This graph here will make it pretty clear what I was talking about earlier. Take special note of the one of a P4, notice how there are two color tones. So in short an athlon is a little more powerful than a P3, but much more bandwidth from the 266 effective FSB. The P4 has more bandwidth, but is only effective running optimized code. Also note that the chip I'm using is rated for 1.8ghz, and likely capable of more if I bothered to use. I am very interested in the correlation of memory bus width in video cards. Nice find gerwin. After a quick look I see that my 9800xt has 256-bit bus.

P4s can be pretty bad. The Willamette P4 is particularly bad because of its low clock rates and 256KB L2. A 256KB L2 is not enough for P4 to run efficiently. And then they'd stick it on PC133 SDRAM mobos on top of that. If you want to play with a P4, what you want is a Northwood P4 + RDRAM or dual-channel DDR. Prescott + DDR2 is another option, but Northwood was really the "golden era" of P4.

Video card memory bandwidth is very important of course. The more fillrate a GPU has, the more bandwidth it needs to use that fillrate. There are 128-bit cards today with more bandwidth than a older 256-bit card because RAM speeds have increased a lot since then. The new GDDR5 is particularly cool because it's actually quad data rate. There is a new 40nm Radeon GPU (RV740) that will be ~$100 when it comes out and it'll have that GDDR5 on a 128-bit bus and have bandwidth similar to high end 256bit cards from just last year.

@prophase_j,

It would be very interesting if you could re-run just the default test (640x480x8), using an inferior graphics card to the 256-bit one you are using ATM. How about something like a TNT2 or a Rage 128, or anything cheap/basic you can lay your hands on from that era?

swaaye wrote:

There is a new 40nm Radeon GPU (RV740) that will be ~$100 when it comes out and it'll have that GDDR5 on a 128-bit bus and have bandwidth similar to high end 256bit cards from just last year.

Any '98 drivers for it? 😉

[qoute]Any '98 drivers for it? Wink[/qoute]

Fat chance.

[qoute]How about something like a TNT2 or a Rage 128, or anything cheap/basic you can lay your hands on from that era?[/qoute]

I have a Voodoo5 and TnT on hand. I might be inclined to do that within the next few days. I was interested to what Swaaye is getting with his souped up P3/BX combo. It's too bad he can't throw a Voodoo5 in there with that 89mhz AGP setting, that would make for a really nice comparison on the actual work the processor was doing.

@gerwin,

Why not try both of your Geforce MX440s on the same rig? That way, you would be able to quantify the difference in performance between these cards.

Just re-ran pcpbench.exe again, using it's default no command line parameters, and got my best result -

(P3, slot 1, 800mhz)

Trio64v+ PCI 2mb = 51.5!

Previous 2 tests were:

ATI Radeon 9250 AGP 256mb = 44.9
Voodoo3 SG-RAM PCI 16mb = 41.2

Edit: Inside my mobo's BIOS, I selected "optimal default values", rather than the "fail safe values" which I always select. I re-ran the test with the Trio64v+ card, and increased the score to 52.2 (from 51.5).

I took the time to pcpbench various AGP video cards on the same system, the results are both surprising and boring:

System: Soyo 440BX; Pentium-III-S 1000/105MHz; 512MB Ram.
with...
-Dell ATI Radeon 9800, 128MB DDR 256-bit: 50.9 fps
-Abit Siluro Geforce 4 MX440, 64MB 128-bit: 50.5 fps
-ASUS V9180SE Geforce 4 MX440, 64MB 64-bit: 50.4 fps
-Sparkle Geforce 2 MX200, 64MB: 50.8 fps
-ASUS V3800C TNT2 M64, 16MB: 50.8 fps
-STB Velocity Voodoo 3000, 16MB: 50.9 fps

Then some additional tests with the Abit Siluro Geforce 4 MX440:
-VBEHZ loaded: 50.5 fps
-Different BIOS settings: 50.5 fps
-CPU underclocked to 633/66MHz: 32.1 fps
-FASTVID set to enable 'linear frame buffer write combining': 135.0 fps joy!

