JonF wrote on 2023-03-21, 04:43:

Thought this belonged here- […]
Show full quote

Thought this belonged here-

I got a block of dry ice in my latest shipment of frozen coffee. Decided to set up my 200mhz Lucky Star setup and see if 240mhz was possible.

Using the dry ice, once I got the processor to ~-40C (haha) and the system turned on and almost started booting. I even tried turning down the memory timing, turning off the cache, removing extra ram sticks. Normally, this chip needs 10C or less to run reliably at 200mhz.

I imagine with a truly unicorn AM5x86 240mhz benchmarking could be possible!

Awesome!

This one.
It has 4 wheels like a Mac Pro for going pretty fast.

My attempt to build HI-END 486 VLB (PL lng): http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1691482911/0#0

mods:
* mobo repair
* mobo factory hw bug
* mobo vcore mod
* mobo cpu mod
* mobo clockgen mod

parts:
* mobo: SOYO SY-25K VLB SiS 85C471/407 (Green PC ISA-VLB 486 Single Chip)
* cache: L2 256K 15ns
* cpu: Am5x86-P75 AMD-X5-133ADW @160
* ram: 2x16MB 60ns SIMM 72-pin FPM (before 16MB 60ns SIMM 30-pin FPM, chips Nanya NT511740C0J-60)
* gfx: Cirrus Logic CL-GD5429 VLB 2MB (1MB 60ns soldered + 1MB 50ns in sockets)

partial log (full log in previous link):
* optimal and stable settings for current platform with cirrus vlb & 2x16MB simm 72-pin 60ns {
* 50x3/150: bios ram timmings slower/0/1 cache 2/2, doom 2134/1635*35=45.7, quake low 13.1, pcplayer low 17.3, speedsys cpu 56.27
* 40x4/160: bios ram timmings fastest/0/1 cache 1/2, doom 2134/1516*35=49.3, quake low 14.7, pcplayer low 19.6, speedsys cpu 59.90
* 33x4/133: bios ram timmings fastest/0/1 cache 1/2, doom 2134/1942*35=38.5, quake low 11.9, pcplayer low 15.8, speedsys cpu 49.84

smola wrote on 2024-03-05, 19:38:

My attempt to build HI-END 486 VLB (PL lng): http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1691482911/0#0 […]
Show full quote

My attempt to build HI-END 486 VLB (PL lng): http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1691482911/0#0

mods:
* mobo repair
* mobo factory hw bug
* mobo vcore mod
* mobo cpu mod
* mobo clockgen mod

parts:
* mobo: SOYO SY-25K VLB SiS 85C471/407 (Green PC ISA-VLB 486 Single Chip)
* cache: L2 256K 15ns
* cpu: Am5x86-P75 AMD-X5-133ADW @160
* ram: 2x16MB 60ns SIMM 72-pin FPM (before 16MB 60ns SIMM 30-pin FPM, chips Nanya NT511740C0J-60)
* gfx: Cirrus Logic CL-GD5429 VLB 2MB (1MB 60ns soldered + 1MB 50ns in sockets)

partial log (full log in previous link):
* optimal and stable settings for current platform with cirrus vlb & 2x16MB simm 72-pin 60ns {
* 50x3/150: bios ram timmings slower/0/1 cache 2/2, doom 2134/1635*35=45.7, quake low 13.1, pcplayer low 17.3, speedsys cpu 56.27
* 40x4/160: bios ram timmings fastest/0/1 cache 1/2, doom 2134/1516*35=49.3, quake low 14.7, pcplayer low 19.6, speedsys cpu 59.90
* 33x4/133: bios ram timmings fastest/0/1 cache 1/2, doom 2134/1942*35=38.5, quake low 11.9, pcplayer low 15.8, speedsys cpu 49.84

It seems waitstates are holding back the memory throughput. With SIS471 at 40MHz 47MB/s memory throughput (values of speedsys) are possible.
The DOS graphic values seem to be ok for a Cirrus VLB. Of course I would not call it high-end.

froller wrote on 2024-03-04, 09:18:

I was experimenting with one of my AMD-X5-133ADZ's and figured out that in Doom, Quake and some other benchmarks 150MHz (50 x 3) […]
Show full quote

I was experimenting with one of my AMD-X5-133ADZ's and figured out that in Doom, Quake and some other benchmarks 150MHz (50 x 3) config works approximately 4-5% faster than 160MHz (40 x 4). Even with halved PCI bus clock.
Some of installed PCI cards can't stand 50MHz bus and crash while booting in Win'98. I'm gonna figure out which one exactly later this week.
Is there any explanation for this?

