I finally finished testing AMD DX4-120 (@100WT, @100WB and @120WB). I was very surprised that I did not had to change memory timings at all (I ran memtest86 before any benchmarks) and even more surprised that I did not had to set any VLB wait states! 40 MHz bus speed was very stable. Not a single crash or any other problem.

Now I started testing AMD 5x86 and I can already say that there have been many small surprises. So far it looks like the Intel DX4-100 WT (the original CPU for this computer) looks like to be worst choice. Intel DX4-100 WB is much better but not in the way you might think. And since AMD DX4-120 was perfectly stable, I really need to retest the Intel DX4 WB @ 120MHz. And I haven't even tried the Cyrix 5x86 yet and AMD 5x86 needs to be tested under- and overclocked. So lots of testing left to do.

Can someone explain why latest Topbench detects my CPU as "X5+ (Am5x86+) "? There are non-plus 5x86 results in the database so how is the "+" is detected? CPU ID is 4F4 as it should be. Bug in Topbench or is there some way to tell apart plus and non-plus CPUs?

aitotat wrote on 2022-06-11, 07:20:

Can someone explain why latest Topbench detects my CPU as "X5+ (Am5x86+) "? There are non-plus 5x86 results in the database so how is the "+" is detected? CPU ID is 4F4 as it should be. Bug in Topbench or is there some way to tell apart plus and non-plus CPUs?

Are you overclocking? I've seen 160mhz 5x86 being referred unofficially as a 5x86+ or a 5x86+P90

No, not yet. This is set to 4x33MHz as it should be.

The Am5x86 is definitely non-plus since when overclocked to 160MHz it requires 3.6V just to be able to boot to DOS. And with 3.6V it is not stable and cannot run Doom benchmark (but strangely it does pass Quake benchmark).

The Cyrix 5x86 I have seems to be a great individual. It works perfectly at 120 MHz with 3.45V. It is Stepping 1 Revision 3 so it is perhaps surprising it does so well at 120 MHz. However the old stepping might be a problem. I cannot change multiplier (from DOS prompt) at all. Setmul tries to change it but it just does not change. Is it only the later revision of Cyrix 5x86 that has this feature?

aitotat wrote on 2021-08-26, 09:19:

I think the Labway will not be the last sound card for this system. When someone makes a new SB Pro and SB 16 compatible card it will be perfect. But for now the Labway is very good here. Orpheus might have been even better but I don't need second intelligent mode midi controller and the Labway sounds very good so no need for Orpheus.

I think you will be very interested in the upcoming MK8330 Sound Card.

Just a quick update: I found out why changing multiplier did not work. It was because I had URAM (UMB driver) loaded. I don't know why it conflicts. More benchmarking left to do now...

Edit: URAM alone was not the problem. I was testing with L2 cache disabled and forgot it disabled when I first tested without URAM. Without URAM but L2 cache enabled, the clock is set to 33MHz but system either freezes instantly or after I try to do something, like access hard drive. Not sure if this is issue with CPU stepping/revision, or maybe the chipset is too old to support Cyrix 5x86 fully (SiS 571 is very good but it is 1994 chipset released before Cyrix 5x86). Well at least I can get some results by testing 33 MHz with L2 disabled and they will show how useful the 1x multiplier could be if it would work properly.

Edit 2: It seems that Cyrix internal registers are accessed by the same port 22h/23h pair that SiS chipset configuration registers are accessed. There should not be conflict since SiS471 does not use the same register indexes as Cyrix but I suspect this is the cause of the problems anyway. UMB driver has to program the chipset through those ports and Setmul changes the multiplier through those ports. Something goes wrong somewhere. I'll have to make my own test program to see if I can change the multiplier without issues.

appiah4 wrote on 2022-06-13, 09:00:

I think you will be very interested in the upcoming MK8330 Sound Card.

Yes, that would be the bugfree SB Pro and SB16 compatible card I was hoping for. And I really would like that card but not for this system.

I've been thinking more and more about period correct builds and I would like to rule out PnP cards since this is a non-PnP system (BIOS does not have PnP support). I know this might be a bit weird since I'm going to use 650W ATX power supply and I have 60 GB 1,8" HDD and support for CF cards.

