Curiously I get a bit better coprocessor score in Landmark, on two different mobos as well, about 144 MHz equivalent of 80287. IIRC there were 2 variants of the Cyrix 83D87, the later one fixes some potential instability with some mobos but at the cost of reduced performance. And for some reason Cyrix also gimped some of the trigonometric functions, perhaps in effort to produce results better matching those of Intel chips. Can you post a photo of your chip? I can't quite make out the numbers on these.

Deunan wrote on 2024-05-25, 11:09:

Curiously I get a bit better coprocessor score in Landmark, on two different mobos as well, about 144 MHz equivalent of 80287. IIRC there were 2 variants of the Cyrix 83D87, the later one fixes some potential instability with some mobos but at the cost of reduced performance. And for some reason Cyrix also gimped some of the trigonometric functions, perhaps in effort to produce results better matching those of Intel chips. Can you post a photo of your chip? I can't quite make out the numbers on these.

Code is:
ABE306A

Thanks. Mine is DBE352B if anyone is interested, looks pretty much identical otherwise.

Tested with Cyrix 486DLC-40 with same setup. Will add results too.

Hi!
an additional cache chip (U32) slightly increases performance.

I know it sounds weird, but the ET4000AX *might* be a bottleneck (depending on brand/BIOS maybe).

On a 10 MHz 286 running Superscape benchmark, I got better score with an OAK OTI-67 once than I did with an ET4000AX.

Both the VBE TSR and the VGA BIOS TSR had been loaded with the OAK card, so ROM code was fully in RAM.

The fastest 286 results so far were with a 16MHz 286 and an Trident 9000.
I didn't do much testing with VGA cards, though.

Anyway, I don't mean to annoy anyone. It just came to mind while seeing this thread.
Maybe it's worth trying another VGA card, too, *just in case*.

I mean, the ET4000AX *is* pretty awesome. It also has buffers for interfacing ISA bus, which I think is neat.

Good luck! 🙂👍

PiotrUU wrote on 2024-05-26, 14:46:

Hi!
an additional cache chip (U32) slightly increases performance.

There is 64K onboard as you see, what is it for?

when there is no U32 - the cache works in Write-Through mode.
when U32 is present - the cache works in write-back mode
see here: https://theretroweb.com/motherboard/manual/33553.pdf

PiotrUU wrote on 2024-05-26, 14:46:

an additional cache chip (U32) slightly increases performance.

Ah, so that's what gave me a higher score- cache in WB mode. Good to know, I was considering buying another Cyrix '87 for testing and now I can avoid getting the exact same chip I already have.

Is it a special chip? Ebay available?

additional SRAM for dirty bit

is Standard SRAM 4X16kBit memory.
I bought 5 pieces for 8.5 euros directly from the seller. It's a bit more expensive on ebay (about 10.5 euros for 5 pieces)
these are 25ns chips - but as you can see, they work even at 40Mhz.
https://shop.tvsat.com.pl/en_GB/p/5pcs-CY7C16 … -CYPPRESS/26020

I've got a similar machine:

AMD 386DX-40
ULSI FPU
Shuttle HOT-327
UMC chipset
128 KB cache
Tseng Labs ET4000/W32i with 2 MB RAM

ULSI FPU is obviously faster than Cyrix.
But I'm confused about CPU results - notably slower in CheckIt, yet about the same in Landmark...

Grzyb wrote on 2024-05-27, 13:43:

ULSI FPU is obviously faster than Cyrix.

That depends on what you are doing with it. Basic math operations? Yeah, ULSI will finish first. Calculating trigonometric functions? Cyrix wins, and by quite some margin too.

I disagree that the ULSI fpu is faster than the Cyrix fpu.
I will say more carefully - there are tests where ULSI performs better than Cyrix 387+... ("black top")
However, remember that Cx387+ has an older brother Cx387 without the plus Cx387 (gray top) and older brothers always win 😉

let's take this test (packed in attachment - cannot add *.com file)

ulsi needs - 2m13s
cx387+ 2m21s
and the old cx387 only needs 2m09s, it beats ulsi by 4s!.
it's not much, but a win is a win.

I fully agree. ULSI does simple operations faster than cx387+, trigonometric functions better (much more accurate) and Cyrix calculates faster.

ps 1. I encourage everyone to play with the koproc.com test.
This is a test written in ASM, you have to calculate the time yourself, but it is small (requires 286/287 and VGA).
ps 2. I will take a photo of the Cyrix and ULSI fpu because there are many varieties of them, so you need to show them carefully.
(IIT FPU is no less than 7!!! varieties, Intel 387 there are 4!!! I don't know how many ULSI varieties there are - I have 3 Cyrix 387 also has at least 3 varieties)

Jo22 wrote on 2024-05-26, 15:09:

I know it sounds weird, but the ET4000AX *might* be a bottleneck (depending on brand/BIOS maybe). […]
Show full quote

I know it sounds weird, but the ET4000AX *might* be a bottleneck (depending on brand/BIOS maybe).

On a 10 MHz 286 running Superscape benchmark, I got better score with an OAK OTI-67 once than I did with an ET4000AX.

Both the VBE TSR and the VGA BIOS TSR had been loaded with the OAK card, so ROM code was fully in RAM.

The fastest 286 results so far were with a 16MHz 286 and an Trident 9000.
I didn't do much testing with VGA cards, though.

Anyway, I don't mean to annoy anyone. It just came to mind while seeing this thread.
Maybe it's worth trying another VGA card, too, *just in case*.

I mean, the ET4000AX *is* pretty awesome. It also has buffers for interfacing ISA bus, which I think is neat.

Good luck! 🙂👍

ET4000X gives the best result on Landmark but I have OTI77 and Trident cards too. Will test.

PiotrUU wrote on 2024-05-27, 16:58:

ulsi needs - 2m13s

My ULSI does it in 2m08s..2m09s.

Depends on the CPU, mobo, cache and BIOS optimizations. Test bed should be the same when comparing one part - FPU in this case.

In real apps the CPU makes much bigger difference then going from slowest to the fastest FPU.