| View previous topic :: View next topic |
| Author |
Message |
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
 Posted: Sun Apr 28, 2019 6:28 am Post subject: SRAM cache chips (discussion moved from sale thread) |
 |
|
| stamasd wrote: | | Fake cache is usually soldered to the MB so if these came out of sockets they should be OK. No, whent he's talking about "Chinese relabeled" he means chips that have recently appeared on ebay, aliexpress etc that are probably NOS real SRAM chips but relabeled to higher specs than they really are (e.g. 10ns when they are in fact 15ns, 128k when they are 32k and so on) |
Yeah, exactly. I've got a bunch of such 128Kx8 which were "rated" 10ns by the Chinese. They can do 3-2-2-2 timings at 66MHz or 2-1-1-1 at 40MHz which is what you can expect of 15ns async SRAM. Real 10ns should be capable of 3-1-1-1 at 66MHz. |
|
| Back to top |
|
 |
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Sun Apr 28, 2019 6:38 am Post subject: |
|
|
| For authentic 128kx8 chips, I suspect 12 ns is the limit. Have you seen real 10 ns variants? |
|
| Back to top |
|
|
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
| Posted: Mon Apr 29, 2019 12:27 pm Post subject: |
|
|
| feipoa wrote: | | For authentic 128kx8 chips, I suspect 12 ns is the limit. Have you seen real 10 ns variants? |
They exist according to the data sheets. It's hard to say if they have been actually produced in a large quantity. I know that narrow DIP-32 is more likely to have propagation issues at 100MHz than SOJ-32. It's also true that by the time these 10ns async SRAMs hit the market, industrial demand switched largely to pipelined burst SRAMs. I think the best bet now is to design a SOJ-32 to DIP-32 adaptor rather than chase the ghost. There are certain mechanical difficulties which have to be addressed, but it's still possible to design one.
I remember our talk the last year when I was going to upgrade cache chips on my LuckyStar LS-486E rev. D. It took me almost 3 months to get all U.S. export docs right and have the customs release the chips. I have replaced all those poor Tayeh 10uF/16V capacitors with tantalum 100uF/10V and 33uF/25V in the meantime, and some 100nF MLCCs with 5uF. All 9 32Kx8 15ns ISSI chips were replaced. 8 10ns Cypress CY7C199D-10VXI + 1 8ns EliteMT LP61256GS-8 for the tags. It runs at 66MHz with 3-1-1-1 read timings. 2x 32Mb Kingston KTM8X32L-70ET EDO. The SiS 496 north bridge can only do write through cache at 66MHz, besides I have to do it anyway in order to get 64Mb cacheable with only 256Kb L2 cache. Cyrix 5x86-120 @ 133MHz S1R3. It works generally at 3.45V until I enable all optimisations known to work. Then it crashes in less than a minute. Looks like it needs 3.6V, but my power regulator has been already reconfigured for either 3.45V or 3.7V, so I use the latter. A custom cooler with a dual ball bearing fan which runs off 5V instead of 12V to reduce noise.
Runs great. Read speed from the L2 cache is exactly 50% of the L1 cache. |
|
| Back to top |
|
|
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Mon Apr 29, 2019 3:02 pm Post subject: |
|
|
Very interesting - this is the first such example of an S1R3 at 133 MHz. You should post in the World's Fastest 486 thread. https://www.vogons.org/viewtopic.php?f=46&t=28748
6.66 in CPUMark99 is impressive. I see that I only scored 6.35 using a Biostar MB-8433 at 133 MHz, but this is S0R5, so no branch prediction. What do you get in DOS Quake demo version, timedemo1? In my thread, I see 19.8 fps.
Your revision of the LS-486E seems to be better than its predicessors because I see on your board that it is using double-banked SRAM rather than single-banked. This allows you to run faster timings. With the original 15 ns ISSI cache, did you need to limit the timings to 3-2-2 and noticed that you could run it at 3-1-1 with the 8/10 ns cache?
Which graphics card are you using? I bet you could run some early 3D games with ease. As your system should be faster than my IBM 5x86c-133, it would be interesting to see your results with a Voodoo2 and compare it with the results I charted here, https://www.vogons.org/viewtopic.php?t=61639 . In that thread, the Cyrix 5x86-120 S1R3 beat the Cyrix 5x86-133 S0R5, but the IBM 5x86c-133/2x S0R5 beat the Cyrix 5x86-120 S1R3. So it would be really interesting to see how a Cyrix 5x86-133/2x S1R3 compares. It should add another 5% to close the gap between the POD100 and the IBM 5x86-133/2x.