The idea of ruling out PnP card from this system started when I found SB16 CT2940 with real OPL3 (those cards usually have CQM). It has the "Vibra Pro" chip and it did not have the Vibra clipping bug. It sounded very good and did not have any serious issues except MPU-401 bugs (I did not even test those) so CT2940 simply was much better choice than the SB AWE64 Gold.

So it was YMF + CT2940 then, both PnP cards (as was the AWE64 Gold). There was not much reason for the YMF except to hold X2GS (because CT2940 has the MPU401 bugs) and it was also useful for MIDIto to emulate dual OPL mode. But to make things simpler, I replaced the YMF with Roland SCC-1A since it was a period correct for this system. X2GS would be better on my Super Socket 7 machine.

So that left CT2940 as the only PnP card on this system. I have another great SB16, CT2290, that suffers from the single-cycle DMA clicking bug. It is a great card since the clicking is much more quieter than on any other clicking SB16/AWE32 cards I've heard and CT2290 is period correct for this system and also sounds really good (it has the best possible DAC, same as on AWE64 Gold). But CT2940 is a great card without the clicking so it simply is a better card so I couldn't really decide.

But I didn't have to, not since I've found GUS Extreme again and that is what I'm going to use. It is a very late card, late 96 or early 97, so it would be somewhere between CT2940 and YMF. But it is a non-PnP card and GF1 is old chip released in 92. ESS1688 is early 95 chip so it is period correct for this 1995 486 system. GUS Extreme is a very good quality card, no noise at all and ESS is very compatible. It does not have real OPL3 though. And since I have SCC-1A, I don't need waveblaster connector that the GUS Extreme does not have.

So now the sound setup is GUS Extreme + Roland SCC-1A + Roland MPU401AT with McCake (or at least it will be once I've finished all the CPU tests. I'm using SB Pro2 CT1600 during the tests to see at what speed real OPL3 starts to fail). So all I'm missing is real OPL3 and I really would want one of these to fix that. There simply aren't a better choice to provide real OPL3 for this system! But I can live without real OPL3 if I have to. ESFM is really good and I don't even know if I can use the Adlib Gold clone. There might be Adlib port conflicts that cannot be solved since the GUS Extreme is NOT a PnP card (now how did it come to this...)? Edit: So whatever I do it always comes to this: It is impossible to decide what sound card(s) to use.

aitotat wrote on 2022-06-14, 05:06:

I've been thinking more and more about period correct builds and I would like to rule out PnP cards since this is a non-PnP system (BIOS does not have PnP support).

There should be a few Vibra models from around that time which are non-PnP and have genuine OPL3 on board.

I have one of those weird SoundForte varieties which uses a CT2501 chip paired with an YMF262 + YAC512 combo. I haven't tested it extensively, but from what I remember, it has the standard MPU-401 bugs, as well as the Vibra distortion issue, but no DMA clicking. The FM synth output is very clean once you disable the on-board amplifier via jumpers.

I do have Sound Blaster 32 CT3930. It has the CT2501 and external Yamaha OPL3. It is a good card but it does have the Vibra clipping bug. I even did some recordings and compared it to the Vibra Pro. CT3930 is one of the best non-PnP cards by Creative but GUS Extreme is so good that I'll gladly go with it.

Sounds like you need to make do with a quality ES688 non-PnP card with onboard (and likely 100% clone) OPL3.

aitotat wrote on 2022-06-14, 07:41:

I do have Sound Blaster 32 CT3930. It has the CT2501 and external Yamaha OPL3. It is a good card but it does have the Vibra clipping bug.

That's a nice card indeed. If you want to pair it with a non-PnP SBPro compatible card, you can also use an OPTi 82C929 or 930. I have the latter and I'm pretty happy with it. You can read my full review here.

BTW, there are brand name varieties of these cards which offer cleaner sound and sometimes even include an OPL4 chip. They tend to be rare nowadays though. The Miro Sound PCM 10 is one such example.