Test setup: AMD-X5-133ADZ, LS-486e Rev.D, 256kB L2 15ns, 2x32MB EDO 60ns, CL-GD5446. 0WS, cache timings turned down to "Faster" (2-2-2-2).

Somehow it is explainable I would say. If you use the exact same timing values for cache and RAM @50 and 40MHz, then cache and RAM is exactly 25% faster @ 50MHz.
But the question is, why you cannot use better values at 40MHz, like cache read 2-1-1-1 and fastest for RAM.

CoffeeOne wrote on 2024-03-05, 20:34:

Somehow it is explainable I would say. If you use the exact same timing values for cache and RAM @50 and 40MHz, then cache and RAM is exactly 25% faster @ 50MHz.
But the question is, why you cannot use better values at 40MHz, like cache read 2-1-1-1 and fastest for RAM.

I can't use 2-1-1-1 timings either @40 and 50MHz.
@40MHz it boots into DOS and fails only on some memory intensive tasks like Quake or booting into Win'98 GUI.
@50MHz it doesn't even get to reading boot sector.

System with 2-2-2-2 timings looks stable at both 40 and 50MHz with barely measurable performance drop comparing to 2-1-1-1.
I mean performance difference between 4x40 and 3x50 is an order of magnitude greater.

The difference in timings and slower CPU clock negates all benefit of higher FSB. You need to push FSB to 60-66Mhz to see performance boost.

The Serpent Rider wrote on 2024-03-06, 14:49:

The difference in timings and slower CPU clock negates all benefit of higher FSB. You need to push FSB to 60-66Mhz to see performance boost.

Really?
2-1-1-1 is 0.2% faster than 2-2-2-2
4x40MHz is 19.4% faster than 4x33MHz
3x50MHz is 31.3% faster than 4x33MHz and 9.8% faster than 4x40MHz

froller wrote on 2024-03-06, 13:14:

I mean performance difference between 4x40 and 3x50 is an order of magnitude greater.

I can't see what you are referring to, as order of magnitude implies 1000%, a whole power of ten above, not the maximum top to bottom 25% lower 33% faster I see on the results you posted.

BitWrangler wrote on 2024-03-06, 15:04:

froller wrote on 2024-03-06, 13:14:

I mean performance difference between 4x40 and 3x50 is an order of magnitude greater.

I can't see what you are referring to, as order of magnitude implies 1000%, a whole power of ten above, not the maximum top to bottom 25% lower 33% faster I see on the results you posted.

I'm referring the performance difference of switching timings from 2-2-2-2 to 2-1-1-1, not the absolute values

CoffeeOne wrote on 2024-03-05, 20:26:

It seems waitstates are holding back the memory throughput. With SIS471 at 40MHz 47MB/s memory throughput (values of speedsys) are possible.

Well, maybe with edo ram, however this mobo doesn't support it, only fpm, I tested about 40 different edo modules and they always freeze @ post.

The DOS graphic values seem to be ok for a Cirrus VLB. Of course I would not call it high-end.

I'm sure there are better gfx than cirrus, however I've own only one vlb card 😉 Actually I built this platform to test repaired vlb cards, so it's not important/comparable to motherboards with pci and edo mem. This one is dedicated to vlb only. And this is main reason why I called it hi-end 486 vlb not hi-end 486 😉

Anyway, I got few conclusions during building/testing this platform, maybe it'll be helpful for someone:
* maximizing timmings (actually reducing them) has a huge impact on vlb performance
* adding cache significantly speeds up the operation of the platform and vlb
* enabling irq for the vlb card improves performance slightly
* for vlb the WB/WriteBack cache mode is slightly faster than WT/WriteThru
* memory bandwidth in speedsys is irrelevant if the comparison concerns cache on/off
* adding additional memory in the card for 3D and game testers in DOS does not bring anything, but in Windows it may be important
* increasing the clock on the AT bus has no impact on vlb performance
* in DOS, Trident 9440 VLB is slightly faster than Cirrus Logic CL-GD5429 VLB
* the choice of vlb slot is important, only 1 of 3 works on this mobo, the one closest to the middle of the board

web:
orig post: http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1617221455/201#201

3x50 is certainly very interesting configuration.
Many of us checked it before. To my best knowledge it is slightly faster in some tests than 4x40 and slightly slower in others.
It will be good if we talk specifics - please share the metrics. This will make the conversation more substantial.

pshipkov wrote on 2024-03-06, 20:00:

3x50 is certainly very interesting configuration.
Many of us checked it before. To my best knowledge it is slightly faster in some tests than 4x40 and slightly slower in others.
It will be good if we talk specifics - please share the metrics. This will make the conversation more substantial.