I have not tried my Cyrix 5x86-120 S1R3 at 133 Mhz. I felt the chip was too rare to try at 133 MHz. The only chips I found which would run reliably at 2x66 MHz were the IBM 5x86c in QFP form. I had one of these IBM 5x86c chips running at 150 MHz in DOS pretty well, until one day it just died.
Are you running 70 ns EDO RAM? Any plans to upgrade to 60 or 50 ns? Were you able to quantify how much EDO RAM improves performance on this board?
How is your heatsink attached? Thermal epoxy?
Why was it so difficult to get the SRAM shipped to you? Which country are you in?
If you want to tune in the voltage some more, you can replace the voltage set resistor with a trim pot like I did. However, after extensive testing on my Biostar board, none of my 2x66 chips were long-term stable at 3.6 V. I would run GLQuake in loop for an hour to fine tune this. For most people, 3.6 V may be good enough for most situations. It is interesting how turning on more features requires more voltage. |
|
| Back to top |
|
|
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
| Posted: Mon Apr 29, 2019 4:57 pm Post subject: |
|
|
| feipoa wrote: |
Your revision of the LS-486E seems to be better than its predicessors because I see on your board that it is using double-banked SRAM rather than single-banked. This allows you to run faster timings. With the original 15 ns ISSI cache, did you need to limit the timings to 3-2-2 and noticed that you could run it at 3-1-1 with the 8/10 ns cache? |
LS-486E rev. C1 / C2: single banked cache, 4 DIP-32 for data + 1 DIP-28 for tags. No EDO support.
LS-486E rev. D: dual banked cache, 8 SOJ-28 for data + 1 SOJ-28 for tags. Supports EDO.
LS-486E rev. F: single banked cache, 4 DIP-32 for data + 1 DIP-28 for tags. Supports EDO. Poor cache layout as they moved the north bridge too far.
The L2 cache didn't work at 66MHz with all 15ns chips. It worked with 3-2-2-2 timings after I replaced the tag chip with 8ns. Started to do 3-1-1-1 after the data chips were replaced with 10ns.
| feipoa wrote: |
Which graphics card are you using? I bet you could run some early 3D games with ease. |
Indeed. 3Dfx Voodoo Banshee (Diamond Monster Fusion 16Mb PCI). It's an overkill for Cyrix 5x86.
| feipoa wrote: |
Are you running 70 ns EDO RAM? Any plans to upgrade to 60 or 50 ns? Were you able to quantify how much EDO RAM improves performance on this board?
How is your heatsink attached? Thermal epoxy?
|
The label says 70ns, but the chips are 60ns. I don't think 50ns will make a difference, but I'll give it a try.
Cache Write Cycle: 3
Cache Burst Read Cycle: 1
L2 Cache/DRAM Cycle WS: 3
DRAM RAS to CAS Delay: 3
DRAM Write Cycle: 0
DRAM Write CAS Pulse: 1
DRAM CAS Precharge Time: 1
DRAM RAS to MA Delay: 1
DRAM Speed: Fastest
DRAM Slow Refresh: Enable
Performance is the same with one or two SIMMs. There are only two slots and SiS 496 doesn't support 64Mb SIMMs, so it cannot be any better.
There is a regular zinc oxide thermal compound under the heat sink which is attached with a help of a steel clip taken from a 486 compatible cooler. The aluminium heat sink itself is Pentium compatible. Had to file it a little.
| feipoa wrote: |
Why was it so difficult to get the SRAM shipped to you? Which country are you in? |
These Cypress cache chips were exported from the U.S. and had a 3A991.b.2.b Export Control Classification Number (ECCN) assigned for some arcane reason. Not the usual EAR99. Therefore all int'l customers had to submit a bunch of government regulated paperwork. End Use & Export Compliance Certification with two annex'es, End User Statement. Not for military use, not for re-exporting to Sudan/Iran/Syria/North Korea, etc. All forms printed out, filled, signed, scanned and sent back. Don't want to do this again.
| feipoa wrote: |
If you want to tune in the voltage some more, you can replace the voltage set resistor with a trim pot like I did. However, after extensive testing on my Biostar board, none of my 2x66 chips were long-term stable at 3.6 V. I would run GLQuake in loop for an hour to fine tune this. For most people, 3.6 V may be good enough for most situations. It is interesting how turning on more features requires more voltage. |
Yes, it may help. Prime95 and 3D benchmarks don't stress Cyrix 5x86 a lot because the FPU isn't the fastest part of it. C compilers do much better. They stress ALUs, branch prediction, load/store units, caches and memory. It takes a few hours to recompile a generic kernel of an older NetBSD, FreeBSD or Linux distribution. If nothing gets broken in the process, consider it rock stable. Additional features enabled make it consume more power and be more sensitive to input voltage. |
|
| Back to top |
|
|
stamasd
Joined: 05 Jun 2014
Posts: 1311
Location: Connecticut
|
| Posted: Mon Apr 29, 2019 5:31 pm Post subject: |
|
|
Heh the old "kernel compile" stress test. I did a version of it just the other day on a modern system, not just the kernel but the whole distribution (Gentoo) on a Ivy Bridge laptop. Turns out that the heatsink compound I applied is much better than the factory one.  |
|
| Back to top |
|
|
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Mon Apr 29, 2019 6:04 pm Post subject: |
|
|
Are you able to run the DOS Quake demo benchmark?
