Nice thread!
Would you be willing to try out some games as well? Wonder if the added cache would make any of them run faster in Software mode or even 3D accelerated!

Interesting! I've always wondered about these 3 variants of PPro.

Not sure about the 1MB variant being slower, perhaps it was really meant to do very heavy server types of data handling and were kinda over-optimized specifically for that?

F2bnp wrote:

Nice thread!
Would you be willing to try out some games as well? Wonder if the added cache would make any of them run faster in Software mode or even 3D accelerated!

of course I would, and I in fact wanted to do so, but time was of the essence (the beer store close at ten) and I couldnt think of any demos or games that would give you an fps. But I know some does, do you have any title in mind ? Maybe big red racing ?

Tetrium wrote:

Not sure about the 1MB variant being slower, perhaps it was really meant to do very heavy server types of data handling and were kinda over-optimized specifically for that?

on other tough maybe intel had to add an extra cache controller to handle the second 512kb half and that would slow down the whole process when using only the first few kb of cache. Because, of course, I dont think nssi, nu8 and pctools benchmark are complex enough to take care of l2 cache difference, but then again, who is ?

Interesting!
I would also appreciate some game benches... Quake would be a welcome title.

I did some read in hope to understand why the cache made so little difference. I have not find any satisfactory information but ideas for some more win98se bench software. Also I did reflect on the F2bnp idea and came to the same conclusion as amadeus7779999. I must have spent an hour to try to make fps appear on timedemo, so finally I used another method that give same results in a fraction of the time. Spoiler : the results are even more boring than my first post. It start to make no sence at all... I'm still open to idea, but unless I can boot a nt4/2000 on that machine I see no point in continuing my investigation.

Quake 1.08
msdos6.22 clean boot
software rendering on matrox g450 pci
quake108 -nosound -nocdaudio
~viewsize 100
~timerefresh at quake gates

1 x 256kb:
watts: 83w
320x200 fps: 43.9
360x400 fps: 24.8
640x480 fps: 8.7
1280x1024 fps: 2.28

1 x 512kb:
watts: 90w
320x200 fps: 46.5
360x400 fps: 25.3
640x480 fps: 8.7
1280x1024 fps: 2.28

1 x 1024kb:
watts: 97w
320x200 fps: 47.5
360x400 fps: 27.1
640x480 fps: 8.8
1280x1024 fps: 2.29

PCMark2002
win98se

1 x 256kb:
watts: 90w
CPU score: 548

1 x 512kb:
watts: 97w
CPU score: 557

1 x 1024kb:
watts: 105w
CPU score: 561

WinBench97
win98se

1 x 256kb:
watts: 88w
CPUMark16: 351
CPUMark32: 514

1 x 512kb:
watts: 93.5w
CPUMark16: 361
CPUMark32: 588

1 x 1024kb:
watts: 102w
CPUMark16: 366
CPUMark32: 601

7zip bench
win98se 7-Zip 9.20 (default benchmark)

256kb

1       speed	cpu usage	rating
2comp	162KBs	100%		158MIPS
3deco	1998KBs	100%		180MIPS

512kb

1       speed	cpu usage	rating
2comp	168KBs	100%		164MIPS
3deco	2023KBs	100%		182MIPS

1024kb

1       speed	cpu usage	rating
2comp	170KBs	100%		166MIPS
3deco	2029KBs	100%		183MIPS

It would be neat to see the CPU cost per fps increase. You could also add the cost of the extra power consumption.

feipoa wrote:

It would be neat to see the CPU cost per fps increase. You could also add the cost of the extra power consumption.

I cannot beleive you can be serious, you must be pulling my legs

Try the 333mhz overdrive. ^.^

isnt that somewhat low for quake?

mrau wrote:

isnt that somewhat low for quake?

Regarding Quake 1 in VGA mode; Swaaye's pentium pro 1MB does 42,7 FPS when underclocked to 133MHz (2x66). source. But it is not really fair because i think he enabled write combining for the VGA range, which is not a reliable setting IMO. Contrary to VESA video modes, where it is surely advisable to enable LFB range write combining, with mtrrlfbe or a similar tool.

gerwin wrote:

mrau wrote:

isnt that somewhat low for quake?

Regarding Quake 1 in VGA mode; Swaaye's pentium pro 1MB does 42,7 FPS when underclocked to 133MHz (2x66). source But it is not really fair because i think he enabled write combining for the VGA range, which is not a reliable setting IMO. Contrary to VESA video modes, where it is surely advisable to enable LFB write combining with mtrrlfbe or a similar tool.

im talking about 640 resolution in particular; iirc someone here on a low end p6 got really fluent quake with a matrox vard (might even have been an older one than g450)

Well, then the last sentence of my previous post applies: 640 resolution in Quake uses a VESA mode, which preferably uses a Linear Frame Buffer (LFB). Write combining greatly speeds up the LFB memory access -> more FPS. Does not matter which graphics chipset, it is a CPU feature.

gerwin wrote:

Well, then the last sentence of my previous post applies: 640 resolution in Quake uses a VESA mode, which preferably uses a Linear Frame Buffer (LFB). Write combining greatly speeds up the LFB memory access -> more FPS. Does not matter which graphics chipset, it is a CPU feature.

no no, write combining speeds up by over 100%? also i remember explicitly the statement that the function in question comes from the GPU

100%, seemingly it goes that far: Benchmarks write combining.

gerwin wrote:

100%, seemingly it goes that far: Benchmarks write combining.

noice, however ive seen many people do not reach that much of an increase; is this similar/complementary/whatever to uncacheable speculative write combining?

I don't know. For starters, where are these other people's results?

gerwin wrote:

I don't know. For starters, where are these other people's results?

on this forum, but don't force me to search, i kinda cannot do this effectively; i'm telling this from memory, so don't crucify me either

It is no problem 😀. Just give MTRRLFBE from Rayer a spin when you are chanced. Other then that, I found it is difficult to keep benchmark results comparable. There is often some detail or setting which complicates it.

I really wonder why BIOS developers generally did not include LFB write combining for DOS. Or maybe DOS games should have enabled this themselves. I found out about it way too late... Maybe in the pentium 2 era DOS was already out of favour for gaming.

gerwin wrote:

It is no problem 😀. Just give MTRRLFBE from Rayer a spin when you are chanced. Other then that, I found it is difficult to keep benchmark results comparable. There is often some detail or setting which complicates it.

I really wonder why BIOS developers generally did not include LFB write combining for DOS. Or maybe DOS games should have enabled this themselves. I found out about it way too late...

would this work on an amd fx 6200 cpu? im not sure how id get dos running here