MTRRLFBE used to work on the Athlon XP. Though its benefit is limited: 640X480 should already run above 60FPS regardless...

I'll give it a try but I didnt care very much about if the result are good fps or not, all I wanted to see is the cache difference of different ppro. I'd give RayeR MTRRLFBE a spin and see if it change anything on the 512kb cpu FPS that's on the computer. If it does I might as well redo the bench to see if it make a difference. I hope I will be able to do that tonight. The GXPRO 200 bios is one of the worst I ever saw.

So these 512K and 1024K PPros were built with some kind of server in mind. Databases?

I wonder if a benefit might show in a turn-based strategy game. One that gets nice long turn processing times as the game progresses.

@gerwin : I think you are onto something with that write combining. With the matrox G450 it told me that LFB wasnt supported or wrong adress, with VGA WC it worked but changed nothing in Quake. So I did try an old mga millenium that accepted the command LFB WC and now at 640x480 I can do 22fps and at 1280x1024 the card achieve 7.3fps! You where right about the vesa too, at 360x400 it made no difference.

I couldnt resist to change the cpu for the 1024 big brother to see and try a test at 640x480 and 1280x1024, here is the result respectively: 23fps and 7.4

@swaaye : At the point where I am I dont think that would be worth the trouble. My test was to see the cache performance on 16bit os. I wasnt expecting such a draw. I'm thinking now to install an hdd on that computer with NT4 or linux and make sure there is a case scenario where the cache make actual difference! Maybe I could try something like freeciv and do some benchmark I think this game is on both os. Do you have any better idea?

ElBrunzy wrote:

Maybe I could try something like freeciv and do some benchmark I think this game is on both os. Do you have any better idea?

Freeciv sounds good to me. I'm not sure how to approach testing it though. These strategy games typically don't have benchmarks built in. Loading a save game and literally using a stopwatch might be the approach.

a bit frustrated that the matrox mga millenium was faster than the matrox g450 under dos I decided to investigate a bit with the univbe/fastvid tools. Fastvid, unlike mtrrlfbe, was able to detect the LFB adress. Then on quake the result of the g450 where the exact same of those with the mga.

Fastvid talk about pentium pro motherboard based on the 82450 and 82440 chipset. I'm not sure if write posting work as I have a 82440 chipset, but fastvid enable it and dont crash. That's for the pointer, the speed increase is simply amazing and I've yet to see any drawback.

Sorry to hear MTRRLFBE could not find the Matrox G450 LFB. But glad you got it sorted it by using Fastvid instead, with the expected increase in framerates.

Your initial benchmarks are fine anyways. it is just good to be aware of the available options, like write combining, in that regard. So once in a while I feel like mentioning that topic.

gerwin wrote:

Your initial benchmarks are fine anyways. it is just good to be aware of the available options, like write combining, in that regard. So once in a while I feel like mentioning that topic.

That is well appreciated in this case. It all make me think of disk cache you see on server hard disk controller. Write-Thru, Write-back, all that, is it the same about the vga card memory and the motherboard memory? Kind of use the vga card ram as a cache ? I wonder why fastvid is so concerned about the fact that I have a pentium pro motherboard, does it say that on every computers ?

I remember the setup process of FastVid to be somewhat weird, compared to MTRRLFBE. Other then that, I don't know.
For general info on the technology involved you can read this:
Wikipedia: Memory type range register
Pentium Pro is mentioned there.