I tested Jazz 1.0 and Tyrian 1.0 and they indeed work fine with MMX 262. That is very good!

I could try those games on a faster system just in case but lets make things more interesting.

Let's try these instead on this motherboard

aitotat wrote on 2023-04-07, 07:15:

I tested Jazz 1.0 and Tyrian 1.0 and they indeed work fine with MMX 262. That is very good!

Very nice! Looks like the wiki page was right about the MMX. I'm wondering if there would be any difference with a classic Pentium 200, but that seems unlikely.

I could try those games on a faster system just in case but lets make things more interesting.

If the wiki is correct again, those K6 CPUs should be good up to 500 MHz. On the other hand, Pentium II and Celeron CPUs apparently trigger the error even at 233 MHz.

Joseph_Joestar wrote on 2023-04-07, 07:44:

If the wiki is correct again, those K6 CPUs should be good up to 500 MHz.

Might very well be. Jazz and Tyrian certainly works with K6-2 set at 6x75=450 MHz. Unfortunately that is the fastest this board can do (well there is the unofficial 83 MHz bus speed but I'm not going to find memory timings etc for that). Well, at least that "200 MHz" bug won't bother this motherboard. Well maybe the L2 cache changes things a bit on the K6-2+. I'll test that once I've run all the benchmarks with non plus.

As I mentioned earlier, I eventually upgraded graphics card to Hercules Stringray 128/3D on my 1996 Pentium. Not the wisest upgrade and the dual planar design caused problems later when I tried to upgrade my computer to K6. The idea was to replace motherboard with a Gigabyte motherboard with Intel TX chipset and install AMD K6 233 MHz to it. I think that was the fastest AMD chip then and best I could afford. Pentium 2 was too expensive.

But the Dual Planar Voodoo Rush card did not fit on any of the PCI slots. CPU was in very bad location. One long PCI card could have been installed but not the dual planar Voodoo Rush. So I had to replace case to ATX and get ATX motherboard instead. So the fact that I chose Voodoo Rush instead of Voodoo 1 costed once again here as well. Voodoo 1 solution would have been cheaper and better in every way. That Voodoo Rush fiasco was the first and last time I preordered anything without reading reviews first.

I liked the 233 MHz K6. I had to go to the BIOS to disable L1 cache to slow down the system. Just like I had to do with the Pentium 120@133. It was a very fast CPU. Eventually I upgraded to Voodoo 2. My friend had Voodoo 2 as well but with a Pentium MMX 200 MHz. I remember it could be easily seen that the MMX ran Quake 2 faster even though it was only 200 MHz while my K6 was 233 MHz. But my K6 was much much better at ZSNES.

I don't have a K6 (the first one without 3DNOW) to test but that does not matter. Underclocked K6-2 should perform just like the K6 when nothing uses 3DNOW instructions. However, the tested K6-2 450 MHz is a Super Socket 7 CPU! How can it work on this motherboard? No 100 MHz bus, voltages are different etc. Luckily this motherboard does support the low voltages, although they are not printed on manual or on to the motherboard. Here are the necessary jumper settings. So to get 2.2V for the K6-2 CPU, jumpers must be placed to 2.5V AND 2.7V. And to get 2.0V for the K6-2-plus CPU, jumpers must be placed to 2.5V, 2.7V, 2.8V and 2.9V.

What about multipliers? There are only two multiplier jumpers so multipliers from 1.5, 2.0, 2.5 and 3 are possible. Pentium MMX and AMD K6 use multiplier 1.5 as 3.5. But the 450 MHz K6-2 should use 4.5 x 100 MHz. There is no multiplier locking on K6 CPUs and they use multiplier 2.0 as 6.0. That makes possible to use 6x66=400 MHz and 6x75=450MHz. So without modding the motherboard (it is possible to add the third multiplier jumper), multipliers 2.5, 3.0, 3.5 and 6.0 are available for the K6-2 on this motherboard. It is possible to use the missing multipliers with a plus-CPU since they allow multipliers to be set with software while system is running. But are they even needed? That needs to be tested first. I haven't tested the plus CPU, only K6-2. And that K6-2+ CPU is unlocked to K6-3+ CPU so those are actually K6-3 results that are missing from now.