Re: The World's Fastest 486

Wouldn't it be these?

thanks.
somehow didn't see that, not sure how, so i kept wondering what percentages are discussed here.

1froller's 3x50 2-2-2-2 numbers: 21.6 54.76 15.4
2froller's 4x40 2-1-1-1 numbers: 20.3 49.94 ?
3     mine 4x40 2-1-1-1 numbers: 24.1 56.6 17.6

Froller's numbers are surprisingly low. Wonder what is the reason for that ?
I recognize that the point Froller is trying to make is about the performance difference between the two modes. However, there are two factors to consider:
- the lower than expected numbers bring the question if the outcome is skewed in some way
- some of the DOS interactive graphics tests benefit from higher FSB, while other, more CPU intensive tests benefit from higher CPU clock - something to keep in mind.

LuckyStar revision D can run with tightest BIOS timings at 4x40 and 3x50 with the default 15ns L2 cache chips and decently ok RAM modules without even trying. It just works.
I can say this with fairly high level of confidence after at least 5 of these boards passed through here over the years. Same applies to LuckyStar C/C2 btw.

I remember at least one other discussion with @maxtherabbit with his LSC boardhere
There was a follow-up on that topic somewhere else in that thread.

Can the amount or type of RAM affect the results? I use 2x32MB 60ns EDO.
@4x40MHz I could max out all timings except one no matter which exactly.

froller wrote on 2024-03-07, 12:40:

Can the amount or type of RAM affect the results? I use 2x32MB 60ns EDO.
@4x40MHz I could max out all timings except one no matter which exactly.
VirtualBox_MS-DOS_07_03_2024_15_41_38.png

64MB RAM is too much for L2 WB in tag 7+1 mode, because the cachable area is 32MB. So you can try to switch to L2 write through and see if values improve.

And your values with 150MHz and 160MHz are really strange. Check the other posting from the polish guy, he has higher values at 160MHz than at 150MHz, which is "normal", at least confirms my findings that 160MHz at fastest settings always beats 150MHz (because 2-1-1-1, 2 cache is not possible at 50MHz.

smola wrote on 2024-03-06, 18:01:

Well, maybe with edo ram, however this mobo doesn't support it, only fpm, I tested about 40 different edo modules and they alway […]
Show full quote

CoffeeOne wrote on 2024-03-05, 20:26:

It seems waitstates are holding back the memory throughput. With SIS471 at 40MHz 47MB/s memory throughput (values of speedsys) are possible.

Well, maybe with edo ram, however this mobo doesn't support it, only fpm, I tested about 40 different edo modules and they always freeze @ post.

The DOS graphic values seem to be ok for a Cirrus VLB. Of course I would not call it high-end.

I'm sure there are better gfx than cirrus, however I've own only one vlb card 😉 Actually I built this platform to test repaired vlb cards, so it's not important/comparable to motherboards with pci and edo mem. This one is dedicated to vlb only. And this is main reason why I called it hi-end 486 vlb not hi-end 486 😉

Anyway, I got few conclusions during building/testing this platform, maybe it'll be helpful for someone:
* maximizing timmings (actually reducing them) has a huge impact on vlb performance
* adding cache significantly speeds up the operation of the platform and vlb
* enabling irq for the vlb card improves performance slightly
* for vlb the WB/WriteBack cache mode is slightly faster than WT/WriteThru
* memory bandwidth in speedsys is irrelevant if the comparison concerns cache on/off
* adding additional memory in the card for 3D and game testers in DOS does not bring anything, but in Windows it may be important
* increasing the clock on the AT bus has no impact on vlb performance
* in DOS, Trident 9440 VLB is slightly faster than Cirrus Logic CL-GD5429 VLB
* the choice of vlb slot is important, only 1 of 3 works on this mobo, the one closest to the middle of the board

web:
orig post: http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1617221455/201#201

No, I was comparing a board with SIS471 with another board with the SIS471 chip. So no EDO and no PCI. FPM and Vesa Local Bus only.
See here for really high end values on an Asus VL/I-486SV2GX4:
Re: 3 (+3 more) retro battle stations

These were my values with the Asus VL/I-486SV2GX4
Re: 3 (+3 more) retro battle stations

I now have slightly higher values, because the Ark 1000 is now running at 80MHz. Ark 1000 is high end for DOS throughput, but not the best card in Windows.
To reach "high end" values, also 1MB L2 cache makes sense, and a Bios that allows to run the write back mode with 7+1 tag configuration.

froller wrote on 2024-03-06, 15:04:

2-1-1-1 is 0.2% faster than 2-2-2-2
4x40MHz is 19.4% faster than 4x33MHz
3x50MHz is 31.3% faster than 4x33MHz and 9.8% faster than 4x40MHz

I'd say to not compare synth mem bandwidth/throughput numbers but real fps in dosbench testers. Bandwidth is just calculated and not measured like throughput. Real life are games/testers and fps 😉
However, I'd suggest you to do recap your mobo. If you wanna go over the spec, then you should have proper equipment. After 30 years hw can be exhausted as a horse after a race 😉
There is more info about it, pl lng ofkoz 😉 http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1642611929/238#238
Most important lines here regarding recap:
* taken 4x470u/6.3V from the donor board, newer Compaq mobo on S775, very good parameters verified with a meter, nominal capacity ESR 0.3 ohm, old ESR a few or a dozen ohms and capacity 5-50% smaller
* recap completed, the board works, ripple ~50mV, previously ~200mV, peaks the same up to 300mV 35KHz, from the power supply, you could add induction to the main 5V line, but maybe one day - it's good and even better than it was 😉

Good luck with your search for perfection 😉

CoffeeOne wrote on 2024-03-07, 21:30:

No, I was comparing a board with SIS471 with another board with the SIS471 chip. So no EDO and no PCI. FPM and Vesa Local Bus on […]
Show full quote

smola wrote on 2024-03-06, 18:01:

Well, maybe with edo ram, however this mobo doesn't support it, only fpm, I tested about 40 different edo modules and they alway […]
Show full quote

CoffeeOne wrote on 2024-03-05, 20:26:

It seems waitstates are holding back the memory throughput. With SIS471 at 40MHz 47MB/s memory throughput (values of speedsys) are possible.

Well, maybe with edo ram, however this mobo doesn't support it, only fpm, I tested about 40 different edo modules and they always freeze @ post.

The DOS graphic values seem to be ok for a Cirrus VLB. Of course I would not call it high-end.

I'm sure there are better gfx than cirrus, however I've own only one vlb card 😉 Actually I built this platform to test repaired vlb cards, so it's not important/comparable to motherboards with pci and edo mem. This one is dedicated to vlb only. And this is main reason why I called it hi-end 486 vlb not hi-end 486 😉

Anyway, I got few conclusions during building/testing this platform, maybe it'll be helpful for someone:
* maximizing timmings (actually reducing them) has a huge impact on vlb performance
* adding cache significantly speeds up the operation of the platform and vlb
* enabling irq for the vlb card improves performance slightly
* for vlb the WB/WriteBack cache mode is slightly faster than WT/WriteThru
* memory bandwidth in speedsys is irrelevant if the comparison concerns cache on/off
* adding additional memory in the card for 3D and game testers in DOS does not bring anything, but in Windows it may be important
* increasing the clock on the AT bus has no impact on vlb performance
* in DOS, Trident 9440 VLB is slightly faster than Cirrus Logic CL-GD5429 VLB
* the choice of vlb slot is important, only 1 of 3 works on this mobo, the one closest to the middle of the board

web:
orig post: http://3dfx.pl/cgi-bin/yabb2/YaBB.pl?num=1617221455/201#201

No, I was comparing a board with SIS471 with another board with the SIS471 chip. So no EDO and no PCI. FPM and Vesa Local Bus only.
See here for really high end values on an Asus VL/I-486SV2GX4:
Re: 3 (+3 more) retro battle stations

These were my values with the Asus VL/I-486SV2GX4
Re: 3 (+3 more) retro battle stations

I now have slightly higher values, because the Ark 1000 is now running at 80MHz. Ark 1000 is high end for DOS throughput, but not the best card in Windows.
To reach "high end" values, also 1MB L2 cache makes sense, and a Bios that allows to run the write back mode with 7+1 tag configuration.

Oh, thank you for info, I'll definitely check it out. There is not many sources with such knowledge.

froller wrote on 2024-03-07, 12:40:

Can the amount or type of RAM affect the results? I use 2x32MB 60ns EDO.
@4x40MHz I could max out all timings except one no matter which exactly.
VirtualBox_MS-DOS_07_03_2024_15_41_38.png

I spent a lot of time with the LSD board at 60 MHz FSB. I think you should be able to get to 2-1-2/Fastest with the right EDO memory module. Best to just use one stick.

Keep in mind that you may need to up the voltage to the CPU. Sometimes this helps with tight SRAM/DRAM timings. I think the LSD only outputs around 3.40V on the 3.3V jumper. I usually take that up to 3.6 V for 160 MHz and 4 V for 180 MHz. However, I think pshipkov is able to run 2-1-2/fastest at 3.4 V at 150/160 MHz.