Regarding the Cyrix/TI 486DLC, it boots up with L1 disabled on a mainboard where the bios do not know this cpu. But I read there is a tool to enable the 1 kB L1 cache.
Anyone has this tool?

You can find the tools you need on this page: http://ps-2.kev009.com/ohlandl/8570_Tim_OConnor/cyrx.htm

Now here's a quick update on what I've been up to:

I finally got my vintage PCs back. I've been playing with the TI486SXL-40 (8kb version of Cyrix 486DLC). Despite my system being "DLC Aware" it really isn't, as it doesn't hook up the FLUSH# pin. I can still enable the CPU cache using the above utility, but when using the BARB cache invalidation method supposedly there is a performance hit. As I was not satisfied with the performance I first overclocked my 40MHz chip to 50MHz using the internal clock doubling feature specific to the TI SXL chips on the 25MHz bus. This actually produced performance WORSE than 1x40MHz. Apparently clock doubling is not very effective on DLC type chips. I used to have a 486DRx2 33/66, but I found that it was only *slightly* faster than the 40MHz chip at CPU intensive activities while the memory scores were substantially worse.
Feeling ballsy, I attempted to push my 40MHz chip to 66MHz (2x33MHz). This was a fatal mistake. The system locked up immediately, and after reverting to 40MHz I was saddened to discover that my L1 cache was no longer working. I fried the on chip cache. A 66MHz version of the SXL chip actually does, exist but in 3.6V form. Apparently 66MHz at 5V was just too much. So let this be a word of warning to 486DLC users who want to push your 40MHz parts to 50. You could fry your cache.

Also, I did get around to testing my clock doubled FPUs. I ran three tests. One at 2x50, one at 2x33, and one at 2x40 (overclocked). The effectiveness of the clock doubling decreases as bus speed increases. At 2x25 the ULSI was 15-20% faster than a standard ULSI @ 25MHz, while at 2x40 it was perhaps only 5-10% faster than a 40MHz part. The IIT X2 was actually faster than the ULSI. At 2x40 it was actually more like 10-15% faster. I was using landmark v6 for the FPU test. I may attempt the test again later using something more conclusive. It's probably not a very good idea to overclock these FPUs too much though, as they are quite rare.

Did you try your fried SXL in other motherboards?

Was the SXL 33/66 QFP only? If so, ever see one on a QFP-to-PGA adapter with a VRM?

Did you run any burn-in tests with a DLC at 40 MHz and the fastest memory timings? I wonder you really need to step the timing down one notch to pass a 24-hr burn-in? If that is the case, then perhaps the DRx2 33/66 with IITx2 is the preferred option?

One other idea about the SXL2-50... I wonder if they run OK at less than 5V? If we run it at, say, 4.5V and try for 33/66 with an over sized fan/heatsink.

I currently only have one 386 board with a PGA socket. I used to have two others, but I sold them off. But you're right, there is the possibility that the board is causing the L1 cache problem....but I highly I doubt it.

As far as I know the TI486SXL2 33/66 was QFP only for 386SX and 486SX platforms (there wasn't one for 386DX), but it's available as PGA for the 486SX pinout (it's a 3.6V part).

The 144pin QFP 486SX part was sold on an adapter by evergreen or improve it technologies for 386DX pinout with VRM. I think a few people over on CPU world have them. I remember seeing one for sale on eBay a while back but I could never quite tell if it was for a 386 or 486 socket so I never went for it.

I haven't yet had much chance to do stability testing on my current board with fast timings, but it was quite solid on my ASI 386/486 hyrbid board I used to have. That board used the same chipset, but was much better with DLC chips if you could figure out the jumpers.

Try SXL overclocked using a lower voltage and see what happens. Currently I'm SOL.

Anonymous Coward wrote:

Try SXL overclocked using a lower voltage and see what happens. Currently I'm SOL.

It would certainly be an interesting test to undertake for any of the 386 CPU's at any frequency.

