VOGONS


Reply 20 of 37, by pentiumspeed

User metadata
Rank l33t
Rank
l33t

Did not see 20ns sram on any thing faster than 33mhz cpus.

That formula does not apply, since the requirement is due to timing window which varies with access timings to sram and CPU. Typically you need margin to have reliablity as well. This is what CPU and chipset, (if there's a cache controller as well), examining the data sheet is required to calculate to arrive at reliable timings and arrive at safe margins without "over taking" the sram's minimum access time.

For example SRAM is rated at 20ns, that's the *least* safe access timing window and that's no margin!, you want margin of about 20% which could be 15ns SRAM is needed. Go any less and you will be forced to insert wait cycles (over-clocking the MHz) or change SRAM chips for 15ns instead (allows adjust to more tighter timings but usually chipset and motherboard design limitations often put crimp on the overclocking success.

Cheers,

Great Northern aka Canada.

Reply 21 of 37, by jakethompson1

User metadata
Rank Oldbie
Rank
Oldbie
pentiumspeed wrote on 2020-07-05, 00:45:
Did not see 20ns sram on any thing faster than 33mhz cpus. […]
Show full quote

Did not see 20ns sram on any thing faster than 33mhz cpus.

That formula does not apply, since the requirement is due to timing window which varies with access timings to sram and CPU. Typically you need margin to have reliablity as well. This is what CPU and chipset, (if there's a cache controller as well), examining the data sheet is required to calculate to arrive at reliable timings and arrive at safe margins without "over taking" the sram's minimum access time.

For example SRAM is rated at 20ns, that's the *least* safe access timing window and that's no margin!, you want margin of about 20% which could be 15ns SRAM is needed. Go any less and you will be forced to insert wait cycles (over-clocking the MHz) or change SRAM chips for 15ns instead (allows adjust to more tighter timings but usually chipset and motherboard design limitations often put crimp on the overclocking success.

Cheers,

Yeah, that's what I mean. If the CPU clock cycle is 20ns (which is the purpose of the formula), and the memory is 20ns, that allows no time whatsoever for the CPU to generate the signals, for them to travel to the memory and back, and for the cpu to recognize them. So 1000/MHz, and then another step faster in speed to allow a margin of safety

Reply 22 of 37, by pshipkov

User metadata
Rank Oldbie
Rank
Oldbie

I have FX-3000 working very reliably at 45MHz with FPU and at 50MHz without FPU.
At 50MHz i had to lower some of the BIOS settings, but cannot remember anymore which ones exactly.
At 45MHz all BIOS settings are to the max.
The mobo is not pretentious about SRAM chips. It takes ISSI, UMC, Winbond brands at 25, 20, 15, 12 ms with ease.
Intel, AMD, DLCs work well too.

So, maybe you got a dud ? 🙁

retro bits and bytes

Reply 23 of 37, by Anonymous Coward

User metadata
Rank l33t
Rank
l33t

Not sure why, but 486s normally need 15ns cache for reliable operation at 40MHz, but on 386s I almost always see 40MHz systems equipped with 20ns cache and they can usually take the fastest timings too. But, I think at 50MHz you're really going to want something better.

Anyone know what die process was used on 40MHz AMD CPUs? I would think they should handle 50MHz no problem if they're not really old.
50MHz could be a problem for the FPU. I recommend a black top Cyrix Fasmath based on 387+ core.

"Will the highways on the internets become more few?" -Gee Dubya
V'Ger XT|Upgraded AT|Ultimate 386|Super VL/EISA 486|SMP VL/EISA Pentium

Reply 24 of 37, by pshipkov

User metadata
Rank Oldbie
Rank
Oldbie
Anonymous Coward wrote on 2020-07-05, 01:46:

50MHz could be a problem for the FPU. I recommend a black top Cyrix Fasmath based on 387+ core.

+1

I think early 386 was something like 1500 nm.
Later models went down to 1000 nm.

So far all 40MHz rated 386 CPUs i tried handle 50MHz just fine.
The main bottleneck is predominantly the motherboard, SRAM, VGA, RAM - in that order.

retro bits and bytes

Reply 25 of 37, by NScaleTransitModels

User metadata
Rank Newbie
Rank
Newbie
pshipkov wrote on 2020-07-05, 01:26:
I have FX-3000 working very reliably at 45MHz with FPU and at 50MHz without FPU. At 50MHz i had to lower some of the BIOS settin […]
Show full quote

I have FX-3000 working very reliably at 45MHz with FPU and at 50MHz without FPU.
At 50MHz i had to lower some of the BIOS settings, but cannot remember anymore which ones exactly.
At 45MHz all BIOS settings are to the max.
The mobo is not pretentious about SRAM chips. It takes ISSI, UMC, Winbond brands at 25, 20, 15, 12 ms with ease.
Intel, AMD, DLCs work well too.

