VOGONS


First post, by konc

User metadata
Rank l33t
Rank
l33t

Hi guys
I recently got a GD5434 PCI to tinker with, out of curiosity.

GD5434PCI_c.jpg
Filename
GD5434PCI_c.jpg
File size
1.36 MiB
Views
347 views
File license
Fair use/fair dealing exception

From the data book and wikipedia

wiki.JPG
Filename
wiki.JPG
File size
13.36 KiB
Views
347 views
File license
Fair use/fair dealing exception

I thought that expanding the memory to 2MB would enable the 64bit memory bus and I wanted to measure just how fast it can be and how it compares to other cards of the time.

I'm confident that the memory upgrade is recognized and used. How's that? Initially I had no FPM DRAM and tried some EDO just in case. Although 2MB were reported, testing a mode that requires more than 1MB with the trustful X-VESA produced a ton of errors. After acquiring the correct FPM chips, the same test finishes successfully.

But to my surprise the common DOS benchmarks (PC player, 3D bench, Doom, Quake) on a PIII 800MHz (trying to eliminate/minimize CPU restrictions to show pure VGA performance) showed exactly zero improvement. What's happening?
-Am I wrong to expect a performance increase with 2MB for this card?
-Is my testing system idea wrong for this kind of test?
-Could it be that the 64bit memory bus is only enabled under Windows with drivers loaded for GUI acceleration?
-Is it possible that the budget manufacturer "FastWare" of this card somehow crippled the feature to save 3 cents?

All ideas/comments are welcome.

Reply 1 of 6, by darry

User metadata
Rank l33t++
Rank
l33t++

-Could it be that the 64bit memory bus is only enabled under Windows with drivers loaded for GUI acceleration?

That was my suspicion. Though it migh also help in VESA 2.0 modes using an LFB. Do try to enable writing combining (with fastvid, for example).

If running Windows 3.x or even 9x , maybe you could try Steve Gibson's Torque utility/benchmark .

Scitech Display Doctor/UNIVESA/UNIVBE has a utility that can run dome benchmarks (vbetest, ,AFAICR).

Reply 2 of 6, by Azarien

User metadata
Rank Oldbie
Rank
Oldbie

-Am I wrong to expect a performance increase with 2MB for this card?

I'm not sure you'll get performance increase, but the list of available resolutions will be larger.
With only 1 MB you can get 640x480 in 24-bit color, but 800x600 only in 16-bit color, and 1024x768 only in 256 colors.
2 MB gives you 800x600 in 24 bpp and 1024x768 in 16 bpp.

Reply 3 of 6, by konc

User metadata
Rank l33t
Rank
l33t
darry wrote on 2024-05-14, 08:14:
That was my suspicion. Though it migh also help in VESA 2.0 modes using an LFB. Do try to enable writing combining (with fastvi […]
Show full quote

-Could it be that the 64bit memory bus is only enabled under Windows with drivers loaded for GUI acceleration?

That was my suspicion. Though it migh also help in VESA 2.0 modes using an LFB. Do try to enable writing combining (with fastvid, for example).

If running Windows 3.x or even 9x , maybe you could try Steve Gibson's Torque utility/benchmark .

Scitech Display Doctor/UNIVESA/UNIVBE has a utility that can run dome benchmarks (vbetest, ,AFAICR).

Yeah that's my conclusion too, unless of course the chipset feature is broken on this specific card. I was never able to get a performance increase no matter what I tried in dos (eg. fastvid), windows XP, or dos software under XP. Admittedly I can't speak about windows 3.x or 9x with their own drivers, I was only trying to see if it's an undercover/little known dos performer. Well I can't complain, I have my results: It's not 😁

Reply 4 of 6, by darry

User metadata
Rank l33t++
Rank
l33t++
konc wrote on 2024-05-16, 10:07:
darry wrote on 2024-05-14, 08:14:
That was my suspicion. Though it migh also help in VESA 2.0 modes using an LFB. Do try to enable writing combining (with fastvi […]
Show full quote

-Could it be that the 64bit memory bus is only enabled under Windows with drivers loaded for GUI acceleration?

That was my suspicion. Though it migh also help in VESA 2.0 modes using an LFB. Do try to enable writing combining (with fastvid, for example).

If running Windows 3.x or even 9x , maybe you could try Steve Gibson's Torque utility/benchmark .

