... and IIRC the later Catalyst drivers removed those 3DNow optimisations so you actually need to hunt.

MattRocks wrote on 2025-11-06, 16:03:

... and IIRC the later Catalyst drivers removed those 3DNow optimisations so you actually need to hunt.

I have used 3DNow! OpenGL at Quake2 with ATi Rage Pro / XL (improvements visible at lower resolutions with AMD K6-II+ 400Mhz)

Quake2 3.20@Rage XL
Quake2 3Dnow ! OpenGL 320x240 all default 34.7
Quake2 default OpenGL 320x240 all default 33.7
Quake2 3Dnow ! OpenGL 640x480 all default 32.2
Quake2 default OpenGL 640x480 all default 31.4

marxveix wrote on 2025-11-06, 16:32:

I have used 3DNow! OpenGL at Quake2 with ATi Rage Pro / XL (improvements visible at lower resolutions with AMD K6-II+ 400Mhz) […]
Show full quote

I have used 3DNow! OpenGL at Quake2 with ATi Rage Pro / XL (improvements visible at lower resolutions with AMD K6-II+ 400Mhz)

Quake2 3.20@Rage XL
Quake2 3Dnow ! OpenGL 320x240 all default 34.7
Quake2 default OpenGL 320x240 all default 33.7
Quake2 3Dnow ! OpenGL 640x480 all default 32.2
Quake2 default OpenGL 640x480 all default 31.4

as i concluded years ago, 3dnow version of quake2 yields significant improvement on software rendering and 3dfx cards, but very slight boost for nvidia and ati cards.

Damn! I forgot about this thread. I have some updates. A new freshly compiled exe for Quake2 with code taken from the #Quake2ToTheMax contest. The exe has a loot of quality if life improvents and speed improvements. I will come back this evening and post the stuff.

Somebody can share some light why Rage3 gets 11.6fps penalty with cl_lights @ multitexture with AMD K6-II?
It gets 4fps gain with same settings, multitexture wins with set cl_lights "0" and loses with set cl_lights "1" 😀

Direct battle,same settings,difference is multitexture

No multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 28.4
vs
Multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 32.4

4fps gain is good results @ rage3 and multitexture
----------------------------------------

No multitexture - Bilinear enabled / Default
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 27.9
vs
No multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 28.4

Difference is small, 0,5fps @ cl_lights on/off ok!
-----------------------------------------

Multitexture - Bilinear enabled / Default
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 20.8
vs
Multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 32.4

difference is 11.6fps , that seems very odd to me 😀

Re: ATi RagePro drivers and software
Re: ATi RagePro drivers and software

Here is the new 3.27 patch. A faster exe that gives you more frames per second, packed with Quality of life improvements
Note You need to patch Quake2 to version 3.20 before using this.
The SSE version ONLY works on computers with SSE

Quake 2 BDGN 3.27 patch notes.

Modern screen resolutions

Cinematic can now be turned off in the option menu so the intro and
movies between missions don play.
It can also be disable from the console using cl_hack_nocinematics 1.

Hiding or drawing the weapon on the left can now be done from the
option menu, no need to go to multiplayer and player setup menu.

There are now 4 sound quality options instead of two.
8khz was added for slower machines and 44kHz for faster machines.
The game only has 22khz sound originally so a high definition
sound pack needs to be added to benefit from 44kHz.

Wall and floor textures can also be disable in OpenGL mode from the
video menu, mainly for very slow machines.
The console can also be used to disable textures, gl_hack_notextures 1.
There is also gl_hack_noimages 1 to turn off even more textures but
the result is ugly.

Zoom was added to the game.
Its 2.4x and is bound to mouse2 but can be rebound to any key in the
option menu and customize controls.
Sensitivity is adjusted automatically when zoom is used.

If you desire custom zoom and sensitivity you can set zfov and
zsensitivity manually in the console.
You should also set dfov and dsensitivity to match your normal
fov and sensitivity cvar.
*(z for zoom and d for default when not zoomed)

All keybindings where updated to modern WASD hack style and more
keybindings are available in the option menu under customize controls,
like weapons, zoom and grenade.

