The Ultimate 686 Benchmark Comparison

[Callahan]

Totally awesome work!
Awesome test results!
Thanks for hard work!


I never would have expected that the top will be occupied mostly by amd ss7 processors.

I have to dust off my ss7 boards

[meljor]

Long live my k6-3+ !


Did you encounter any stabillity problems during testing? I wonder because when i had my first pentium era systems (when they were still tha bomb) i allways had problems and i really hated windows 95 and 98(se) because of the many lockups and bluescreens. Gaming was fun back then but the many clean installs wasn`t.

Now that i have my retro systems i barely have problems. I mostly use Asus boards now and good psu`s, i guess that really makes a difference: back in the days i used cheap pcchips boards etc. and didn`t know anything about psu`s.

I know for a fact that my first Intel cpu was way more stable than my cyrix, but it must have mostly been about cooling? Didn`t know that in the days either so that was probably the problem....

[meljor]

btw, i love my k6-3+ @ 600mhz system (it`s my favorite) but i gets murdered in every single benchmark i do by my p3-700. In most cases it has a VERY hard time coming even close to my p3-450 cpu.....

Is it the software you are using? It`s perfectly fine for games of that era but the k6 series fpu really isn`t up to par compared to p3. Integer is boss tough.

Only when 3dnow! is used properly it gets high scores (3dmark 99 and ofcourse Quake2)

[feipoa]

I encountered stability problems with many of the processors above their rated frequencies. They are noted in the spreadsheet, usually by means of extrapolated data, "c". My K6-3+ at 600 MHz had trouble.

It is only natural that different hardware combinations will cause better/worse performances. I wish that Vogons benchmark archive link was available before starting this project. The more benchmarks suites the better.

[feipoa]

This post is a note to self

Next revision should include:

1) Add K6-III-400 and 450 results for the i430tx board.

2) Enable FastVid or MTRRLFBE for CPUs which support write combining and retest effected programs (benchmarks which run at resolution >320x200 ?)

3) Enable Quake2 3DNow! patch and retest AMD K6-2/3, VIA C3, and Winchip2 CPUs.

4) Corrections: There are two Intel P55C, 262 MHz listings on the Overall chart with different scores. This is correct, however one of them should have an "SS7" suffix to denote it was run on a Super 7 board.

5) Investigations: Why did the C3 Nehemiah at 600 MHz, 66 FSB score 0.5 points higher than the C3 nehemiah at 600 MHz, 100 FSB in the Overall Performance chart?

6) CPUs to add: Intel Pentium II Dechutes - 450, Winchip2 W2B, Rise MP6-333

I will not be doing this anytime soon.

[gerwin]

2) Enable FastVid for CPUs which support write combining and retest effected programs (benchmarks which run at resolution >320x200 ?)

In addition to the LFB range one can enable write combining for the VGA range too, but I found Keen games malfunctioned when I did that, and never bothered with it again. Also since it is kinda pointless IF 320x200 games are already at 60 FPS without write combining. Except for benchmarks of course... Be sure to try MTRRLFBE too, I think it beats FastVid in ease of use.

[5u3]

Many BIOSes enable write combining for the standard VGA range by default (as long as they detect a capable CPU).

[feipoa]

Many BIOSes enable write combining for the standard VGA range by default (as long as they detect a capable CPU).

So it is possible that using FastVid or MTRRLFBE will have no impact on some, or all, of the utilised motherboards and CPUs?

[gerwin]

Good point 5u3. I noticed BIOSes have code in regard to CPU write policy registers, but cannot say much on what to expect from it.

As for the LFB range, I found most mainboards show a significant difference in FPS after setting write combining manually: Example. If the BIOS had already done that for the LFB range, there would be no difference in FPS.

[sunaiac]

Hello feipoa,

I just tested 133MHz iDX4s, and didn't have very good results till now :
- One won't start at 66x2, even at 4 volts.
- One will, and will complete pcpbench and 3dbench, but quake won't start and doom will make an error, leading the demo to happen in a different way than the intended one.

So i have the following questions :
- what is the BIOS revision on your MB8433 ?
- how did you validate system stability ?
- did you have to try several CPUs ?

[feipoa]

I think of the 3 I tested, they all had the same stability issues. I'm not sure if I used 4 V or 5 V. I think for the Quake test, I had to put the chip in the freezer first. There are several Windows-based tests which would not complete.

I never established stability at 133 MHz with an Intel DX4.

I am using the BIOS version found in the World's Fastest 486 thread.

Be sure to set your cache to 3-2-2-2 and your read read/write wait to 2/0 WS.

[sunaiac]

Oh OK, I thought you got a stable one
Well I guess then 120 is the best for those chips
They have a much bigger die than the DX5 after all.

