I'm not sure about the UMC491, but the 480 chipset is decent, though probably not the fastest.
The SiS Rabbit was up there, but it has A, B and C revisions, so the C is probably the one you want. I've also heard claims that one of the later Forex chipsets is supposed to be the speed king. We have a thread of 386 Speedsys benchmarks. You can check it out and see for yourself.

Chipsets that I believe support interleave are SiS Rabbit, VLSI Topcat, and probably Symphony Hadyn. In theory the C&T PEAK/DM chipset should support it as well, but I have yet to see a board that actually uses it. My Abit FU3 for example has the option hidden, and when I enable it with AMISETUP the system is unstable.
I think you'll find DRAM interleaving is mostly on the pre 1992 boards, where the later boards support external writeback and/or double banked cache to make up the difference.

Yes, 386SX will run Windows 95...slowly.

The fastest is probably one of those 386/486 combo boards like we were talking about the other day:

386/486 boards? Motherboards with a socket for both a 386 and 486...

The UM481 and UM491 don't support 7+1 tag ram (as I found from one of my 486 boards) so if you get a board that uses those and no dirty RAM it's likely "Always Dirty" and slower.

Isn't 7+1 the scheme where one bit is borrowed from the tag RAM to create a dirty bit? I would imagine it should still be possible to have an extra chip to provide the dirty bit.

Anonymous Coward wrote on 2021-07-31, 05:33:

Isn't 7+1 the scheme where one bit is borrowed from the tag RAM to create a dirty bit? I would imagine it should still be possible to have an extra chip to provide the dirty bit.

Yeah, we've talked about the one board I have (EFA 4DMU-HL3S) that has the UM491, but does not have a socket for a dirty RAM. So it seems the only way to get rid of the 'always dirty' strategy would be to go after the board with a soldering iron to modify it to support a dirty chip. I don't want to do anything involving the UM491 pins for fear I would ruin it.

ok guyz, isn't that just the difference between write through and write back?
i would rather stick to WT because while WB is a few percents faster, i want larger cacheable range as i am stuffing 32mb into it to run win95 and quake2.

noshutdown wrote on 2021-08-01, 09:13:

ok guyz, isn't that just the difference between write through and write back?
i would rather stick to WT because while WB is a few percents faster, i want larger cacheable range as i am stuffing 32mb into it to run win95 and quake2.

I like to call them "read-only" and "read-write" caching. You can think of it as different philosophies of how to approach the issue of "slow" DRAM becoming out of sync compared to the cache RAM.

With write-back, if the CPU writes to a location that happens to already be in the cache, DRAM is allowed to become out of date and syncing it with the cache is delayed as long as possible. So if a memory location is being used as a spillover location to hold a register value in a tight loop, etc., this could avoid a lot of needless updates to memory. On the other hand, it needs additional logic as, if a device like a floppy controller wants to read directly from memory without going through the CPU, the chipset has to intercept this access and make sure there isn't yet-to-be-written-back, aka dirty, data still sitting in the cache before allowing this access.

Write-through is simpler and just says not to let this happen. When reading memory, the cache will speed up accesses if information is already there, but when writing it, both the cache and DRAM get updated right-there-and-then. That means that memory reads are sped up by the cache, but memory writes are not. This means the only special case is when the floppy controller writes to memory, that part of the cache has to be invalidated so the CPU is forced to go to DRAM to read that data rather than getting out of date information from cache.

Because DRAM can't be out of sync with write-through, you don't need a "dirty" bit to track the fact it has become out of sync, and that is why switching to it doubles your cacheable area for those caches that rob the dirty bit from the tag RAM.

Because L1 is much faster than L2, the benefit of write-back is higher for L1, and I know there are several members who run write-back L1 and write-through L2 to maximize their cacheable area.

Are you sure you want to run Windows 95 and Quake(!) on a 386?

386DX40 can run some 486 games at playable speed like Settlers 2, Command & Conquer, Lion King, but not 3d shooters or plane simulators. Windows 95 is going to be sluggish. I would not recommend installing it on anything slower than 486DX2 66. I remember back in the day when I had 386DX40, 486DX33-50 was used to run Windows 3.11 as 386 was deemed to slow for that. 3.11 wasn't too bad on 386 but there wasn't much usable software for it, unlike for Windows 95 and later.

Win95 and a 386sx? Why? Just because lolz ofc.

I still have that one coming up one day myself. For unobvious reasons.

Practically, no you wouldn't. For the best experience I think win 9x belongs on a Pentium MMX or faster, win 3.1x on a slow Pentium and power DOS on anything slower than Pentium. Even when Windows 98 will boot fine on any 486 as well.

Ok, win 3.1x on a 486 will run fine, and is usually ok for some early Windows games like SimCity 2000 or simtower. Also it provides some nice networking capabilities.

Are we talking "fastest" just in terms of CPU and RAM, or also video options?

For anything involving graphics, the fastest 386DX40 will be one with VLB slots not limited by ISA bus bandwidths, i.e. on a 386/486 motherboard, such as one with OPTi 495SX chipset. A 32b bus (could also be EISA or MCA) will also make I/O a lot faster.

I know for a fact, that larger Command&Conquer maps are unplayable with a fast 386, hey even a 25/33 MHz 486 will be slugish.
The graphics is basic, but the AI will take your cpu cycles.

dionb wrote on 2021-08-01, 21:13:

Are we talking "fastest" just in terms of CPU and RAM, or also video options?

For anything involving graphics, the fastest 386DX40 will be one with VLB slots not limited by ISA bus bandwidths, i.e. on a 386/486 motherboard, such as one with OPTi 495SX chipset. A 32b bus (could also be EISA or MCA) will also make I/O a lot faster.

i want "pure" 386 boards that don't involve 486 socket or rapidcad/486dlc upgrades.
the chipset may have 386/486 hybrid support, but if it has both 386 and 486 sockets, its considered a 486 board.

noshutdown wrote on 2021-08-02, 06:45:

dionb wrote on 2021-08-01, 21:13:

Are we talking "fastest" just in terms of CPU and RAM, or also video options?

For anything involving graphics, the fastest 386DX40 will be one with VLB slots not limited by ISA bus bandwidths, i.e. on a 386/486 motherboard, such as one with OPTi 495SX chipset. A 32b bus (could also be EISA or MCA) will also make I/O a lot faster.

i want "pure" 386 boards that don't involve 486 socket or rapidcad/486dlc upgrades.
the chipset may have 386/486 hybrid support, but if it has both 386 and 486 sockets, its considered a 486 board.

Correction: *you* consider it a 486 board. You can't go blaming others for not guessing prejudices you may have without actually making them explicit.

If it has a 386DX40 running on it, I consider it a 386 board. There definitely exist some where the 486 bits aren't implemented and there's just an onboard 386DX/40. How about that?

But regardless, it's your build.