Scitech Display Doctor/UNIVESA/UNIVBE has a utility that can run dome benchmarks (vbetest, ,AFAICR).

Yeah that's my conclusion too, unless of course the chipset feature is broken on this specific card. I was never able to get a performance increase no matter what I tried in dos (eg. fastvid), windows XP, or dos software under XP. Admittedly I can't speak about windows 3.x or 9x with their own drivers, I was only trying to see if it's an undercover/little known dos performer. Well I can't complain, I have my results: It's not 😁

There are some PCI registers that you could check to see if 64-bit memory access is enabled and possibly how to enable it (by writing proper values into these registers).
Maybe the VBIOS is expected to set them and somehow isn't doing it ? Or maybe they are implicitly set if 2MB of memory is detected and there is just not a significant performance difference .

reg1.png
Filename
reg1.png
File size
248.84 KiB
Views
187 views
File license
Fair use/fair dealing exception
reg2.png
Filename
reg2.png
File size
183.53 KiB
Views
187 views
File license
Fair use/fair dealing exception

Source : https://dn790008.ca.archive.org/0/items/bitsa … l_4ed_Feb95.pdf

Reply 5 of 6, by konc

User metadata
Rank l33t
Rank
l33t
darry wrote on 2024-05-17, 03:00:

There are some PCI registers that you could check to see if 64-bit memory access is enabled and possibly how to enable it (by writing proper values into these registers).

I found the idea fascinating and decided to explore it, hence the late reply. So after dusting off my C skills,

1MB default configuration soldered on the card: As expected from the data sheet (bit 3 disabled, bit 4 enabled)

1MB.png
Filename
1MB.png
File size
817 Bytes
Views
109 views
File license
Public domain

2MB: the register's bit is set internally and automatically when the memory is expanded (bits 3 & 4 enabled). So the 64-bit memory data bus does get enabled!

2MB.png
Filename
2MB.png
File size
800 Bytes
Views
109 views
File license
Public domain

Since we are it, what about bank switching (bit 7)? It shouldn't matter but let's enable it too...

2MB_2.png
Filename
2MB_2.png
File size
1.27 KiB
Views
109 views
File license
Public domain

Yeah still no difference. So, it appears that the 64-bit memory data bus is enabled with 2MB of VRAM, but it makes exactly zero performance difference under DOS. Not even a tiny one within the small margin of repeated tests, the numbers of 4 different benchmarks are exactly the same to the decimal digit. Now that it's confirmed that the bus width changes correctly when expanding the memory, maybe it makes sense to test with 1. a much less powerful system and 2. in windows 3.x/9x with drivers installed.

Anyway, I had a great time tinkering with this, which was the reason I got the card in the first place. And I'm left with my own DOS software to manipulate registers 😁

Reply 6 of 6, by darry

User metadata
Rank l33t++
Rank
l33t++
konc wrote on 2024-05-21, 08:47:
I found the idea fascinating and decided to explore it, hence the late reply. So after dusting off my C skills, […]
Show full quote
darry wrote on 2024-05-17, 03:00:

There are some PCI registers that you could check to see if 64-bit memory access is enabled and possibly how to enable it (by writing proper values into these registers).

I found the idea fascinating and decided to explore it, hence the late reply. So after dusting off my C skills,

1MB default configuration soldered on the card: As expected from the data sheet (bit 3 disabled, bit 4 enabled)
1MB.png

2MB: the register's bit is set internally and automatically when the memory is expanded (bits 3 & 4 enabled). So the 64-bit memory data bus does get enabled!
2MB.png

Since we are it, what about bank switching (bit 7)? It shouldn't matter but let's enable it too...
2MB_2.png

Yeah still no difference. So, it appears that the 64-bit memory data bus is enabled with 2MB of VRAM, but it makes exactly zero performance difference under DOS. Not even a tiny one within the small margin of repeated tests, the numbers of 4 different benchmarks are exactly the same to the decimal digit. Now that it's confirmed that the bus width changes correctly when expanding the memory, maybe it makes sense to test with 1. a much less powerful system and 2. in windows 3.x/9x with drivers installed.

Anyway, I had a great time tinkering with this, which was the reason I got the card in the first place. And I'm left with my own DOS software to manipulate registers 😁

Thank you for testing and for sharing the results. I tend to use RWEverything or just the plain DOS Debug utility to manipulate registers (not that I often need to), but that's just because I'm not a programmer. Custom DIY stuff is always cool.