mbarszcz wrote on 2023-03-19, 21:07:

The cards are paired with a DX4-100 and 32MB of EDO RAM.

Wasn't there some limitation with the Cirrus Logic VESA frame buffer?
https://nerdlypleasures.blogspot.com/2014/01/ … nk-systems.html
"UniVBE 6.53 reports that Linear Frame Buffer will not work reliably with CL 542x chips with more than 14MB of RAM"
But in a VLB system, the VESA speed in DOS may not matter much, since the games there are mostly 320x200 resolution.

That's not that much slower. But you're comparing 5429 with the best version of ET4000/w32p. Might not be the best version of the card but non the less it's one of the fastest chipsets in DOS.

Adding memory won't do anything performance wise. Even Tseng's 2MB performance benefit only shows in Windows GUI.

542x chips are MUCH slower than Tseng. The only limitation is 486.

Yeah a 2MB ET4000 W32p was hot stuff back then. It's faster in DOS than almost everything else VLB. It's also a fine Windows card too of course.

Interesting TV tuner version.

The complete CL-GD542x series only has a 16-bit bus interface. It uses the VLB signal to tell the processor: "I only handled 16 bits. If this was a 32-bit request, please repeat the request for the high 16 bits". That's the BS16 (Bus size: 16) signal. This alone causes some slowdown. Furthermore, the Cirrus Logic RAM interface is not the fastest stuff man ever built (whereas the Tseng stuff is hot, Tseng's marketing claimed that their DRAM-based cards rival with the performance of competitors' VRAM-based cards).

Looking at my thread Tool to adjust the memory clock on Cirrus Logic CL-GD542x graphics cards, it seems 6.6MB/s for 16-bit or 32-bit access is in the low side. You should expect 6.6MB/s at byte access, or 13.3MB/s at word access assuming the memory clock is fast enough at FSB33. Those values are not stellar, but acceptable. If you have a 5429, check whether the BIOS initialized the memory clock to 50MHz. The 5429 permits up to 60 MHz officially, and higher frequencies (i.e. overclocking) likely work, too. You still won't reach ET4000/W32p levels, though.

I'm also running GL 5429 albeit in a AMD 120mhz, your numbers seem a bit low. Have you tweaked the bios settings? Results can be very sensitive.

For comparison:
Mem bandwidth - 10,200 kb/s
Chris normal - 47.4 fps
Chris 640x480 - 14.8 fps

mkarcher, I gave your clmclk a try and it is reporting the current MCLK at 50.12 MHz. Upping it to 60MHz gets me a bandwidth of 8437KB/s (up from 7672 KB/s).

What should I look for in the BIOS? My motherboard is a PC Chips M912 v1.7. It used to have fake cache, but now it has 512K of real cache.

My CPU is a DX4-100, so a 33.3MHz Bus x 3.

1cache read hit wait state =  2-1-1-1
2write hit wait state = 0 WS
3dram wait state = 0 WS
4DRAM Page Mode = Fast
5AT Clock = PCLK/4
6Local Ready Delay Setting = Delay 1T
7CPU ADS# Delay 1T or Not = No Delay
8Alt Bit in Tag SRAM = 7+1 Bits
9Signal LDEV# Sample Time = in T3
10ISA Bus Refresh Disable Bit = Fast
11DRAM Refresh Method = RAS Only
12Divider for Refresh = 1/1
13Video ROM Cache C000, 32K = Enabled
14System ROM Cache F000,64K = Disabled
15LOWA20# Select = Chipset
16RC Reset Select = Chipset
17IO Recovery Time Select = 5BCLK

It certainly seems that compared to yours, my 5429 is leaving some performance on the table.

You have a jumper block absent, I think the wait state is strapped in.

The issue with Cirrus Logic VLB cards is that internally use 16-bit design (for example a 32-bit write is converted to two 16-bit writes internally), so they only benefit from faster VLB bus speed. That's why those are usually slower compared to other VLB cards.

gerwin wrote on 2023-03-19, 22:11:

Wasn't there some limitation with the Cirrus Logic VESA frame buffer? https://nerdlypleasures.blogspot.com/2014/01/ … nk-systems […]
Show full quote

mbarszcz wrote on 2023-03-19, 21:07:

The cards are paired with a DX4-100 and 32MB of EDO RAM.

Wasn't there some limitation with the Cirrus Logic VESA frame buffer?
https://nerdlypleasures.blogspot.com/2014/01/ … nk-systems.html
"UniVBE 6.53 reports that Linear Frame Buffer will not work reliably with CL 542x chips with more than 14MB of RAM"
But in a VLB system, the VESA speed in DOS may not matter much, since the games there are mostly 320x200 resolution.

This issue is directly related to what I said before. Having a 16-bit design (based on the ISA bus), those cards cannot address VRAM above 24-bit addresses, which is in fact the 16 Mb limit. If you want to use LFB with those cards, stick with 12Mb of RAM and it will work properlly. Anyway having LFB doesn't help much, because internally they are still 16-bit.

