debugfs to be built into the kernel to test. For example: # cat /sys/kernel/debug/dri/0/pstate 07: core 324 MHz memory 648 MHz AC DC * 0a: core 324-692 MHz memory 1620 MHz 0d: core 549-1137 MHz memory 7000 MHz 0f: core 549-1137 MHz memory 7000 MHz DC: core 324 MHz memory 648 MHz # echo 0f > /sys/kernel/debug/dri/0/pstate # cat /sys/kernel/debug/dri/0/pstate 07: core 324 MHz memory 648 MHz 0a: core 324-692 MHz memory 1620 MHz 0d: core 549-1137 MHz memory 7000 MHz 0f: core 549-1137 MHz memory 7000 MHz AC DC * DC: core 1072 MHz memory 6999 MHz The first 4 lines represent the different reclocking states and the last line shows the current state. This could help the performance with nouveau a lot. However I think the kernel in 14.2 might be too old, I recall there was a lot of improvements to reclocking around 4.10 (?). It should be pretty trivial to build the kernel from current with Pat's config for 14.2 to test this however. This helper function could also be added to /root/.bashrc to make it easier to control. pstate='/sys/kernel/debug/dri/0/pstate' pstate () { if [ "${1}" ] ; then case "$1" in -1) echo 07 > "$pstate" ;; -2) echo 0a > "$pstate" ;; -3) echo 0d > "$pstate" ;; -4) echo 0f > "$pstate" ;; -l) cat "$pstate" ;; -h) echo 'pstate [-1][-2][-3][-4][-l][-h]' ;; *) echo 'unrecognized option, use -h for help' ;; esac else echo 'use -h for help or -l to see the pstate file' fi } # pstate -l 07: core 324 MHz memory 648 MHz AC DC * 0a: core 324-692 MHz memory 1620 MHz 0d: core 549-1137 MHz memory 7000 MHz 0f: core 549-1137 MHz memory 7000 MHz DC: core 324 MHz memory 648 MHz # pstate -3 # pstate -l 07: core 324 MHz memory 648 MHz 0a: core 324-692 MHz memory 1620 MHz 0d: core 549-1137 MHz memory 7000 MHz AC DC * 0f: core 549-1137 MHz memory 7000 MHz DC: core 1072 MHz memory 6999 MHz Note the values can differ between cards so its important to not simply copy/paste this function and make sure the values are correct first.

I may fall back and try that in a couple of days. Thanks for the info! I didn't realize that level of tweaking was possible.

And I ran Blender with cuda support.

1. The result was really, really fast.
2. The result was also low-quality crap rendering. I mean 256 colors bad rendering.

I find the Blender UI to be almost a crime but I haven't done very much with it; lots of people love it, so I assume I'm the outlier. I assume there's something I can tweak to improve the rendering quality. Any pointers?

As an aside, nouveau reclocking is not available with Slackware64 14.2's kernel.

When the next version of Slackware is released, I'll re-run the test and modify the nouveau clock speed if possible.