Then the slower ASUS V9180SE Geforce 4 MX440:
-FASTVID set to enable 'linear frame buffer write combining': 134.0 fps

FASTVID 'linear frame buffer write combining' gives some visual errors in pcpbench. You can sometimes see the starry sky through the ground. Using VBEHZ this gets a lot less. Starting System Shock CD with the same config: no visual errors.

Okay, these are the measurements, I don't know if I can explain them.

EDIT:
-AMD Sempron 3000+/333MHz with Geforce 7600 GS: 81.8 fps
-Same with all FASTVID options set: 222.8 fps

Don't worry about the visual glitches in PCPBench (flickering, banding, etc.). These are a peculiarity of the benchmark program, usually they happen when the frame rate is higher than the screen refresh rate.

Here are my PCPBench results (K6-3+ 500 MHz, Voodoo 5):

1VESA mode       | linear framebuffer | write-combining | FPS
2----------------+--------------------+-----------------+------
3100 (640x480x8) |      disabled      |     disabled    | 44.9
4100 (640x480x8) |      enabled       |     disabled    | 53.4
5100 (640x480x8) |      disabled      |     enabled     | 63.5
6100 (640x480x8) |      enabled       |     enabled     | 80.5

In order to interpret your readings, it helps to know the following things:
Write-combining is only possible on CPUs that provide MTRRs (Pentium Pro/II, AMD K6-2 (CXT core), Cyrix 6x86 and newer).
The BIOS usually enables write combining automatically, but only for the VGA framebuffer below 1 MB.
Since this range is also used for the banked VESA modes, running banked modes actually gives you faster results, unless you set the MTRRs to cover the LFB range as well.
There are plenty of DOS utilities available for that, well known ones are FASTVID for Intel CPUs and MXK6OPT for AMD.
Windows video drivers released after ~1997 set the MTRRs automatically.

The performance gain you get when enabling write combining is not as high as this synthetic benchmark suggests, but it still pays off in high-res DOS games.

5u3 wrote:

Don't worry about the visual glitches in PCPBench (flickering, banding, etc.). These are a peculiarity of the benchmark program, usually they happen when the frame rate is higher than the screen refresh rate.

Yes, I kinda suspected something like that.

Thanks for the useful info, although I am not yet too familiar with the inner workings of these matters.

I see you get 53.4 fps in the situation where I get 50.5 fps, yes?, but your system has only half the CPU speed (in MHz at least)? I wonder what makes the difference here. See how the fps dropped when I underclocked the FSB to 66MHz.

EDIT:
This page I found describes a pentium III system.
He gets 89.6 fps in the same benchmark test. hmm... with or without FastVid? do I compare this to 50.5 or 135.0? nevermind. 😖
AXCEL 216 benchmarks

@prophase_j,

Were you using FASTVID when you ran your tests?

@gerwin,

Please can you tell me a bit more about your AMD Sempron 3000+/333MHz system? I have never bought AMD (but will soon), and just wondered what types of slots this mobo had? Does it have no ISA slots? Does it support windows 98SE? Thanks.

@everyone,

Your posts are a treasure-trove of excellent information.

😀

retro games 100 wrote:

Please can you tell me a bit more about your AMD Sempron 3000+/333MHz system? I have never bought AMD (but will soon), and just wondered what types of slots this mobo had? Does it have no ISA slots? Does it support windows 98SE? Thanks.

I always like bragging about my own stuff, sure. 😀 I will put in a topic at the system specs subforum with some pictures.

Sorry it has taken me a minute to set it up. I was kinda nervous about taking my g-rigged 9800 out, had to find a floppy drive... floppy boot disk. I'll test my Voodoo5, a GF4 mx440, and a TNT. I'll test the default, and 800x600, 1024x768 at 256-colors and 800x600 and 1024x768 at 16bit. Test results are very interesting. Also, when I first tried the bootdisk I got a much slower score with the 9800xt than before when I simply exited from Win98. I suspect this is because of the PCI latency patch and 4-in-1 drivers. After I ran fastvid, I got a stellar result. Well over 200!!! So when I'm testing I'll do the default with no changes and then with fastvid, but the higher res/bit depth ones will be applied. Also regarding the coruption display errors: I'm satisfied with the banding being a side effect of super fast FPS. The junk in the lower part of the screen as being a quirk of this bench.