So, maybe you got a dud ? 🙁

I highly doubt it, all 3 FX-3000's I tried to OC work well at 40mhz with cache. One of them even came new with the original box and manual. However, one of them (not the new one) gives strange errors with both my AMD 386's, yet is fine with an Intel chip running at 40 O__o 🤔

Anyway, I just ordered some 15ns cache chips, so we'll see how those hold up. Hopefully then we can at least rule out the cache as the problem.

Black top Fasmath = one of these? https://www.ebay.com/itm/NOS-Cyrix-CX-83D87-4 … pc/382774566990 Or the plastic-surface-mount-on-PCB version?

pshipkov wrote on 2020-07-05, 03:35:

So far all 40MHz rated 386 CPUs i tried handle 50MHz just fine.
The main bottleneck is predominantly the motherboard, SRAM, VGA, RAM - in that order.

Did you try both the ceramic PGA version and the plastic PQFP version? I wasn't sure which one to pick: it seems like the ceramic one would be easier to cool, but I read that the plastic version tolerates higher temperatures. Also just a speculation: maybe the plastic version is a later revision that would be more tolerant of an OC? Both 40mhz chips I have at the moment are ceramics w/ identical white lettering, and both slow down by the same amount when OC'd to 50.

Builds:

  • ECS FX-3000; 386DX-40@50; ET4000AX, ISA 1mb
  • Acer VI9; 486DLC-40; Mach32, VLB 2mb
  • Chicony CH-471A; CX486s-40; Mach32, VLB 2mb
  • Gateway 2000 P5-60; Pentium-60@66; S3 928, PCI 3mb
  • DTK PKM-0033S; AM5x86-133@160

Reply 26 of 37, by Anonymous Coward

User metadata
Rank l33t
Rank
l33t

The later ceramic Fasmaths typically had the 387+ core, which was reduced in complexity to allow it to scale better to higher frequencies. The older 83D87 core is said to have topped out at 33MHz. The 387+ models are far more common because they were dirt cheap.
I don't know much about the PQFP version. The only one I have is soldered to an upgrade, so I haven't been able to try overclocking it.

"Will the highways on the internets become more few?" -Gee Dubya
V'Ger XT|Upgraded AT|Ultimate 386|Super VL/EISA 486|SMP VL/EISA Pentium

Reply 27 of 37, by amadeus777999

User metadata
Rank Oldbie
Rank
Oldbie
pentiumspeed wrote on 2020-07-05, 00:45:
Did not see 20ns sram on any thing faster than 33mhz cpus. […]
Show full quote

Did not see 20ns sram on any thing faster than 33mhz cpus.

That formula does not apply, since the requirement is due to timing window which varies with access timings to sram and CPU. Typically you need margin to have reliablity as well. This is what CPU and chipset, (if there's a cache controller as well), examining the data sheet is required to calculate to arrive at reliable timings and arrive at safe margins without "over taking" the sram's minimum access time.

For example SRAM is rated at 20ns, that's the *least* safe access timing window and that's no margin!, you want margin of about 20% which could be 15ns SRAM is needed. Go any less and you will be forced to insert wait cycles (over-clocking the MHz) or change SRAM chips for 15ns instead (allows adjust to more tighter timings but usually chipset and motherboard design limitations often put crimp on the overclocking success.

Cheers,

The original ratings seem to only apply to very good batches.

@NScaleTransitModels - at (close to) 50mhz the 386 may even be a "good" cpu for the original Doom.

Reply 28 of 37, by pshipkov

User metadata
Rank Oldbie
Rank
Oldbie

@nscaletransitmodels

I posted some notes about fx-3000 here:
Re: 3 (+3 more) retro battle stations

So far all 40mhz rated PGA and PGFP 386 cups I tried tick reliably at 50mhz. In fact they cope with 55mhz as well, but at that frequency the 386 class hardware is at the limit. Bios settings have to be relaxed which actually leads to lower perf than optimized system running at 50 and even 45mhz.
I was unable to see any termal stability diff between the two packages.