Freelook is now on by default and a lot of legacy keybindings and mouse
settings where removed.

Chris Dohnal SMP patch for Quake 3 was added to the video menu as a
driver option, SMP OpenGL.
Only OpenGL is available as a menu option because the 3dfx
implementation was intended to use 3dfxvgl.dll from the 30202 voodoo2
drivers and using that or other minigl driver offered no benefit in any
testing we did.

Using the SMP Opengldriver option in the video menu I saw a gain
from around 160fps to 220fps on my Dual Athlon MP 2Ghz with the patch.
On my Dual Pentium 3 1Ghz I saw no gains, so results may very so make
sure to benchmark before and after using timedemo 1 and demo1.dm2
in the console.
Also Make sure you actually have 2 or more processors and are running
windows NT, this is not for single processor systems with windows 9x,
it will crash but the game should launch just fine again.
Processors with SMT like Pentium 4 with Hyper Threading should work as
long as you are on some Version of Windows NT.
The patch should be mostly useful if you are CPU limited and not GPU limited.

Here is the latest Quake2 speed patch that we have done.
It contains the the above posted 3.27 patch with the same faster quake2 exe, dll:s etc
a optimised autoexec.cfg file for maximum FPS,
pak7.pak file contains smaller uglier textures for low vram computers
pak8.pak contains low quality 4khz mono sounds for those really slow computers

Thank you, i will try some of it. Do i need to do something extra or just copy files over Quake2 v3.20 files if same named files exist?

marxveix wrote on 2025-11-11, 14:53:

Thank you, i will try some of it. Do i need to do something extra or just copy files over Quake2 v3.20 files if same named files exist?

All you need to do is to place the files in your Quake 2 folder

Necrodude wrote on 2025-11-11, 14:57:

marxveix wrote on 2025-11-11, 14:53:

Thank you, i will try some of it. Do i need to do something extra or just copy files over Quake2 v3.20 files if same named files exist?

All you need to do is to place the files in your Quake 2 folder

Thanks, later today i will check it out.

It's not always as simple as swapping drivers on Socket 7.

My view is that the bottleneck is the L2 cache that could be 256K, 512K, or 1024K. And, the bottleneck is rarely the cache bandwidth. The approach I take is to estimate OS + OpenGL driver + Q3 app footprint in L2 cache. I want them all to remain resident in the L2 cache because if those three softwares are evicting each other then my K6 would be polling system RAM - and that is a disaster for FPS.

Tuning was quite straightforward in DX5 era when all T&L is handled by the K6. Voodoo3 is the easy win.

DX7 makes it complicated because T&L drivers need a big chunk of L2 cache - and those drivers do nothing useful because Quake III does its T&L on the CPU anyway. DX8 makes it even more complicated.

The fun challenge is, can you get anything newer to outpace a Vooodoo3?

I doubt any of the "cool cards" would score highest FPS because they all promised extra driver features that Quake III cannot use - bigger drivers. I'd bank on RV100 (smaller pre-Catalyst drivers), RV200 (extra hardware features might help), or GF4MX (balanced).

GeForce 3 Detonator drivers fill L2 cache with a lot of T&L fancy stuff that Quake III won't use - it's the wrong card to match with a K6.

The more nuanced problem is a Voodoo 3 is synchronous VPU, meaning is runs commands lock-step with the CPU. If you can speed up the CPU commands/second, you speed up the Voodoo commands/second - Glide is L2 cache friendly. All the rival GPUs are asynchronous, meaning the CPU is pushing commands to memory in batches and GPU is pulling commands from memory in batches - that's a head scratcher if big CPU batches and GPU batches are evicting each other from the L2 cache.

But in all cases, regardless of VPU/GPU, the first thing to do is stop other junk filling your L2 cache: Disable anything you don't use in BIOS. Disables any left over hardware devices. Disable all unused Windows services.