There were some surprises with the K6-2. To get better DOS performance, write combining should be enabled for the K6 CPUs. Here are everything that is needed. I didn't know about that when I had the K6 CPU. So I did couple of benchmarks without enabling write combining and write combining is something that is better to keep enabled.

Although K6 has weaker FPU than Pentium MMX or even Pentium Classic, it is no surprise that 400 MHz is enough to make it faster that any other tested CPU. At least with L1 enabled. K6 suffers from disabled L1 more than any other tested CPU so even at 450 MHz (remember, 75 MHz bus here, not 100 MHz), the CPU is slow enough for everything. There are some surprises here. For some reason the original "386" version of Indy3 works with K6 and L1 cache disabled! That sure was surprising, especially since a-parameter is needed for the "486" version! But there are more surprise results. With L1 enabled and CPU set to 400 MHz, Indy "486" works with a-command line parameter and Indy "Pentium" work fine. But not with 200 or 233 MHz! What is going on here? And I double checked the results.

Must be something to do with the high multiplier, synchronizing the CPU clock to FSB? I have no explanation, especially since K6 @ 262 (3.5x75) work like the 400 MHz so better than 233 MHz.

Anyway, K6-2 400 or 450 would be an excellent CPU if the CPU does not need to be period correct. I expect the plus-CPU to be even better. But I think the K6-2 deserves a proper Super Socket 7 motherboard (but the SS7 chipsets have their own problems).

Here are the results.

Edit: I forgot to mention that I also tested the K6-2 450 (6x75) with Voodoo 3 2000 PCI. The idea was to try how much a very fast 2D PCI card would help. Surprisingly little, actually. Even though the Voodoo 3 has superior memory speed but that does not help here. PCI is the bottleneck because everything is rendered with the CPU and Virge is already fast enough. But I'm going to test other video cards. Maybe there will be some surprises.

Also, for the K6-2 testing, I updated motherboard BIOS to the 0207_J4 found here.

Started testing K6-2+, or rather K6-III+ since it is unlocked. This is going to take a while. I kind of assumed that with L2 cache disabled it would perform the same as K6-2. But no, it does not. So for all the possible combinations, I would have to test this with all multipliers, both caches enabled, both caches disabled and with only L1 or L2 cache enabled. And then repeat all that with 75 MHz bus speed.

There seems to be very little difference if L2 is enabled or not when L1 is disabled. Is this normal? I expected that with L2 only I would get quite nice slowdown results but that does not seem to be the case. In fact, it looks like K6-2 with L1 disabled is a bit faster than K6-III+ with L1 disabled and L2 enabled.

Edit: With L2 cache I mean the L2 cache inside the plus-CPU. I'm not going to disable motherboard cache in any of the tests.

aitotat wrote on 2023-04-12, 04:12:

I'm not going to disable motherboard cache in any of the tests.

Why not?

On my MMX 166 system, I get 486 speeds with L1 disabled and 386 speeds when both L1 and L2 are disabled. Benchmarks here. And Phil has a much more extensive analysis of how motherboard cache affects things on this page.

Because I would have to go to the BIOS to disable motherboard cache. That is too much trouble. I want everything to be available from command line.

About that page you linked. It shows those nice Pentium MMX test register features that I did not test when I tested Pentium MMX. But it seems that Pentium Classic also supports BPD and VPD (I looked at Pentium datasheet). It does not support CCD and DCD, those are for MMX only. However, I think Setmul only supported L1D and L1E for Pentium Classic. So even more investigating and testing ahead...

aitotat wrote on 2023-04-12, 05:43:

Because I would have to go to the BIOS to disable motherboard cache. That is too much trouble. I want everything to be available from command line.

Heh, too much trouble to go into the BIOS and change a single setting? Your choice, but you're deliberately excluding a method which would allow you to play some early 90s games like Wing Commander at the correct speed.

Personally, I just toggle this when a game absolutely needs those early 386 speeds. I don't use it to get more range from the test register settings as Phil does.

aitotat wrote on 2023-04-12, 05:43:

That is too much trouble.