The older cyrix fpu core is slightly faster, clock to clock, than the newer one, but does not overclock past 45mhz.
Pick your poison.

retro bits and bytes

Reply 29 of 37, by Anonymous Coward

User metadata
Rank l33t
Rank
l33t

I believe am386 was built using .8 micron fabrication technology, so 50MHz really shouldn't be much of a big deal. It was the same process used on the Intel DX2 chips, and 386s don't have internal cache to worry about.

"Will the highways on the internets become more few?" -Gee Dubya
V'Ger XT|Upgraded AT|Ultimate 386|Super VL/EISA 486|SMP VL/EISA Pentium

Reply 30 of 37, by NScaleTransitModels

User metadata
Rank Newbie
Rank
Newbie
pshipkov wrote on 2020-07-05, 07:57:
@nscaletransitmodels […]
Show full quote

@nscaletransitmodels

I posted some notes about fx-3000 here:
Re: 3 (+3 more) retro battle stations

So far all 40mhz rated PGA and PGFP 386 cups I tried tick reliably at 50mhz. In fact they cope with 55mhz as well, but at that frequency the 386 class hardware is at the limit. Bios settings have to be relaxed which actually leads to lower perf than optimized system running at 50 and even 45mhz.
I was unable to see any termal stability diff between the two packages.

The older cyrix fpu core is slightly faster, clock to clock, than the newer one, but does not overclock past 45mhz.
Pick your poison.

Thanks for the link! Great info on the FX-3000, and I also enjoyed reading about how other boards also overclock (or don't)! And that DTK board is something else: almost 11 FPS in DOOM?!

I see... I wonder if my two chips were both from the same (very) unlucky batch, or if the previous owner was an extreme overclocker who went too far🤣 (they were both from a personal collection). I ordered a PQFP part yesterday from a major reseller just to be sure it's not the CPU.

I'll probably pass on the FPU for now, I'd rather get my setup running stable and with the fastest timings possible first. Then I'd go with the newer core... if it doesn't work, I have my modern PC for CAD and rendering 🤣

Builds:

  • ECS FX-3000; 386DX-40@50; ET4000AX, ISA 1mb
  • Acer VI9; 486DLC-40; Mach32, VLB 2mb
  • Chicony CH-471A; CX486s-40; Mach32, VLB 2mb
  • Gateway 2000 P5-60; Pentium-60@66; S3 928, PCI 3mb
  • DTK PKM-0033S; AM5x86-133@160

Reply 31 of 37, by mkarcher

User metadata
Rank l33t
Rank
l33t
Anonymous Coward wrote on 2020-07-05, 01:46:

Not sure why, but 486s normally need 15ns cache for reliable operation at 40MHz, but on 386s I almost always see 40MHz systems equipped with 20ns cache and they can usually take the fastest timings too.

The 486 supports a burst read on its front-side bus. The fastest burst supported by the 486 processor is called the 2-1-1-1 burst, because it takes two clock cycles for the first 32-bit value, and 1 clock cycle for the remaining 3 32-bit values. The 386 on the other hand does no bursts on its frontside bus, so you have two clock cycles for each 32-bit value.

Reply 33 of 37, by NScaleTransitModels

User metadata
Rank Newbie
Rank
Newbie

I received the rest of the parts today and no dice. 15ns cache, no errors in cachechk, benchmarks sometimes start and sometimes won't. Windows 95 is a no-go; I get different symptoms for each of my 3 CPUs (2 ceramic and 1 PQFP):
1. Windows protection error on VCACHE
2. Windows protection error on VXDLDR
3. Indefinite loading at the Win95 splash screen; the bar spins but no disk activity

Also tried various cache, memory, and bus timings. The 89.2 and 92.9405 mhz crystals won't work at all.

Time to gut the case for a socket 7 build and throw the 80mhz crystal in my DX-33 system 🤣 🤣

Builds:

  • ECS FX-3000; 386DX-40@50; ET4000AX, ISA 1mb
  • Acer VI9; 486DLC-40; Mach32, VLB 2mb
  • Chicony CH-471A; CX486s-40; Mach32, VLB 2mb
  • Gateway 2000 P5-60; Pentium-60@66; S3 928, PCI 3mb
  • DTK PKM-0033S; AM5x86-133@160

Reply 34 of 37, by NScaleTransitModels

User metadata
Rank Newbie
Rank
Newbie

Revisiting after 2 years (excuse me for the necropost), I succeeded in pushing the 386 to 50mhz, with good timings no less. The speed difference is very noticeable in games and in Windows. Ultimate Doom became ALMOST playable without reducing the screen size. Benchmark results really improved as well; using a Diamond Speedstar 24 ET4000AX and the timings shown in below photos:
3DBENCH 1.0c = 20.9
PC Player = 4.7
Doom = 8179 realtics (9.13 fps)
Speedsys CPU= 11.59

Hardware-wise, no new magic. I'm using the same setup, albeit different serial #s: the FX-3000 motherboard and CPUs I originally tested are long gone (sold). But after coming across another FX-3000 and more DX-40 parts, I decided to give it another shot.