Voodoo5 AGP VBE 3.0 16mb: (Notes: This thing was a freakin champ! No rendering/display errors side from the hyper drawing effect at mass FPS.)
No changes/default: 68.8 LNB Enabled: 227.6
800x600x 8: 137.5 800x600x16: 81.6
1024x768x 8: 90.7 1024x768x16:52.4
1028x1024x8: 57.7 1028x1024:16: 32.5

Nvidia Vanta AGP 16MB (NV6) VBE 3.0 4mb:
No changes/default: 74.6 LFB Enabled: 238.0 < Yowser!
800x600x 8: 146.0 800x600x16: 88.3
1024x768x 8: 97.3 1024x768x16: 29.2
1028x1024x8: 33.4 1028x1024:16: 18.0

Edit:So I can't find the VGA adapter to DVI for my 440. It is a LP variant so its gonna have to wait.
2nd Edit:Looks like my TNT was actually a Vanta (NV6) it only has a 64-bit bus, basicly a value TNT2.

gerwin wrote:

I see you get 53.4 fps in the situation where I get 50.5 fps, yes?

Correct.

gerwin wrote:

but your system has only half the CPU speed (in MHz at least)? I wonder what makes the difference here. See how the fps dropped when I underclocked the FSB to 66MHz.

CPU speed doesn't matter in this case, because with an uncached framebuffer, the frame rate only depends on the bus and video memory speed, which is about the same both on our machines (~100MHz). That's why the frame rate drops when you lower the FSB to 66 MHz.
With write-combining, the picture changes: My K6-3 only manages 80.5 FPS, while your P3-1000 gets 135.

prophase_j wrote:

Also, when I first tried the bootdisk I got a much slower score with the 9800xt than before when I simply exited from Win98. I suspect this is because of the PCI latency patch and 4-in-1 drivers.

It's also because the windows drivers enable write-combining for the LFB range, but the BIOS only enables it for the standard VGA framebuffer.

Okay, I didn't get a chance to use the MX440 but I think I have established some interesting trends. I don't think that the memory bandwidth of the card has half as much to do with it as I thought, at least once you have a ton of processing power. I was really amazed to see that Vanta. It was faster than the Voodoo until you get up 1280x1024.

retro games 100 wrote:

@KAN,

Isn't that Soyo board one of those "industrial" things? (I think I'd be happier to go for something that the "masses" use.)

Sorry for the very late reply, I've been down with illness. I don't think it's an industrial mobo though; it seems to be an ordinary intel 845-based mobo, but with ISA slots. Also, it has on-board AC97 audio; is it common for industrial mobo to have on-board audio?

Kreshna Aryaguna Nurzaman wrote:

retro games 100 wrote:

@KAN,

Isn't that Soyo board one of those "industrial" things? (I think I'd be happier to go for something that the "masses" use.)

Sorry for the very late reply, I've been down with illness. I don't think it's an industrial mobo though; it seems to be an ordinary intel 845-based mobo, but with ISA slots. Also, it has on-board AC97 audio; is it common for industrial mobo to have on-board audio?

Hope you're feeling better KAN! 😀
Do you know if 845-based mobos have win98 chipset drivers?
Thanks.

retro games 100 wrote:

Hope you're feeling better KAN! 😀

Thanks! 😀

retro games 100 wrote:

Do you know if 845-based mobos have win98 chipset drivers?
Thanks.

I haven't tried my mobos yet; they're still safely stored in my vault. However, I think generic 845 chipset driver should work, shouldn't it?

5u3 wrote:

CPU speed doesn't matter in this case, because with an uncached framebuffer, the frame rate only depends on the bus and video memory speed, which is about the same both on our machines (~100MHz). That's why the frame rate drops when you lower the FSB to 66 MHz.
With write-combining, the picture changes: My K6-3 only manages 80.5 FPS, while your P3-1000 gets 135.

Very clear, thanks!
The video memory speed, I would think so too, but I measured the Radeon 9800 makes no difference to a TNT2 budget version. Any ideas on that matter?

gerwin wrote:

The video memory speed, I would think so too, but I measured the Radeon 9800 makes no difference to a TNT2 budget version. Any ideas on that matter?

Same reason. All those cards you tested without FASTVID get ~50 FPS, because they are bottlenecked by the uncached framebuffer.