That Miro sure would be nice to try. It even looks really nice. I've never even tested any card with OPL4. Same goes for those old ESS cards. The ES1688F is the oldest ESS sound chip I've tested. I'd like to build some nice 386 system someday (I have a 386/486 hybrid board but I don't like it at all because it is essentially a 1994 486 board that just happens to support 386). So I'd like to have some quality non-PnP sound cards to test for it and maybe for this system as well. But I'd like it to be a 1992/1993 sound card for a 386 system and that limits a lot.

Speaking of quality sound cards. Since I now have SB Pro 2 CT1600 during the CPU tests, I was wondering why I hear very faint Topbench clicking sounds when I do not have PC Speaker connected. I very soon understood that I now have "Creative quality" installed.

But back to Cyrix 5x86. I made my own test program to set the multiplier to 1x. Results were the same as with Setmul. I also tried the Evergreen ET5x86 utility to modify the required register but no difference. To rule out any conflicts between SiS471 and CPU I tried to set the multiplier to 3x when it was already 3x. Nothing happened. If the CPU would have messed the SiS registers, then it would have most likely messed it again. But no, so it likely is not address conflict. CPU might have passed the data on the bus (although it should not according to datasheet) and maybe SiS mirrors some of its registers there but that was not the case. And if it were, changing the multiplier did work when L2 was disabled and URAM not loaded.

I started to experiment BIOS settings. I found out that when I disable "DRAM Write Burst" from Chipset settings, I can keep L2 enabled and multiplier can be changed to 1x and back. But I still cannot use URAM. When URAM is loaded, changing to 1x multiplier instantly freezes the system. I managed run all the benchmarks without any issues with "DRAM Write Burst" enabled so why does it have to be disabled when changing to 1x multiplier?

CPU must do something else in addition to just changing the multiplier. And the 1x multiplier is not really even required. Disabling L1 alone will be very useful and Turbo button can do the rest as with all the other tested CPUs. Multiplier 1x in combination to turbo button provides some speeds that are not available for other tested CPUs but those are not required. I think Cyrix 5x86 is best with those last 486 chipsets with PCI support. But I would like to test the S0R5 version of Cyrix 5x86. Maybe it behaves differently. I'm going to test some other UMB driver and just forget the 1x multiplier if it changes nothing.

I think I've tested 1x multiplier enough. I tested another real mode UMB driver (UMBSIS) and no difference. Then I tested emm386 to provide the UMBs and it worked so 1x multiplier does work on V86 mode.

So back to real mode testing. I adjusted UMBs to only D or E segment but that did not help. Next I loaded DOS low and something changed. I was able to successfully set multiplier to 1x but loading norton commander did not work (DOS file buffers were still in UMB segment). But I was able to use setmul and set multiplier back to 3 and then everything worked again. For some reason the CPU cannot properly access UMBs provided by chipset when in real mode and multiplier is set to 1x. I did look a SIS471 chipset datasheet and D and E segment UMBs are always uncached (at least only C and F segments have bits to make them cacheable). So either there is compatibility issue with CPU and chipset or this revision of CPU is buggy in real mode with 1x multiplier.

I did test some of those performance options from CPU control registers (well, I disabled everything) and it did not help. So I suppose I have to accept that Cyrix 5x86 1x multiplier and real mode UMBs are not possible together on this motherboard. I really would like to test the other Cyrix 5x86 CPU revision someday.

Time for some CPU results. I'll just post the result now and analyze the CPUs later. But first about the test setup. It looks like this.

I'm using the Corsair PSU I modified before. A smart fuse is connected between mainboard and the PSU. I thought one of those smart fuses might be a good way to protect the motherboard and other components. I don't want faulty PSU to break all the rare and expensive stuff. But smart fuse just makes the ATX to AT cabling hell even more worse and there is no good place to connect it. It should have been made to fit 3.5" or 5.25" drive bay and even then that would not solve the mess with cables. But at least the display would remain visible and buttons could be operated. It would have been useful since one of the buttons is ATX soft power switch. But now it is better as a test bench equipment. It displays power consumption and I used it to get peak power consumption from 5V rail. The reading includes not only the CPU but motherboard, VGA, multi I/O and the Lo-tech card for XTIDE Universal BIOS and Sound Blaster Pro 2 CT1600, so everything that is connected to mainboard.

