LFB is indeed active. Linear and page mode drivers are two different drivers. Furthermore, if you set the buffer start to 32 MB, the image on the monitor is expectedly corrupted. I don't have 1 MB memory modules, only 4 MB. Therefore, I can't use less than 16 MB of memory.

Darmok wrote on 2025-12-25, 18:23:

The motherboard is based on the opti495slc chipset. This chipset supports 64 MB of memory. The BIOS sets the actual memory size to 32 MB. Accessing addresses above 32 MB triggers data transfers with devices on the ISA bus. Since the addressable size of the ISA bus is 16 MB, it appears twice in addresses above 32 MB. The megabyte immediately after the thirty-second megabyte contains 384 KB of ROM and the video page buffer. The next megabyte is free. Megabytes start counting from 0, so I set the LFB to start at 33 megabytes. The BIOS doesn't have an option to specify a memory hole at address 15 MB.

Cool, is that motherboard information available on TheRetroWeb ?

I don't know the motherboard manufacturer. I have a manual for it, which lists it as model MA013.

Looks a lot like this: https://theretroweb.com/motherboards/s/edom-wintech-ma013

Sorry, I just now remembered that I set the LFB address to 34 MB. This is because the OTI-087 supports 2 MB of video memory. Therefore, the Windows driver has a strange limitation on the LFB address: it must start with an even megabyte number.

maxtherabbit wrote on 2025-12-24, 23:27:

How many ISA VGA really did anything useful with a LFB? Seems like few to none

Mach64 - Re: VLB with 386DX-40 ?

Many late ISA cards are based on chips designed for various buses, so they of course do support LFB.
But software support for this use case seems rare, indeed.

I may misremember it, but I think that UniVBE can enable VBE and LFBs on ISA/VLB/PCI VGA cards.
Provided that they have 512 KB or more Video RAM.
For some reason, UniVBE is a bit discriminatory and refuses to support any 256 KB SVGAs,
despite them being able to do certain VBE modes with their own VBE utilities.

On my 386sx with 16mb ram LFB can’t be enabled on a 5434 with UniVbE.

Would be interesting to see benches of ISA LFB on vs off. I normally hear that on ISA no real performance difference.

Marco wrote on 2025-12-26, 15:08:

On my 386sx with 16mb ram LFB can’t be enabled on a 5434 with UniVbE.

Would be interesting to see benches of ISA LFB on vs off. I normally hear that on ISA no real performance difference.

1) Try to set a memory hole in BIOS.
2) Try it with 8 MB RAM.

Will try. Bios hole can’t be selected in my bios. Therefor B)

There were low power 386sx class CPUs that had a 25 bit address 386s and 386sl come to mind

Marco wrote on 2025-12-26, 15:08:

On my 386sx with 16mb ram LFB can’t be enabled on a 5434 with UniVbE.

Would be interesting to see benches of ISA LFB on vs off. I normally hear that on ISA no real performance difference.

It makes a big difference in Duke3D. Double FPS, where the CPU is not limited. Tested on a Chips 65545.

The application most likely needs to work with VESA drivers, or specically written for LFB, such as Win3.1 drivers.

rmay635703 wrote on 2025-12-27, 06:36:

There were low power 386sx class CPUs that had a 25 bit address 386s and 386sl come to mind

It was the embedded version 386EX. It has 26 bit address space for 64 MB RAM.

MikeSG wrote on 2025-12-25, 15:14:

The Vesa drivers for the Chips 65545 ISA card contains an LFB mode, which uses the 15-16MB memory hole.

Duke3D is twice as fast at 320x200, and enables 320x400 for double the resolution at the same frame rate as 320x200 used to operate. Highly recommend. I tested this on a 386DX but a 16 bit bus is a 16 bit bus.

Duke3D requires a 486 processor. It's very strange that you doubled your FPS in 320x200 and 256 colors using LFB. In this mode, the required video memory is 64 KB, exactly the size of one video buffer page. There should be no difference between linear and paged modes. In higher resolution modes, the ISA bus bandwidth begins to be insufficient. The theoretical ISA bus bandwidth at 8 MHz is 8,000,000 x 2/4 = 4,000,000 bytes per second. For a resolution of 640x480, 307,000 bytes per frame are required. The maximum theoretical FPS will be 13. For 320x200 mode, the maximum FPS is 62. Linear addressing avoids switching video buffer pages, which is a rather slow operation since it requires writing to ports. However, this is only effective when a significant amount of video memory is required, such as at high resolutions or with a large number of colors. In this case, VLB or PCI is still required.

Indeed I remember that LFB increased slightly 320x200 fps during pcpbench on a pentium 100 using pci vga. For Isa and 386 I would also be surprised about a doubling.

Darmok wrote on 2025-12-27, 12:32:

In higher resolution modes, the ISA bus bandwidth begins to be insufficient. The theoretical ISA bus bandwidth at 8 MHz is 8,000,000 x 2/4 = 4,000,000 bytes per second. For a resolution of 640x480, 307,000 bytes per frame are required. The maximum theoretical FPS will be 13. For 320x200 mode, the maximum FPS is 62. Linear addressing avoids switching video buffer pages, which is a rather slow operation since it requires writing to ports. However, this is only effective when a significant amount of video memory is required, such as at high resolutions or with a large number of colors. In this case, VLB or PCI is still required.

Your math looks good, but your max frame rate only holds true if you redraw every pixel every frame. Games from this era did some tricks to reduce the total number of pixels that had to be drawn for each frame. Nothing like mpeg, but things like the hud and borders were not redrawn every frame. Could linear addressing help the game identify unchanged pixels?

VESA modes work a bit different in the hardware itself, since VGA modes have a bunch of hardware in the way that go all the way back to EGA times while VESA modes are just framebuffer scanout without any bitplane logic and other EGA/VGA specific things in the way. This is where the increases in performance come from.

Tiido wrote on 2025-12-27, 13:50:

VESA modes work a bit different in the hardware itself, since VGA modes have a bunch of hardware in the way that go all the way back to EGA times while VESA modes are just framebuffer scanout without any bitplane logic and other EGA/VGA specific things in the way. This is where the increases in performance come from.

LFB isn't suitable for real-mode games because it requires 32-bit addressing. In the case of DOS, these are games with a DOS extender. In most cases, these are 3D shooters. To improve performance, the developers used planar modes to avoid the overhead of bit planes. Planar VGA modes are similar to planar VESA modes. There shouldn't be a significant difference in FPS. Perhaps VESA modes are better optimized in the game code.

Workaround for 386dx40 is to oc the isa bus to 12Mhz, to get about 50% boost which is significant. It makes some games like Transport Tycoon playable, but not enough for Settlers 2.

Tiido wrote on 2025-12-27, 13:50:

VESA modes work a bit different in the hardware itself, since VGA modes have a bunch of hardware in the way that go all the way back to EGA times while VESA modes are just framebuffer scanout without any bitplane logic and other EGA/VGA specific things in the way. This is where the increases in performance come from.

And this is what s3spdup does: it places the VESA screen buffer at the standard addresss A000:0000, so while real mode VESA games still need to do bank switching, they get the "direct video memory access" bonus you would otherwise only get using the LFB.