Giving this thread a bit new life:

same mainboard for all results:

386DX-40 with Cyrix FasMath FPU

Ti486DLC L1 off with Cyrix FasMath FPU

Ti486DLC L1 off with IIT4C87

Ti486DLC L1 on with IIT4C87

Seriously? The FPU type affects the CPU score?

I'm really starting to think that Speedsys isnt' suitable for anything slower than a 486.

Yes thats why for CPU benchmarks with seperate CPU/FPU scores are more suited, like NSSI that gives Dhry and Whetstone.
Also VESA Speed doesn't say much with Speedsys.
f.e. I get there on my 486VLB with the Spea Vega about 8 to 9 MB/s and about 25 MB/s for the ET4000.
Vidspeeds transfer scores however show that the ET4000 is only slightly faster than the Vega which is also reflected with more real world benches.

What is especially problematic is mixing of read and write transfer scores in one result. Read is not relevant for graphics cards. This problem has also Topbench for it's video score. VGA cards with high Read speeds score unbalanced high.

Here one of my 386 boards!

AMD 386DX-40
8x1MB memory 70ns
128KB Cache
Chipset is FOREX
Graphics card is a Tseng ET4000
I use an ISA controller / serial / parallel
2GB CF card with 512MB FAT16 partition
BIOS defaults

And here my other 386DX-40 system.

FX-3000 Rev 1.0
UMC UM82C482AF chipset
8x 1MB 70ns memory
256KB 15ns Cache
Diamond Speedstar 24X
IDE controller with floppy emulator and 2GB CF card with 512MB FAT16
BIOS defaults

Thanks for posting these. I was really curious about the Forex. It doesn't seem to live up to it's performance legend as far as I can see though. The UMC is also very much appreciated.

What performance legend 🤣

Are the specific BIOS settings I should try out? I always test with BIOS defaults.

The Forex achieved 16.1 in 3dbench, the UMC is faster achieving 16.6. That's with BIOS defaults on both systems. I don't have 15ns cache chips on the FOREX however and only 70ns memory.

EDIT:

FX-3000 board with tuned BIOS settings. I put all the WS to the lowest settings that the system would accept.

EDIT:

Is there a better / other program to test memory throughput? Having to have 8MB in the system is annoying 😀 Also Cachechk comes up with different figures. Which ones are correct?

And here the 3dbench score with tuned BIOS:

These results look fairly low to me.

Could you display the UMC Speedsys screenshot with the CMOS memory/cache timings optimised?

feipoa wrote:

Could you display the UMC Speedsys screenshot with the CMOS memory/cache timings optimised?

I have, just one post above.

16.6 in 3dbench is very high according to the VOGONS 3dbench database. There are only 2 faster systems, one being VLB.

I found DOOM -timedemo demo3 very good at showing BIOS settings differences. Speedsys and CACHECHK show different results so not sure which ones are accurate.

Does using a 486DLC or 468SXL in the UMC board improve memory performance in any way?

I can tell for my SIS board that it does a bit, according to the scores from 24.96 MB/s to 26.31 MB/s.

Try tuning the forex.

Could you guys run 3dbench 1.0 on you DX-40 systems with and without cache?

Tuning isn't really my thing. I prefer BIOS defaults and stable operation 😀

To me the vesa speeds of 5-5.7meg are impressive if that's at standard bus speeds and the 3Dbench scores are very good, I never got my 386 systems those levels, what scores do you get for the doom timedemo? I ignore the ram speeds I prefer to guage actual performance, the cpu score of 9.33 I don't think ive seen that high for a 386DX-40 either, seems like both systems are good to me.

CPU score is not accurate in speedsys. As you see above the 486DLC-40 upgrade CPU reaches only 9.00, while it is definitely faster than a 386DX-40. Then the score goes significantly down just by plugging a FPU. So you may compare these values only for the same CPU/FPU setup.

From my experience tuning the waitstates on a 386 system isn't much of a risk to get a unstable system. If you have too fast values it will just freeze after BIOS. I never got the case that it was "just a bit" unstable. You have also to consider that nowadays you usually don't see the 80ns SIMMs anymore. They are mostly 70 ns or 60 ns which is fast enough to lower the WS.