CT1600 was used in the tests because I wanted some old card with Yamaha OPL3 in it to test speeds where problematic games start to behave badly. OPL3 is not to blame, the problem is that the games do not have proper delays between writes to the OPL. OPL2 is much more sensitive than OPL3 and I wouldn't use OPL2 cards on DX4 or DX2 systems.

There are other speed issues than just the OPL3 issues. For example many Sierra games are speed sensitive and you can have weird issues that does not seem to relate to too fast CPU unless you know to suspect it. I did not test them here but I know from experience that the DX4-100 WT slowed down with de-turbo (1/3) makes the system slow enough for Sierra games. And that is the very reason I like DX4 systems so much. DX4 systems can be slow enough and fast enough for a very wide range of DOS games and turbo button on the computer case makes things easy and 486 systems have that good old feeling that 286 and 386 systems have but Pentiums do not. But the DX4-100WT I've used on this computer (it is the original CPU) isn't ideal for the OPL3 speed issue. De-turbo (1/3) alone is not enough and neither is disabling L1 cache. So I need to use them both and that makes the system much slower than necessary. The idea for this test was to find out ideal solution for that problem. Since there are two de-turbo modes on SiS471 chipset (1/3 and 2/3 but BIOS setup does not let to choose the one to use) I thought that de-turbo 2/3 + L1D might be a very good fix. And it is. It will double the speed! But it turned out that EVERY other tested CPUs are slow enough for OPL3 problems with L1D alone! And there are other small surprises as well.

So here are the results and I'll analyze the CPUs on later posts.

aitotat wrote on 2022-06-18, 06:58:

DX4 systems can be slow enough and fast enough for a very wide range of DOS games and turbo button on the computer case makes things easy and 486 systems have that good old feeling that 286 and 386 systems have but Pentiums do not.

I'm not sure about the original Pentiums, but a Pentium MMX can be slowed down to a wide range of speeds using SetMul. Phil's showcases that in a video and runs a ton of benchmarks in another video.

As an example, my MMX 166 can be slowed down to 386DX-25 speed with both caches disabled and 486DX-33 speed with L1 off and L2 on. It can also reach some other speeds like a 486DX2-66 or a Pentium 90 by toggling various test registers via SetMul, though I haven't experimented too much with that.

Yes, Pentium MMX is excellent for slowing down and those systems are more common than good 486 systems or good SS7 systems with K6 plus CPUs. So it would be cheapest and easiest to go with MMX Pentium. And Pentium MMX systems would be great for Voodoo 1 and that has more relevance with DOS than any other 3D card. But for some reason I like 486 more, and 486 with VLB to be exact (I don't want 486 with PCI). I can't really explain it.

I used to have Pentium 120@133 back in the days. It slows down nicely by disabling L1 Cache. But Setmul was not available then so I always had to go to BIOS setup to disable L1 cache. Setmul sure is useful, and not just for Pentium MMX. I eventually upgraded my Pentium system to K6-233. It was much faster than my friends P200MMX (at least with ZSNES that we used a lot then, with Quake 2 it is the other way around). Perhaps that is one reason why I don't feel comfortable to use Pentium MMX. One other reason is that they are more of a Win9x systems. I like DOS.

aitotat wrote on 2022-06-18, 08:18:

Perhaps that is one reason why I don't feel comfortable to use Pentium MMX. One other reason is that they are more of a Win9x systems. I like DOS.

True, the MMX was released in early 1997 when DOS gaming was on its last legs and Win9x games really started picking up speed. There were still a few high profile DOS titles released during that year (e.g. Dungeon Keeper, Carmageddon, Shadow Warrior) but it was essentially one last hoorah before the end.

And I fully agree that a 486 is more period correct and fitting for DOS. I don't personally have any nostalgia for that platform since I started off with a Pentium 133, but I still appreciate the role that 486 systems played in PC gaming history.

I just added two bonus benchmark results. I was about to disassemble my test setup and though of a test I have wanted to do a long time ago. I have Diamond S3 Trio64 VLB graphics card and I wanted to test if it is better than the #9 S3 Vision 864 I have been using (original card for this system).