I thought people who don't like "troubles" wouldn't even visit this site or call retro computing as a hobby. 😅

It was too much trouble when I had to go to the BIOS to disable L1 cache on my 1996 system back then (but I did that when needed). That is why I wanted a 486 system for DOS games when computers were too fast and ISA was gone. Turbo button made things much more easy plus I could use ICD and ICE from command line. They did not work on later systems. But now with SETMUL, Pentium makes much more sense, especially with MMX CPU.

And trying to get the speed down to exact speed (25 MHz 386 most likely for Wing Commander) has a problem when you slow down CPU but use very fast VGA card. It is just not the same. Much more fun trying to build a period correct 386 then for that special purpose (I'd like to build a 386 next since I finally found a nice late 1991 motherboard with 386DX-33. But I have many nice 1993 VGA cards and no where to put those).

But it sure would be nice to be able to disable motherboard cache easily. But how necessary that is? Are there many other games that have so specific performance requirements as Wing Commander? Most games just work or don't. Is there a game or other test I should try here? Any synthetic benchmark that could be used to find sweet spot for Wing Commander?

aitotat wrote on 2023-04-12, 07:44:

It was too much trouble when I had to go to the BIOS to disable L1 cache on my 1996 system back then (but I did that when needed).

Maybe if you're frequently jumping between games which need different speeds. But if you're doing a dedicated playthrough of Wing Commander, you just need to disable L1+L2 in the BIOS before starting the game. Several days later, after completing the game, you return the settings to normal. I honestly don't see this as being a hassle.

Are there many other games that have so specific performance requirements as Wing Commander? Most games just work or don't. Is there a game or other test I should try here?

I don't personally play many games from that era other than Wing Commander, but I have heard that Test Drive III is also difficult to control at speeds higher than a 386. Some of the other games listed on the Vogons Wiki page as "needs a 386" may have similar issues.

Any synthetic benchmark that could be used to find sweet spot for Wing Commander?

I think 3DBench 1.0c is fine. You get playable speeds in Wing Commander with scores between 10 and 15.

Joseph_Joestar wrote on 2023-04-12, 08:13:

Several days later, after completing the game, you return the settings to normal. I honestly don't see this as being a hassle.

That makes sense.

I think 3DBench 1.0c is fine. You get playable speeds in Wing Commander with scores between 10 and 15.

I had to try something with the plus-CPU. I set bus to 50 MHz to try to get lowest possible speed with both CPU caches disabled and multiplier set to 2 and no write combining. I got 16.0 from 3DBench 1.0c. That would be the slowest score so far. Motherboard cache was enabled so even slower is possible if needed. 50 MHz bus speed limits the maximum clock to 300 MHz but it was surprisingly fast. That would mean faster than MMX 233 in every benchmark and overclocked MMX 262 was faster only in Quake. That CPU L2 cache sure helps with low bus speed.

But I have a lot more testing to do. I think I'm going to re-test most interesting CPUs with motherboard cache disabled to see if 10-15 is reachable from 3DBench 1.0c. I suppose I should also test those MMX test register features for even better comparison. Maybe not all combinations but enough to get slow scores from 3DBench.

Some K6-III+ benchmarks added. Results with 75 MHz bus are missing. I'll do those next week but not with that many multipliers. It can be seen that it is enough to test 6x, 2x and maybe some from between. I didn't test every possible combination with 66 MHz bus since it does not make sense to test those half multipliers. They don't really make any difference.

Ability to change multipliers seems to be not necessary with 66 MHz bus. With 100 MHz bus things are likely different and you can easily avoid the "200 MHz bug" (500 MHz bug on a K6-2?) by changing multipliers. But that is not needed here. You need to disable CPU cache(s) for perfect OPL3 compatibility (and that about 486DX33 speed is also great for Sierra games with speed bugs). And to get down to the 386-25 level for Wing Commander you need to disable motherboard cache as well. And those are available with every CPU. So the ability to change multipliers is not needed at all (maybe there are some rare cases where it helps. Let me know what those are).