Enabling various motherboard chipset features can slow down a K6 because every extra feature means another driver claiming extra space in the L2 cache. And, whatever config works best for you will be determined by how much cache do you have to play with: 256K, 512K, 1024K?

marxveix wrote on 2025-11-11, 13:36:

Somebody can share some light why Rage3 gets 11.6fps penalty with cl_lights @ multitexture with AMD K6-II? It gets 4fps gain wit […]
Show full quote

Somebody can share some light why Rage3 gets 11.6fps penalty with cl_lights @ multitexture with AMD K6-II?
It gets 4fps gain with same settings, multitexture wins with set cl_lights "0" and loses with set cl_lights "1" 😀

Direct battle,same settings,difference is multitexture

No multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 28.4
vs
Multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 32.4

4fps gain is good results @ rage3 and multitexture
----------------------------------------

No multitexture - Bilinear enabled / Default
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 27.9
vs
No multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 28.4

Difference is small, 0,5fps @ cl_lights on/off ok!
-----------------------------------------

Multitexture - Bilinear enabled / Default
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 20.8
vs
Multitexture - Bilinear enabled / set cl_lights "0"
Quake2 v3.20 @Rage XL FastPerf-OpenGL setting
Quake2 3Dnow ! OpenGL 640x480 all default 32.4

difference is 11.6fps , that seems very odd to me 😀

Re: ATi RagePro drivers and software
Re: ATi RagePro drivers and software

You have found a concrete example of the challenge I explained in my last post.

On a K6, each VPU special effect requires a CPU driver call + VPU driver call. Those calls are thrashing the L2 cache. When you make too many driver calls, the commands spill over from L2 cache into much slower system RAM - and that is causing your observed FPS drop!

I remember all this stuff because, back in the day, I was probably the last person still stuck on Socket 7. I remember dragging my cranky K6-2 with 3Dfx Banshee into rooms filled with Athlons and GeForce256s. I had to squeeze every last bit out the L2 cache, and each OS or driver update only made that demand harder.

MattRocks wrote on 2025-11-12, 14:46:

You have found a concrete example of the challenge I explained in my last post.

On a K6, each VPU special effect requires a CPU driver call + VPU driver call. Those calls are thrashing the L2 cache. When you make too many driver calls, the commands spill over from L2 cache into much slower system RAM - and that is causing your observed FPS drop!

I remember all this stuff because, back in the day, I was probably the last person still stuck on Socket 7. I remember dragging my cranky K6-2 with 3Dfx Banshee into rooms filled with Athlons and GeForce256s. I had to squeeze every last bit out the L2 cache, and each OS or driver update only made that demand harder.

I look into it and i dont remember Pentium 3 doing that kind of behaviour with same graphics hardware and same vga drivers.

marxveix wrote on 2025-11-12, 16:18:

I look into it and i dont remember Pentium 3 doing that kind of behaviour with same graphics hardware and same vga drivers.

The Socket 7 boards have L2 cache on the motherboard between the CPU and the RAM, clocked to the FSB - that architectural design dates back to the 486. Super Socket 7 K6-III was the very last CPU to be paired with that architecture. When the L2 spills, the CPU waits for the memory controller to synchronously write and read from main memory.

Pentium 3 has a totally different architecture and totally different character. It still has a L2 cache but it's operating very differently because not just because its clocked to the CPU (like K6-III L3 cache) but also because the Pentium III memory controller on the motherboard has higher bandwidth, asynchronous prefetching, etc. - it is behaving completely differently.

Those differences were not observed straight away (early K6-2 vs Pentium II reviews) because the OS and drivers and apps that they were using had all been designed to work within Socket 7 Level2 caching constraints - although Quake III was developed on later architecture I assume they backtested it to keep Socket 7 friendly.

marxveix wrote on 2025-11-11, 15:00:

Necrodude wrote on 2025-11-11, 14:57:

marxveix wrote on 2025-11-11, 14:53:

Thank you, i will try some of it. Do i need to do something extra or just copy files over Quake2 v3.20 files if same named files exist?