Diamond had 1MB RAM installed so I first tested with it. Then I upgraded RAM to 2 MB. I had Am5x86-133 installed so I used it to do the tests. The results? Almost as I expected. Upgrading RAM helped only SVGA benchmark. And the Trio64 is supposed to be Vision 864 with integrated RAMDAC so there should not have been any differences. And there weren't, except for Topbench. Maybe the #9 have tighter timings although it has 70ns RAM while the Diamond has 60ns. There could have been clock rate differences as well but that would have shown on all benchmarks. I know PCI Trios and Virges do not always use same clocks among different manufacturers and there are differences in image quality as well even though they have integrated RAMDACs.

The external RAMDAC on S3 Vision cards can be better than the integrated S3 and I know this #9 is a very good quality card. I wouldn't have changed it anyway since it is the original card for this system (and looks much better, more professional).

Lets start with Am486 DX4-120 with 8k WB cache. Unfortunately I forgot to run speedsys twice so I only have 100 MHz WB results. I wanted 100 MHz WB and WT results for better comparison with Intel DX4.

Clock for clock this is the slowest CPU I tested. At 100 MHz with 8k WB cache it is slower than Intel with 16k WT cache. At full 120 MHz it is only a little faster than Intel CPUs at 100 MHz. But even at full 120 MHz it consumes less power than Intel CPUs, thanks to 500nm core (Intel has 600nm) and of course half cache means less transistors to produce heat. AMD also has 500nm core with 16K cache but it is supposed to be rare. More likely AMD DX4 CPU with 16k cache is actually 350nm 5x86 core and that might be a bad thing because it has a problem when disabling L1 cache.

But nothing wrong with this CPU, no issues or crashes at all. It is simply slow and needs the extra 20 MHz to keep up with 100 MHz Intel.

Next the original CPU from this system, Intel DX4-100 with WT cache.

As I mentioned on some previous post, to be able to slow down this for most problematic OPL-games, de-turbo (1/3) or disabling L1 alone is not enough. Both are needed and that makes the system much slower than necessary. I started tests with this CPU and was really happy that the new de-turbo 2/3 -mode doubled the speed over the previous 1/3+L1D combination and was slow enough for Cycles and likely other problematic games as well.

When testing the other CPUs it came clear that for ALL the other CPUs, even the overclocked ones, it is enough to disable L1 alone. Now how is that possible? I blame chipset. Motherboard has jumpers JP5 and JP6 (hardware trap) that are used to group different types of CPUs. In first group there are all traditional SX, DX, SX2, DX2 and DX4 CPUs with only WT cache. Next group has Cyrix CPUs and last group is for Intel and AMD WB CPUs. So from all tested CPUs this is the only one from group 1. I'm sure the chipset does something differently in this group. This also means that if you have motherboard with some other chipset, then likely the usable slowdown options are different. But most likely L1D + de-turbo combination is enough.

But this was not the only issue with this CPU. It overheated during the tests. At one point system crashed. I rebooted and it crashed very soon again. Heatsink was quite warm but not too hot to touch. Then I turned the PSU as seen on the picture: it sucks air through the CPU heat sink. No problems after that. This CPU was tested with different heat sink than the other CPUs. This one has original heat sink glued on. I don't know what glue is it or how good is it when it has gotten old. Maybe it acts more as thermal insulator than thermal conductor by now.

So I'm going to replace the CPU from this system. The latter problem could be resolved by removing the heatsink and glue but I can't do anything with the L1D + de-turbo issue. Chipset simply handle it better for other CPUs. But otherwise Intel DX4-WT is fast and good CPU but it does use more power than any other non-overclocked CPU and so it produces more heat.

Next we have Intel DX4-WB:

Little faster than WT model and consumes little less power and L1D alone is enough for the problematic OPL games. So in every way better than WT model. But at 120 MHz this is the second fastest CPU I tested. Cyrix has faster FPU but that is not really needed for 486 games and Cyrix does win more tests but loose some. Overclocking does increase power consumption and heat but only little above Intel DX4-100 WT. Very good CPU. I was surprised how good it does at 120 MHz.

I did have one crash during the tests but I think it was my fault. This was the second CPU I tested and I sometimes left the OPL3 in error state after unsuccessful test with Cycles and Indy3. I then stopped doing that (I reset the system before starting new tests if OPL test did not pass) and there were no problems after that with any CPU. Overclocking test had no problems at all.

