Nexxen wrote on 2026-10-31, 02:08:

Duron 1000, that ups some but not in an interesting way.
Doom max details runs slower than a P166.
IDK but looks like it needs older boards??? 😀

ISA performance got de-prioritized since introduction of pci

Nexxen wrote on 2026-10-31, 02:08:

I soldered a 40MHz crystal to Mclk and results were the same.
Later I'll do 50MHz but I'm not confident it'll make it go any faster.

finally checked and ISA pin B8 is not routed at all = cant support 0 wait states ;(

Nexxen wrote on 2026-10-31, 02:08:

60ns, 70 or 80 makes no difference

the only difference ram rating makes is if its working ok or artifacting.

Nexxen wrote on 2026-10-31, 02:08:

I use VESA.COM to enable Vesa. Version is 1.2D, I'll look for 1.3 or better to see if there is a difference.
This allows 640*480 to work as it finds Vesa to launch benchmarks that need it.

There's no VBE 1.3. Next is VBE 2.0 and the latest is VBE 3.0. For DOS gaming VBE 2.0 is enough.

You can try some "universal" VBE TSR like SciTech UniVBE ver. 6.7. According to its read-me it supports UM85C408 and can provide VBE 2.0 and VBE 3.0. I don't like these TSR VBE solutions and that's why I want BIOS with VBE.

Nexxen wrote on 2026-10-31, 02:08:

I didn't find any difference in performance except for the cpu, Duron 1000, that ups some but not in an interesting way. Doom ma […]
Show full quote

I didn't find any difference in performance except for the cpu, Duron 1000, that ups some but not in an interesting way.
Doom max details runs slower than a P166.

I tried the P166 on a GA-5AA and results were worse.
I soldered a 40MHz crystal to Mclk and results were the same.

An ISA videocard on "too modern" motherboard/system (including Pentium class) usually results in very sluggish video performance. Don't ask how I got to this experience 😀

Nexxen wrote on 2026-10-31, 02:08:

IDK but looks like it needs older boards??? 😀 60ns, 70 or 80 makes no difference - something is wrong and it could be 9502 related?
Later I'll do 50MHz but I'm not confident it'll make it go any faster.

I had the impression, that the PLL chip on your card is still alive and only the 14.31818 MHz crystal resonator is dead. Did you managed to trace the pinout of UM9505?

Ok:

1. I'll stick to 486 - never thought it could be that bad on modern hw... comes as a surprise.
2. I learnt about VESA
3. I have to do some tracing
4. IDK about the PLL - I have a batch of faulty crystals as none is ouputting the 14mhz... Imma put a 16 and see what happens, if it works at all this PLL
5. I'm taking a day or two off to regroup ideas

analog_programmer wrote on 2026-10-30, 15:48:

Mov AX, 0xDEAD wrote on 2026-10-30, 14:58:

VGA 320x240x256 performance on 386-40Mhz (ISA 10Mhz)

What kind of VGA test utility is this?

GSPEED, vidbench v.0.6, VideoSpeed 0.42 and VIDSPEED 4.0 give same results, my is from GSPEED.

Report shows that increasing DRAM(or videomemory width bus?) to 1024K gives 20-40% boost in CPU->VGA memory access speed. Also some OTI cards (OTI-067) have the same dependence of speed on DRAM quantity.

rasz_pl wrote on 2026-10-31, 02:33:

finally checked and ISA pin B8 is not routed at all = cant support 0 wait states ;(

But it still be one of fastest "pre 0ws" card

Generic cards:
Trident 9000i
1487W 1140R Bytes per millisecond
2976W 2281R 16 bit writes/reads
3238W 2648R 32 bit writes/reads

OAK OTI067
1539W 1192R Bytes per millisecond
3079W 2384R 16 bit writes/reads
3331W 2741R 32 bit writes/reads

PVGA1
1479W 1232R Bytes per millisecond
2959W 2464R 16 bit writes/reads
3120W 2829R 32 bit writes/reads

Our hero UM85C408
2365W 1513R Bytes per millisecond
4731W 3026R 16 bit writes/reads
5317W 3527R 32 bit writes/reads

0WS/FIFO cards:
Trident 8900CL
3198W 1208R Bytes per millisecond
6385W 2416R 16 bit writes/reads
7613W 2690R 32 bit writes/reads

WD 90C30
3198W 1522R Bytes per millisecond
6395W 3045R 16 bit writes/reads
7674W 3568R 32 bit writes/reads

Mov AX, 0xDEAD wrote on 2026-10-31, 14:09:

analog_programmer wrote on 2026-10-30, 15:48:

Mov AX, 0xDEAD wrote on 2026-10-30, 14:58:

VGA 320x240x256 performance on 386-40Mhz (ISA 10Mhz)

What kind of VGA test utility is this?

GSPEED, vidbench v.0.6, VideoSpeed 0.42 and VIDSPEED 4.0 give same results, my is from GSPEED.

Report shows that increasing DRAM(or videomemory width bus?) to 1024K gives 20-40% boost in CPU->VGA memory access speed. Also some OTI cards (OTI-067) have the same dependence of speed on DRAM quantity.

Thanks for this information!

I also have one OAK OTI-067 slow ISA VGA abomination with 512 KB VRAM, but it has no place (trough-holes for DIP sockets) for memory upgrades. But I still think the UMC thing does have such option, even if it is not intended to use 8-bit chips.