All you need to do is to place the files in your Quake 2 folder

Thanks, later today i will check it out.

Forgot to mention that the config has a few lines that you need to know about.
\ not supported in vanilla q2
set gl_hack_noimages "1"
set gl_hack_notextures "1"
set cl_hack_nocinematics "1"
set s_khz "8"
noimages and notextures turns of the textures. If you set them to 0 you will see the games textures as normal
set khz "8" means low quality sound. other values are 11 22 and 44. 22 is default.

I have so far got the best results on my Pentium MMX 233@266 with a PCI riva 128. The FPS went up from 11 fps with vanilla Quake 2 to 33 fps with this patch.

I got good results on my k6-3+ machine. But the texture pak file and the sound pak file only gave me negligable results.

However sad as it is the 3dnow patch for quake2 is still faster then the 3.27 patch. But it is fully possible to combine the autoexec.cfg file and the pak files with the 3dnow patch. Another option is to use the 3.27 patch and use the ref_glam.dll as renderer. It will give a boost to the 3.27 patch, but some features wont work as for example turning of textures.

Necrodude wrote on 2020-02-19, 22:08:

Please test with a similar cpu. The components in my system is as follows. Motherboard: Asus P5A Cpu: Amd k6-3+ @ 617 112x5, […]
Show full quote

Please test with a similar cpu.
The components in my system is as follows.
Motherboard: Asus P5A
Cpu: Amd k6-3+ @ 617 112x5,5
Graphics Card: Geforce 256 DDR
128mb Cl2 PC133
AWE 64
Intel® PRO/1000 GT network card

Do you have a Audigy 2 ZS?

Appreciating this was not put to me (and I don't have my Socket 7 systems to hand) but my memory is clear as day.. your ask reflects the tragic irony of 1990s computing :'(

Before I elaborate, the quick trick to getting lower FPS is to have less sound effects. And you want sound, so it seems to me the sound should be nominal - less is more.

Quake3 uses software synthesis and pushes WAVs via DirectSound so from a computational perspective the K6 does the heavy lifting.

Your AWE64 is already receiving and playing Quake3's WAVs with small driver footprint and low bus activity. What the Live/Audigy cards would do is add EAX colouring to WAVs, but that's adding work that worsens your FPS. And, this is where the tragedy begins..

What you really want, and what everyone wanted, was silicon to accelerate the sound maths. Quake3 and others did support A3D for that. But, not all A3D are equal. You can get software emulated A3D, and hardware accelerated A3D - you want the latter.

You want an Aureal. Quite rare. It's a lot of hunting for a tiny FPS gain, but it existed and it no longer exists.

It's a lot like the time Netscape threatened to make software that interfered with Microsoft's platform domination. Microsoft's answer? "Sue the intruder! Sure, the better product will win the legal case, but strategic delays and mounting legal costs will kill it off before it does!"

Creative Labs had the same idea. Aureal went bust while winning their right to exist. Now we all have worse sound cards.

here we go again.... see page 4.

Page 4 doesn't show you how to get >5 FPS boost 😉

The thing is you want sound, and you don't want to waste CPU cycles on it. That requires an Ensoniq PCI card (or rebadged Ensoniq like CT4750) - not something anyone admitted at the time because they were cheap and ugly and didn't add anything, but that's the whole point if you are chasing FPS.

MattRocks wrote on 2025-11-13, 01:59:

The thing is you want sound, and you don't want to waste CPU cycles on it. That requires an Ensoniq PCI card (or rebadged Ensoniq like CT4750) - not something anyone admitted at the time because they were cheap and ugly and didn't add anything, but that's the whole point if you are chasing FPS.

I have one Audio PCI 3000 with Creative ES1370 with AK4531. You think this works faster than ESS Solo1?
I tested K6II with Creative Live card,but it does not like my MVP3 or MVP3 motherboard does not like Live.