Next Am5x86-133:

133MHz 4..no 5x86 sure sounds impressive, doesn't it? Too bad the AMD DX4 core is inferior to Intel, at least if speed is the main thing and by naming it to 5x86 is surely must be.

At worst the extra 33 MHz is needed to match 100 MHz Intel DX4 WB. But at its best Doom was 17% faster and Doom is a real world benchmark. So it is nowhere near as fast as 133 MHz suggest it to be. But it was never really slower than Intel DX4-100 WB but uses a lot less power so it produces a lot less heat. This is the best feature of Am5x86. There is very little speed difference between 120 MHz and 133 MHz. What you lose from CPU clock you gain from memory clock. But with L1 disabled, the 120 MHz is the faster one. So overall, Am5x86 is faster at 120 MHz than at 133 MHz. And as a bonus, it then identifies "correctly" as Am486 instead of Am5x86. I really hate to call this 5x86.

But this CPU has issue with disabling L1 cache. I observed this with maybe at three different speeds so this is not a one time glitch. Disabled L1 cache seems to partially enable randomly. I had to retest several L1D tests because the results were 50-80% too high and this very little L1 cache did not get enabled at same point every time.

For example when overclocked to 160 MHz, I started from L1D test the usual way. First Quake or Doom, then the other of those two. They are so sloooooww so I'll just do something else while the tests are running. Next Doom with min graphics, then PC Player tests (first SVGA), then 3DBench and very last Topbench. It was 3Dbench where the false readings started. But on other tested speeds it was at some other test at much earlier.

And the 50-80% is too much of an increase in speed that there will be OPL related issues on the most problematic games. No other CPU had this kind of issue and it happened many times. Too bad, Am5x86-133 would have been great CPU otherwise.

Last Cyrix 5x86:

Really good CPU. Fast with little less power consumption than Intel CPUs. Overclocked really well to 120 MHz: completely stable with 3,45V and power consumption increased only little. PC Player benchmarks favors Intel quite a lot for some reason. Other than that, the Cyrix is faster and with some control register tuning it gets even faster. But still not fast enough to beat Intel at PC Player benchmarks. Overall, Cyrix is the fastest 100 MHz CPU and fastest 120 MHz CPU. 100 MHz is not enough to outperform Am5x86-133 but 120 MHz is enough for that, even without tuning.

But there is the issue with 1x multiplier. I'm unfair here because it is a feature that the other CPUs do not have at all. And even the Cyrix does not really need it since same L1D trick slows down just like with the other CPUs (expect Intel WT that needed de-turbo as well). So first I needed to disable the certain BIOS setup feature to be able to use 1x multiplier with L2 cache enabled. Even then it did not work with UMB driver but it did work if EMM386 was used to create UMBs.

I think I have explanation for that. Cyrix 6x86 does have a feature in its control registers to write protect 640k-1MB memory area. That is where the UMB area is. Cyrix 5x86 does not have that bit in its registers but that write protect would explain perfectly the issues I'm having. With EMM386 those UMBs are not actually located in the 640k-1MB area but remapped to somewhere else. The physical UMB area is reserved by chipset unless real mode UMB driver is used.

So what I think is that this S1R3 CPU I'm having somehow manages to enable the write protection while entering 1x multiplier mode and when switching back to 3x multiplier mode, the write protection also disables. So a CPU bug. That is why I would be very interested to try the S0R5 CPU (that should be later even though it is Stepping 0).

Since I would most likely use the 1x multiplier only when using EMM386 as well (so I can use MIDIto) this is not a big issue.

I don't like 486 systems with PCI but I understand the benefits: much easier to find a good VGA card, integrated I/O so BIOS can enable PIO4 and one big card less and so on. If I would build a PCI 486 system, I'd definitely go with Cyrix 5x86. It is a very good CPU and would be a perfect match for a late 486 PCI motherboard.

But what do I choose for this system? I'll think about it some more. It is either Intel WB or Cyrix. Am5x86 doesn't belong here, not with its L1D issue. I do admit that lower power consumption would be nice.