But the plus-CPU is the fastest you can install on this motherboard and it runs cooler than regular K6-2 or K6-3 so that makes the plus-CPU best because you can have everything with it, slow speed and fastest speed. And if you don't care about speeds above MMX-233, then you can use 50 MHz bus with the plus-CPU. That way you can get 16.0 from 3DBench (multiplier 2) and speed above MMX-233 (multiplier 6). Then you don't need to disable motherboard cache and that is the only case when ability to change multiplier is useful.

Well, almost the only case. For some reason the system does not post when multiplier is set to 2 (=6). It does post with K6-2 but not with the plus-CPU. So I set it to multiplier 1,5(=3,5) and then used SETMUL to set the CPU to 6x. Maybe a BIOS issue?

I forgot to test one more possible scenario. What if you just keep motherboard cache disabled and never enable it at all? I suppose this is better than setting bus speed to 50 MHz. There is very little performance drop since the K6-3+ has 256k internal L2 cache. So disabling motherboard cache harms only a little. You will also get 3DBench score of 13,7 just by disabling L1 cache (and L3 of course remains disabled). Just perfect for Wing Commander! (6x66=400MHz here). Or alternatively you could disable L2 as well to get 3DBench score of 11,9.

But for some reason Cycles fails even though otherwise it is a very slow speed with L2 enabled. Cycles can be fixed by disabling L2 or by setting multiplier to 2.0x. Anything above and it fails.

But that is about it. You get very slow and very fast speed but nothing useful in between. So no way to get that other useful close to 486 DX-33 speed. Motherboard cache would have to be enabled for that one.

I'd really like to get this system ready. These CPU tests are taking forever (and I happened to find one more graphics card for the next phase of tests...) and I'm not going to use K6 here anyway.

CPU tests are now complete. Again I tested more than I thought I would but I really wanted to know how Cyrix 6x86 pr200+ performs with L2 disabled, for example. I also just had to test some of those test register combinations with MMX 166 and MMX 200. It was really needed because I got some interesting results with Pentium Classic 200. And of course I did many tests with K6-3+ with 75 MHz bus.

Cyrix 6x86 pr200+ got some nice results with L1 and L2 caches disabled. 3DBench score of 16,9 makes it perhaps a bit too fast for Wing commander but all the OPL3 tests succeeded. So the performance is somewhere between the "safe" 486 speed and "Wing Commander" 386 speed.

Pentium classic does not have those CCD and DCD features that the MMX has. I tested and they do nothing. That does not matter since they simple disable half of the L1 cache so speed will be above the "safe" 486. I don't know what would need that kind of speed and not work with full Pentium speed.

BPD and VPD seem to do very little when L1 is disabled, at least they affect speed very very little. In fact, I started testing those with L1 disabled and thought they do not work on a Pentium Classic even though according to datasheet Pentium Classic should support them. And it does. Even though they affect speed very very little when L1 is disabled, disabling Branch Prediction does one great thing. It makes Indy "386" to work. So to slow Pentium Classic 200 to 486 levels, I recommend to use L1D BPD parameters for setmul instead of just L1D. It makes sense that disabling branch prediction helps here because it will slow down software timing loops.

What about when L2 is disabled? Pentium Classic 200 is great here also, 3DBench scores are 11,3...11,7 depending on what combination of L1D, BPD and VPD is used. Makes very little difference but for some reason disabling all three seems to be the fastest. This is very strange but the difference is very very small. But Pentium Classic 200 is excellent CPU to get down to 386 and 486 levels of performance. Just what I wanted! And it gets even better. Even though this CPU had no overclocking potential at all, it seems to run stable with 3,3V (it should be used with 3,5V). So I'm very glad that this last and best of the 1996 Pentiums is such a great CPU!

But what about MMX? MMX-166 can do all the great things Classic 200 does. There are minor speed differences (3DBench score of 12,4 when it is at slowest speed, so more than with Pentium Classic 200) but nothing important. With full speed it is a bit slower than Classic 200 but if you want to run Windows 9x games as well, then the MMX instructions can be useful and makes the MMX CPU faster. And MMX-166 runs cooler so overall it is a better CPU. But it was released in 1997 and best of Pentium Classics sure sounds much more nicer than slowest of the Pentium MMX CPUs.