With a BIOS update, the 4DPS, which is a very similar board running the SiS 496 is able to use 64 MB SIMMs. I ran 128 MB on my board. Have you tried the newer Tomato 4DPS BIOSes? |
|
| Back to top |
|
|
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
| Posted: Mon Apr 29, 2019 6:31 pm Post subject: |
|
|
I ran both Quake and GLQuake on it. FPS was playable and about the same. I'll do timedemo's and report the results.
Zida 4DPS uses Winbond W83787IF super I/O. My board has UMC UM8669F.
_________________
My Active Sales |
|
| Back to top |
|
|
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Mon Apr 29, 2019 9:19 pm Post subject: |
|
|
| Did you try the Funworld 4DPS BIOS? It has worked on most of the SiS 496 boards that I've tested it on, though sometimes you need to edit the PCI routing table. |
|
| Back to top |
|
|
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
| Posted: Tue Apr 30, 2019 1:48 am Post subject: |
|
|
| Can you give me a BIOS file to try? |
|
| Back to top |
|
|
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Tue Apr 30, 2019 2:43 am Post subject: |
|
|
I've attached the two latest BIOSes for the 4DPS. 4DPS2013 is the Funworld BIOS, v1.72, but I have removed that annoying Funworld logo and set some CMOS defaults. 4DPS400A is supposedly the latest 4DPS BIOS, but I have had issues getting it running properly. Others have reported that it works on their system, and other had my issue with it.
It is the 4DPS2013 BIOS that I've used on other SiS 496 boards, but I've had to match the PCI routing table on the 4DPS2013 BIOS to those that are on your existing board's BIOS. The 4DPS has PS/2 support, so you can remove that feature in the BIOS if you wish by using Modbin. |
|
| Back to top |
|
|
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
| Posted: Tue Apr 30, 2019 5:07 pm Post subject: |
|
|
Quake 1.06 timedemo1 = 17.1 fps
GLQuake 1.09 timedemo1 = 20.6 fps
Banshee or Voodoo3, doesn't matter because it's FPU limited. I don't think 3Dfx year 2000 drivers are optimised for 486 either. Maybe Voodoo2 can do better, have to give it a try.
It also seems to work fine with LINBRST enabled.
Last edited by ph4nt0m on Fri May 08, 2020 9:23 am; edited 2 times in total |
|
| Back to top |
|
|
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Tue Apr 30, 2019 5:44 pm Post subject: |
|
|
Yeah, I think it is BWRT that doesn't work on S1R3 CPUs.
Is Speedsys misreporting the speed of your system?
Your Quake timedemo1 score seems lower than it should be for 133 MHz on a Cx5x86. Did you run -nosound -nomouse ? |
|
| Back to top |
|
|
ph4nt0m
Joined: 01 Jan 2018
Posts: 812
Location: Europe
|
| Posted: Thu May 02, 2019 5:05 pm Post subject: |
|
|
Speedsys misreports speeds sometimes. I've attached another screen shot with a correct display.
The 4DPS2013 BIOS seems to work. The 4DPS400A BIOS cannot boot DOS off a floppy. Overall 4DPS2013 is better than the original LuckyStar BIOS because it's Y2K compliant and inits PnP devices properly. May need minor tuning though. What's your issue with 4DPS400A? |
|
| Back to top |
|
|
feipoa
Joined: 08 Mar 2011
Posts: 553
Location: Canada
|
| Posted: Thu May 02, 2019 5:20 pm Post subject: |
|
|
I don't recall the issue with the 400A BIOS; it has been more than 5 years since I tested it.
Was there any performance boost with the 2013 BIOS? Can you send me the LuckyStar BIOS you were using? Sometimes if the PCI routing table is not correct, the PCI devices won't work when you load drivers in Windows 9x. |
|
| Back to top |
|
|
|
![[ Hidden ]](https://cdn.cpu-world.com/forum/templates/CPUWorld/images/lang_english/icon_email.gif "[ Hidden ]"){kind=icon}