Mov AX, 0xDEAD wrote on 2026-10-31, 14:18:

But it still be one of fastest "pre 0ws" card […]
Show full quote

rasz_pl wrote on 2026-10-31, 02:33:

finally checked and ISA pin B8 is not routed at all = cant support 0 wait states ;(

But it still be one of fastest "pre 0ws" card

Generic cards:
Trident 9000i
1487W 1140R Bytes per millisecond
2976W 2281R 16 bit writes/reads
3238W 2648R 32 bit writes/reads

OAK OTI067
1539W 1192R Bytes per millisecond
3079W 2384R 16 bit writes/reads
3331W 2741R 32 bit writes/reads

PVGA1
1479W 1232R Bytes per millisecond
2959W 2464R 16 bit writes/reads
3120W 2829R 32 bit writes/reads

Our hero UM85C408
2365W 1513R Bytes per millisecond
4731W 3026R 16 bit writes/reads
5317W 3527R 32 bit writes/reads

0WS/FIFO cards:
Trident 8900CL
3198W 1208R Bytes per millisecond
6385W 2416R 16 bit writes/reads
7613W 2690R 32 bit writes/reads

WD 90C30
3198W 1522R Bytes per millisecond
6395W 3045R 16 bit writes/reads
7674W 3568R 32 bit writes/reads

You're amazing! I also have Trident 9000i-1 (probably very similar to 9000i) ISA VGA card, so now I can see which of these slow ISA videocards is "the best amongst the worst ones" 😁 I also have Trident 8900C, but it should be inferior compared to 8900CL.

5x86 120MHz
1196 in Landmark. Results were a little under P75.
I'll wait for the 25MHz crystal.

For the time being I'm happy with the results so far.
There are tons of things to know and you all gave me glimpses of it, a collective thank you!!

Nexxen wrote on 2026-11-01, 01:02:

I'll wait for the 25MHz crystal.

28.322!

rasz_pl wrote on 2026-11-01, 03:06:

Nexxen wrote on 2026-11-01, 01:02:

I'll wait for the 25MHz crystal.

28.322!

Yes 😀 My mistake.

analog_programmer wrote on 2026-10-31, 14:22:

I also have one OAK OTI-067 slow ISA VGA abomination with 512 KB VRAM, but it has no place (trough-holes for DIP sockets) for memory upgrades. But I still think the UMC thing does have such option, even if it is not intended to use 8-bit chips.

As far as I know, the OTI-067 is a shrunk down chip that only supports 16-bit RAM at 512KB. To get 32-bit RAM at 1MB, you need the full chip, the OTI-077. Sockets wouldn't help on the 067, it only has 16 data pins.

mkarcher wrote on Today, 12:44:

analog_programmer wrote on 2026-10-31, 14:22:

I also have one OAK OTI-067 slow ISA VGA abomination with 512 KB VRAM, but it has no place (trough-holes for DIP sockets) for memory upgrades. But I still think the UMC thing does have such option, even if it is not intended to use 8-bit chips.

As far as I know, the OTI-067 is a shrunk down chip that only supports 16-bit RAM at 512KB. To get 32-bit RAM at 1MB, you need the full chip, the OTI-077. Sockets wouldn't help on the 067, it only has 16 data pins.

I have 6 sockets, adding 2 to 4 would make it 768KiB.
That is an odd quantity.
24 bit would it be.

Nexxen wrote on Today, 12:59:

I have 6 sockets, adding 2 to 4 would make it 768KiB.
That is an odd quantity.
24 bit would it be.

I've seen graphics cards from that era with 6 sockets, but I didn't yet see any clear sign that these cards actually support 24 bit bus width. On the other hand, UMC and Trident cards are known to be designed for partially faulty RAM chips, using 3 bits from 4 of the RAM chips an 2 bits from the remaining two, resulting in 3*4+2*2 = 16 bits total. Possibly your OAK OTI-067 card is also designed like this, but they had fully working RAM chips when your card was assembled, so they used jumpers / solder bridges to configure it for 4 fully working chips instead of 6 partially broken chips.

mkarcher wrote on Today, 12:44:

As far as I know, the OTI-067 is a shrunk down chip that only supports 16-bit RAM at 512KB. To get 32-bit RAM at 1MB, you need the full chip, the OTI-077. Sockets wouldn't help on the 067, it only has 16 data pins.

Thanks for the info. This explains why OTI-067 card has no spare memory sockets, through-holes or pads left on the PCB.

mkarcher wrote on Today, 13:12:

I've seen graphics cards from that era with 6 sockets, but I didn't yet see any clear sign that these cards actually support 24 bit bus width. On the other hand, UMC and Trident cards are known to be designed for partially faulty RAM chips, using 3 bits from 4 of the RAM chips an 2 bits from the remaining two, resulting in 3*4+2*2 = 16 bits total.

What about two 1Mx8 chips (or four 512x4) chips = 16-bit bus for the UMC card? Probably we may figure out something like VRAM-upgade-adaptor with all those bridges and sockets on the PCB.

analog_programmer wrote on Today, 13:23:

What about two 1Mx8 chips (or four 512x4) chips = 16-bit bus for the UMC card? Probably we may figure out something like VRAM-upgade-adaptor with all those bridges and sockets on the PCB.

This is unlikely to work. Typically, the adressing capability of the entry level SVGA chips of that era end at 256K*something, by using 9 address lines. These lines are used both for the column and row adress, so you get 512 rows * 512 columns = 262144 memory locations. If the UMC chip works like this, 1Mx8 or 512Kx4 can not be addressed by that chip.

The CL-GD5420 is a notable exception, it has a tenth memory address pin and supports (according to the data sheet) two 512K x 8 chips for 1MB at 16 bit bus width. 512K*4 are not mentioned in the Cirrus data sheet, but if those chip do exist, they should work as well. Actually, I don't really get the technical point of the 5420, as it uses the same pinout as the 32-bit capable 5422. It seemst to be nearly(?) the same die, just missing the bond wires for the second 16 data pins. Most likely, the 5420 exists only for marketing reasons, to have a 16-bit chip at a lower price point than the 32-bit chips.