What about MMX-200 then? Perfectly usable. You can get it down to 13,3 (3DBench) so it should be fine. But you'll need to disable L2 cache to get Indy "386" working. That is not necessary with with Pentium Classic 200 or with MMX-166.

One thing to note. Take a look at the Phils results here and compare with mine. Clock-for-clock my fastest are faster than his and my slowest scores are slower (that is a great thing, of course). So other components matter here as well so I recommend to experiment when building a system.

But nothing beats K6-3+. With 66 MHz bus you can get 3DBench score down to 11,2 and there will be many combinations to get great "safe" 486 speed. But no matter what the speed, Indy "486" always requires the "a" command line parameter when using K6. With 75 MHz bus you get 3DBench score down to 11,7. Again a great result. And at their fastest speeds K6-3+ is in completely different performance group than Pentium Classic or MMX. So if you have a motherboard that supports the plus CPUs, then the K6plus CPUs are simply the best.

But I'll choose the Pentium Classic 200. It is a great 1996 CPU. But wait, what about the Pentium 150 with 2x75? It seems 75 MHz bus is too fast to get Indy "386" to work with 486 speed. You need to go down to 386 speed to get it working and that requires disabling L2.

About L2 (L3 with K6plus CPUs). Disabling it is necessary to get down to 386 speeds. I'll investigate it further. I'm sure some software solution can be made to disable it without need to go to the BIOS to do it. But that can wait until the system is build and otherwise ready. Next there will be many graphics cards to test (with the Pentium Classic 200).

Edit: One more thing. K6-3+ 450 MHz (6x75) was free from the "200 MHz" Turbo Pascal bug (Runtime Error 200) . So it did not affect any of the tested CPUs.

aitotat wrote on 2023-04-21, 14:46:

One thing to note. Take a look at the Phils results here and compare with mine. Clock-for-clock my fastest are faster than his and my slowest scores are slower (that is a great thing, of course). So other components matter here as well so I recommend to experiment when building a system.

Yeah, component selection and BIOS settings can slightly affect overall speeds. Some motherboard chipsets are a bit faster than others, and the same goes for graphics cards.

I think it might be possible to get some more slowdown by loosening memory timings and disabling video BIOS caching or shadowing, but I never bothered with that on my MMX 166. Once I got to the desired 386 and 486 speeds, I didn't really feel like I needed anything else, at least for the games that I usually play.

BTW, great work on all the CPU tests. That's a very comprehensive analysis! Those numbers should be pretty helpful to people who are planning to build a DOS gaming PC, and want to consider various options.

I haven't started testing video cards yet. It is easy to do benchmarks but I'd like somehow compare picture quality on a CRT and test how well they work with LCD. I'm not going to use LCD monitors with my DOS computers but it might be useful information. Image quality with CRT will be quite hard to compare and I'll have to think how to do that.

But before that it is time to install motherboard back in to case. I used PSU from this case to do all the CPU tests. Modern ATX power supply might be safer to use but I can't use one that has 120 mm fan unless I modify the case (like I did on my DX4 system). I should have modified the case when I grinded the rust off. So at least for now, I'll just use the old PSU since it worked fine. I replaced the old fan with better and much more quiet Nexus. It is last of the 80mm Nexus fans I bought years ago. They were recommended fans then.

So I opened the PSU, cleaned it and removed the old fan. It was soldered so this was a good time to do a simple mod.

I soldered two resistors to be used as a fan header. Those are 56 ohm resistors. With them, the fan gets 8V. I was hoping to get 10V, like with same kind of fan on my 286. 10V is very quiet but pushes perhaps more air than the original fan, at least on that other PSU. Maybe I should have used 47 ohm resistors or 33 ohm. Something around there would be the best. I don't remember what I used on the other PSU. I didn't modify it like this. I think I used fan controller there and adjusted it so fan gets 10V. With 8V the fan seems to push out enough air (I hope) and I'm going to use another fan anyway so I'll leave it like this for now.