[feipoa]

I have added the Excel sheet to the original post. Column and row headers locked for easy viewing.
download/file.php?id=16770

To avoid having a bunch of outdated Excel documents floating around on the net, I was originally waiting until I made some updates to the file before posting it, however I do not think these updates will come anytime soon.

[vetz]

I have added the Excel sheet to the original post. Column and row headers locked for easy viewing.
download/file.php?id=16770

To avoid having a bunch of outdated Excel documents floating around on the net, I was originally waiting until I made some updates to the file before posting it, however I do not think these updates will come anytime soon.

Why not upload it to Google Docs like Phil has done with his sheets?

[feipoa]

Google Docs is too newschool for my tastes.

[matze79]

Why is a K6 called 686 ? its not fully P6 combatible
Its just a i586 Core.
Like 6x86, K6-2, K6-3 etc.

[386SX]

I was reading the 3DNow! benefits in Quake2 as I remember, it was really incredible. 31% in K62-300... man..

[feipoa]

When using a K6-III-450, I found the 3DNow! patch to be unremarkable. Did I do something wrong? For example, using Quake II (800x600x32) as the benchmark in Win98SE, I obtained

ATI Rage 128 VR 32 MB
38.7 fps before 3DNow! patch
40.0 fps after 3DNow! patch

3DLabs Oxygen VX1 32 MB
29.2 fps before 3DNow! patch
30.4 fps after 3DNow! patch

NVIDIA TNT2 M64 16 MB
28.7 fps before 3DNow! patch
29.3 fps after 3DNow! patch

NVIDIA TNT 16 MB
41.5 fps before 3DNow! patch
42.2 fps after 3DNow! patch

Matrox G450 32 MB
20.1 fps before 3DNow! patch
20.0 fps after 3DNow! patch

GeForce2 MX400 64 MB
48.8 fps before 3DNow! patch
53.7 fps after 3DNow! patch

Voodoo Banshee 16 MB
Would not run with 3DNow! optimisations when using 3dfx-glide mode.

[386SX]

When using a K6-III-450, I found the 3DNow! patch to be unremarkable. Did I do something wrong? For example, using Quake II (800x600x32) as the benchmark in Win98SE, I obtained

ATI Rage 128 VR 32 MB
38.7 fps before 3DNow! patch
40.0 fps after 3DNow! patch

3DLabs Oxygen VX1 32 MB
29.2 fps before 3DNow! patch
30.4 fps after 3DNow! patch

NVIDIA TNT2 M64 16 MB
28.7 fps before 3DNow! patch
29.3 fps after 3DNow! patch

NVIDIA TNT 16 MB
41.5 fps before 3DNow! patch
42.2 fps after 3DNow! patch

Matrox G450 32 MB
20.1 fps before 3DNow! patch
20.0 fps after 3DNow! patch

GeForce2 MX400 64 MB
48.8 fps before 3DNow! patch
53.7 fps after 3DNow! patch

Voodoo Banshee 16 MB
Would not run with 3DNow! optimisations when using 3dfx-glide mode.

That's is strange... maybe it was coded only for the K62 and can't recognize the K63 as 3DNow! enabled cpu?

[falloutboy]

The Quake 2 3DNow! patch was optimised for the K6-2 with a Voodoo2 (SLI). It brings an 3DNow optimised minigl driver for the Voodoo2. The optimised software renderer came with later patches.
It also works very well with the V3, V4, V5 and I guess with the banshee as well. But here you need a different minigl driver. I always use the minigl 1.49.
Rename the minigl driver in to "opengl32.dll" and place it in the game directory. In the game you select "3DNow OpenGL" as the video renderer.
My K6-III+@550 MHz + Voodoo5 pci gave me at 640x480 102 fps in the demo1 bench. Voodoo3 is almost equal at this resolution.
The Voodoo2 SLI should be faster in low resolutions.

Found this pictures in the voodooalert forum, it was the work of the user Turrican.

Epox EP-MVP3G5 2MB L2-Cache
320MB (2x128, 1x64) Siemens PC 100 2-2-2-5 @ 66Mhz/100Mhz
Riva TNT2 32MB & 2x Voodoo 2 12MB SLI @ 108 MHz (3dfx 3.02.02)
Creative Soundblaster 4.1 Digital
Windows 98 SE
Quantum Fireball 6.5 GB UDMA 66

Quake II 3.20 Demo1 /w 3DNow!-Patch for K6-2/K6-III
640x480

Look at the K6 450 vs K6-2 450 result. (yes he overclocked the K6 to 450 MHz Vapochill LS ftw.)

And for comparison.