I have the same board too and I'm assuming these are correct settings.

Anyone with further advice, please chime in! I had to manually re-set these after the battery died and only ran some quick testing.

JP16 closed (> 33 mhz)
JP17 open (Vesa 0 WS)

Write Through Cache Enabled
2-1-1-1
Cache Write 0 WS
DRAM Wait 1 WS
DRAM Page Fast
AT Clock PCLK/4
No Delay
No Delay
LDEV In T3
ISA Refresh Slow
DRAM Refresh RAS Only
Divider 1/1
Video ROM Cache Enabled

(Omitted settings match what you have above)

As already said the Cirrus Logic is not disappointing or that slow, the Tseng is very fast.

mbarszcz wrote on 2023-03-19, 21:07:

Would populating the empty 1MB help the performance at all? The general consensus seems to be "no", because dos is never going to use the extra 1MB, but would it provide more memory bandwidth?

Generally speaking yes, and especially for 1->2MB upgrades this is almost always the case. But there are some cases where this isn't true: VGA VRAM size and performance discussion

drosse1meyer, what do your jumpers look like on your card? On mine there are 2 jumpers (4 pins) and they are soldered in (as BitWrangler mentioned).

mbarszcz wrote on 2023-03-21, 14:28:

drosse1meyer, what do your jumpers look like on your card? On mine there are 2 jumpers (4 pins) and they are soldered in (as BitWrangler mentioned).

Nothing of interest. A jumper to set IRQ. And some headers for 'green' control which I assume is some sort of early low power mode thing. No idea how its supposed to work.

Did you try tweaking BIOS and/or jumper settings?

mbarszcz wrote on 2023-03-21, 04:25:

mkarcher, I gave your clmclk a try and it is reporting the current MCLK at 50.12 MHz. Upping it to 60MHz gets me a bandwidth of […]
Show full quote

mkarcher, I gave your clmclk a try and it is reporting the current MCLK at 50.12 MHz. Upping it to 60MHz gets me a bandwidth of 8437KB/s (up from 7672 KB/s).

What should I look for in the BIOS? My motherboard is a PC Chips M912 v1.7. It used to have fake cache, but now it has 512K of real cache.

My CPU is a DX4-100, so a 33.3MHz Bus x 3.

Copy code to clipboard

1Local Ready Delay Setting = Delay 1T
2Signal LDEV# Sample Time = in T3

It certainly seems that compared to yours, my 5429 is leaving some performance on the table.

Keep in mind that my diagram has been obtained at 40MHz VL clock, while you are operating at 33MHz VL clock. That's why there is a hard limit at 13MB/s you can't exceed in any mode at any memory clock on your system: The Cirrus Logic card will not perform a single 16-bit cycle in less than 5 VL clocks, so theoretical max bandwidth is 2bytes/5 * 33MHz.

Furthermore, the two BIOS settings I quoted are the "safe but slow" settings. Try setting them to "No Delay" and "in T2". It's quite likely this provide more benefit to the ET4000 than to the Cirrus card, though. The theoretical absolute maximum of the ET4000/W32p, according to my experiments is 3 VL clocks for 32 bits, so 44MB/s in your system. I don't know whether a DX4/100 is able to provide data that fast, though.

Thanks everyone for all the interesting information in this thread. Even if I can't get it figured out, I learned a few thigs about how it all works.

I'll have to take another look at things tomorrow, but I can confirm that my motherboard is set to no VL wait states since I installed the DX4-100 (with its 33MHz Bus). There isn't any way to change wait states on the card is there (jumpers? video bios?)

I had previously tried Local Ready Delay Setting = No Delay with no improvement to performance. I have not tried Signal LDEV# Sample Time = in T2 yet. Other than that though, I think my BIOS settings are about as fast as they're going to get. Perhaps the difference does come down to the 33MHz Bus rather than the 40MHz, but I have to wonder if the card is somehow inserting wait states. I may try putting the DX2-80 back in to see how that changes things if I feel like monkeying around with it.

swaaye wrote on 2023-03-19, 23:28:

Yeah a 2MB ET4000 W32p was hot stuff back then. It's faster in DOS than almost everything else VLB. It's also a fine Windows card too of course.

Interesting TV tuner version.

Does video capture as well.

Not without drivers it doesn't

mbarszcz wrote on 2023-03-22, 12:22:

Not without drivers it doesn't

I have 2 of those cards with original disks around here somewhere.

Unknown_K wrote on 2023-03-22, 14:38:

mbarszcz wrote on 2023-03-22, 12:22:

Not without drivers it doesn't

I have 2 of those cards with original disks around here somewhere.

😮 I'd love if you could upload them. Information and drivers on the mirovideo is completely unobtanium.