Since I'm not using the exact same motherboard, 15ns SRAMs, or SIMMs, they can't be ruled out. But I'm almost certain CPU's themselves were the limiting factor in previous attempts.

I tried 4 CPU's this time, 2 CPGA's and 2 PQFP's. Both the CPGA's pass at 50mhz, while both the PQFP's fail. Last time I had 2 CPGA's and 1 PQFP (all different), all of which failed.

So takeaways for anyone losing sleep over a 386DX-50 build:
-FX-3000 motherboards are your best bet, as they officially support 50mhz bus speed and not TOO rare
-When playing the CPU lottery, ceramic parts may have a higher success rate than PQFP parts (and are easier to cool)
-15ns cache and 60ns memory are hard requirements

Now for the fun part— photos!

Screenshot_20220913-013853_Gallery.jpg
Filename
Screenshot_20220913-013853_Gallery.jpg
File size
580.75 KiB
Views
658 views
File license
CC-BY-4.0

The glorious 386DX-50 setup. Note that I used a 12ns tag SRAM to be safe, but I think 15ns will do. Heatsink + fan are necessary however; fortunately 486 coolers can be permanently mounted using a custom adapter bracket.

20220912_143229.jpg
Filename
20220912_143229.jpg
File size
1.5 MiB
Views
658 views
File license
CC-BY-4.0

Tightest timings, almost: DRAM needs 1 WS. Maybe 50ns SIMMs can do 0 WS, if 30-pin versions even exist? ISA bus overclocks to CPUCLK/4 = 12.5mhz, really boosting video speed.

20220912_134213.jpg
Filename
20220912_134213.jpg
File size
1.69 MiB
Views
658 views
File license
CC-BY-4.0

Windows 98 (yes, 98) runs surprisingly decent, though it does take some tricks to install on a 386. The hardware also passes a final stability test: installing NT 3.51.

20220911_122017.jpg
Filename
20220911_122017.jpg
File size
685.83 KiB
Views
658 views
File license
CC-BY-4.0

Overclockers on the right, non-overclockers on the left

20220912_144648.jpg
Filename
20220912_144648.jpg
File size
916.14 KiB
Views
658 views
File license
CC-BY-4.0

A very clean AT case to top it off

Builds:

  • ECS FX-3000; 386DX-40@50; ET4000AX, ISA 1mb
  • Acer VI9; 486DLC-40; Mach32, VLB 2mb
  • Chicony CH-471A; CX486s-40; Mach32, VLB 2mb
  • Gateway 2000 P5-60; Pentium-60@66; S3 928, PCI 3mb
  • DTK PKM-0033S; AM5x86-133@160

Reply 36 of 37, by Anonymous Coward

User metadata
Rank l33t
Rank
l33t

Somebody on VOGONS made their own 40ns SIMMs. Maybe they still have some for sale.

"Will the highways on the internets become more few?" -Gee Dubya
V'Ger XT|Upgraded AT|Ultimate 386|Super VL/EISA 486|SMP VL/EISA Pentium

Reply 37 of 37, by NScaleTransitModels

User metadata
Rank Newbie
Rank
Newbie
kixs wrote on 2022-09-13, 09:50:

Now put the Ti486SXL2-50 with default 1X multi. This should be faaaast 🤣

I made the mistake of selling my SXL2-66 earlier this year. Now I'll have to look out for another one 🙃

Anonymous Coward wrote on 2022-09-13, 09:56:

Somebody on VOGONS made their own 40ns SIMMs. Maybe they still have some for sale.

He still has 50ns SIMMs for sale, but from his video, looks like 50ns won't cut it at 50mhz. I'll have to ask if he's still making 45ns or 40ns modules.
Custom made, 50ns, FPM, Simm 30, memory modules with parity.

Builds:

  • ECS FX-3000; 386DX-40@50; ET4000AX, ISA 1mb
  • Acer VI9; 486DLC-40; Mach32, VLB 2mb
  • Chicony CH-471A; CX486s-40; Mach32, VLB 2mb
  • Gateway 2000 P5-60; Pentium-60@66; S3 928, PCI 3mb
  • DTK PKM-0033S; AM5x86-133@160