I installed biggest socket 7 heatsink I got. That is meant for passive cooling but I don't know for what CPU exactly. But in practice passive cooler needs some air flow or it gets very hot. That is the reason for the 80mm case fan. It was very easy to install it in this case because there are plastic clips to mount a 80mm fan. And motherboard has a fan header in a very good location. That fan is just temporary solution. It is just some random 80mm fan with leds. I don't want leds, just a normal silent good quality fan. Can anyone recommend any reasonable priced but good fans? Noctua should be good but it is expensive. I'll have to experiment how well this solution cools the CPU. It is not in optimal position as can be seen. But I think this will work fine. I do not want one of those old CPU coolers with 40mm very loud fan (I used original CPU cooler when testing CPUs, horrible noise, but it cooled even the K6-2 non plus just fine). But those can be made much better with modern silent 50mm fan. There is one thing that worries me: case front bezel has very small air holes so the fan won't work as well as it should.

Just remember that the current draw of the fan affects the voltage drop across the resistors, hence also the available voltage to the fan, so you'd better measure the voltage with the fan you intend to use 😀

That fan with leds was unusably loud! And I thought the old original socket 7 cooler from this system sounded bad... I ordered two new Noctua fans (NF-R8 redux 1200). Those should be very good. Maybe I put the other one inside the PSU and use the Nexus on some other project. I hope I don't have to do any voltage adjustments with these. We'll see.

I tried to test CPU temperature with and without the loud fan. With fan (and open case), the CPU (or heat sink actually) stays cool. Heat sink can be touched easily and it is slightly warm. Without the case fan (again open case) the heat sink gets really hot. You can touch it for a few seconds but not more. So whatever the temperature is, it is not critical. But with case closed and no case fan, it will get higher than that. So any case fan, even a very quiet fan, should help here a lot. Just what I expected. The air coming out of PSU is not too warm but it should get warmer when case is closed.

I decided to choose a disk drive and CD-ROM-drive. I just retrobrighted 8x-speed Creative (CD821E) and Sony (CDU311) drives. Then I also have a 4x-speed NEC CDR-C251. What a magnificent drive! There are some youtube videos about that NEC to see that changer in action.

For a 3,5" disk drive I chose Mitsumi. First, the case front bezel is whiter than any drive I have (I had to retrobright power button. It was very yellow) so no drive is a perfect match for it. But the Mitsumi is closest. Originally this system had Panasonic disk drive but that did not work and it was less white than the Mitsumi (and I think I did retrobright the Panasonic last summer). Mitsumi is only cleaned. It did not require retrobrighting at all. Again a Mitsumi drive turned to be more quiet than most drives but I had to clean, lubricate and fix that. Someone has dropped the Mitsumi and one corner was bent so that disk did not fit inside. I was able to fix that. I'm happy with the Mitsumi. It now works, has closest color with case bezel and is also quietest.

I found that NEC CD-changer drive from flea market before Christmas. It was expensive, 20 euros, but 4x-speed drives are so rare and that was special in other ways as well. That changer mechanism is very cool and there is no yellowing at all. But the drive is more gray than the case. Same problem with Creative drive. The Sony is more beige than gray. So no drive is perfect here. I could also remove 8x-speed Mitsumi from my DX4 system but that one is gray just like the Creative drive.

But that NEC would look awesome in DX4 system if I paint the buttons orange. The NEC is easily quietest of the CD-ROM-drives (but the changer mechanism is noisy when changing discs. But that does not matter) as expected from 4x-speed drive. Sony is second best. That unit is manufactured in 1997 but that model was available in 1996 so I can use it here. Creative is much more noisier when changing tracks. The Sony feels that it is better quality drive than Creative. Did Creative manufacture that themselves? It has Sony and Oak chips in it and Creative was written on the PCB. I know Creative used Panasonic and Funai drives earlier (dual speed drives with Panasonic interface).

Since the case bezel does not have anything extra and unnecessary (just like I prefer on this system), the Sony is the best option here. I'm going to put the NEC on my DX4 system even though multi-disc changer is more useful with Pentium era games. Hard to decide but perhaps this is the best. Without the Creative logo and 8Xspeed text, the Creative drive would have looked the best. But maybe I do some better comparison with the Mitsumi and Sony when I remove the Mitsumi from DX4 system. 8x speed is the fastest drive I accept. Faster are too noisy. I'd like these 8x drives would be a bit more quiet but these I can tolerate. Perhaps 6x drive is ideal but I've never had one of those.