Makes sense.

==========

VLB EIDE/SCSI controllers are interesting subject to me.
Keep updating documented data here with new findings.

Recently that was Promise DC440 - a combined IDE (PDC20330A, PDC20430) / SCSI-2 (NCR 53CF94-2) VLB controller from year 1993.
Apparently a high-end product for its time. Priced at $190.

Full package in its original plain white box.

Plugged it into the fast Asus VLI based PC (running at 40MHz base frequency - P24T@100 or Am5x86@160) that was used for testing all other VLB EIDE/SCSI controllers in the link above.

Couldn't find any description of the PDC20430 chip.
I suspect it is an EIDE controller, but the word "enhanced" is not used anywhere in the manual (the booklet).
There are no drivers for it on the floppy discs, or that particular diskette is missing. Also, none of the existing/known DOS and Windows drivers support it. Finally - no jumpers for speed settings (something common on EIDE controllers).
At the same time there are PDC20230 chips that implement EIDE interfaces, so it does not make sense for this one to be plain old IDE.
But without a driver IDE mode is in order. Because of that - performance is very low.

The SCSI-2 interface does bus mastering, so L1 cache must be set to Write Through mode, otherwise boot does not complete - the controllers hangs.
On-board jumpers provide access to BIOS memory location and IRQs.
JP5 can be set to position 1-2 for "base frequency smaller than 33MHz". Supposedly improves performance, but didn't notice any difference between positions 1-2 and 2-3.
SCSI setup menu can be invoked with F2 during POST. Interface is similar to the ones know from the later Promise UltraIDE PCI controllers.
Sync mode is faster than Async.
Used 10mb/s mode.

Both IDE and SCSI-2 performance is similar between CF cards and mechanical HDDs.
Indication that bottleneck is the controller itself.
Performance metrics provided in the link above.
Basically 1.6Mb/s IDE and 3-5Mb/s SCSI-2, reported by SpeedSys and Coretest.
Much slower than most other VLB EIDE controllers (missing driver).
Similar SCSI speed to Adaptec 2840VL.

I think i made all the right moves when testing the device, but still something feels off.
1. The booklet talks about up to 10Mb/s theoretical speed for the SCSI interface, but i don't get even half of that, while the system is capable of reaching 12-18Mb/s with other EIDE controllers.
2. PDC20430 is most likely an EIDE interface ? Anyone ? Driver ?

Neat. Looks like it has some SRAM onboard. Is the SRAM for the SCSI or IDE? I wonder if the SRAM is slowing it down when using a CF or SD card? Pretty neat that they went full board: SCSI, IDE, LPT, FLOPPY, and 2x SERIAL. The only VLB SCSI card I've been able to use L1:WB on is the Adaptec AHA-2842A. Do IDE VLB cards typically suffer the fate of SCSI cards with respect to DMA and L1:WB ?

I was unable to test the EIDE interface because cannot find a driver.
No idea how the cache chips are used.
SCSI perf is a bit lower than Adaptec 2840VL, so unlikely the cache is doing something extraordinary there.
Otherwise the whole package is nice - the entire I/O shebang on a single card.

There is another card based on UMC + Adaptec (some early versions of the) chips that also offers the entire package.
May give it a try too.
I think it will be similar to this D 440 performance wise.

So, how did you make the Adaptec 2840 play nice with L1 cache in WB mode ?

I'm not sure about the 8240, but on the Adaptec AHA-2842A, there is this magic jumper J5 which lets L1:WB work properly, at least it did for me with an Am5x86 at 160 MHz on an Asus VL/I-486SV2GX4.

I messed up the number of the SCSI controller - i meant 2840.
Remember trying that jumper but experienced reliability issues on the same board with same CPU.
That Adaptec SCSI controller was slower than good VLB EIDE ones from that time, so i moved on and didn't look at the problem since then.
Are you still using it that way ?
Would be interested to know what perf it delivers on your system and if it is stable-stable, especially with CF cards.

Yes, I still have it in that system and never had any issues whatsoever with the Am5x86. The system is buried pretty deep in the closet right now, but I can pull out a dozen system to get to it if there is a specific failure mode you wanted me to verify. I have it setup with a Seagate U160 or U320 drive. I think 4 GB w95c, 4 GB NT4, and 65 GB TEMP. I haven't run CF cards on it. I assume you mean through an ACARD bridge adapter? I have a few generic Winbond/UMC/whatever VLB IDE controllers. Are these faster than the 2842A? I should have a Promise EIDE 2300 Plus sitting at AVICC, but not in hand. Probably will ship my AVICC stuff in January.

I sort of recall the 2842A not running properly with a Cyrix and L1:WB, but it was such a long time ago now that I wouldn't trust this recollection.

Yes, using Acard scsi to ide adapter with CF cards. Strongly prefer CF cards over mechanical HDDs. CF are just too practical to avoid.

I think the issue with L1 WB cache and Adaptec AHA-2840# is specific to this motherboard (or maybe the SIS 471 chipset).
At least with several other boards/chipsets - all is good.
But I don't have a complete picture.
Just noticed that SIS496/7 and UM8881 based boards that have VLB slots do just fine with L1 Wb and SCSI bus mastering.

But yes, the 2840 is actually much slower compared to many EIDE adapters.
Stare for a moment at the charts in the link I provided in my previous-previous post.
Notice Promise 2300Plus, Holtek 5660B, Appiah ADI2, DTCw278E next to AHA-2840.

Sorry, I thought you were talking about L1:WB with the 2842. With respect to L2:WB and the AHA-2842, I haven't done extensive testing, but with what testing I have done, there doesn't appear to be an issue. If you have a specific test you know failed, I'll run it. One of the issues with the L2:WB BIOS hack is that selecting L2:WT in the BIOS results in a non-booting system, so it is hard to go back and forth with testing WB/WT unless I swap EEPROMs.

I had seen your charts earlier and didn't think the 2842 performed all that bad. My results were a bit better than yours: 9336 KB/s buffered read in Speedsys. In Windows 95, Adaptec's SCSIBench32 returns 9546 KB/s with a 256 KB transfer size and sequential I/O.

I was planning on testing various VLB IDE controllers when I did my 386+VLB build, which will contain an IBM BL3 upgrade unit. Unfortunately, that system only worked with the BL3 at 3x75, so the bus will be slow at 25 MHz only. I was going to use the Promise 2300 in there with a replica PowerGraph VLB card. However, after looking at your chart, it appears as if the Holtek HT6560B takes the lead over all others when a mechanical HDD is used.

Looking through my drawer, I recalled that I have two VLB IDE controllers pulled from some cheap/free PC chip systems. So cheap that the controller cards only contain a single IDE port. One is based on the Winbond W83759F, the other on the Holtek HT6560A. In looking at the datasheets, it is not clear to me what the difference is between the *A and *B variants of the HT6560. Maybe the *B supports faster PIO modes, CD-ROMs, pipelined pre-fetch data reads, 4-layer FIFO. All of those were omitted from the *A datasheet. Any chance you tested an *A card?

Would you be willing to share your drivers for the W83759F and HT6560?

Any chance the Promise 2300 has BIOS support for >8.4 GB HDDs?

My last few posts were messy.
First i talked about Adaptex AHA-8240 instead of 2840, then kept talking about L2 WB issue with the SCSI adapters instead of L1 WB cache issue.
Got too checked-out for Christmas.
Fixed the occurrences.

---

Good chance you got higher AHA-2840 results because you managed to get it working with L1 cache in WB. But also - some controllers show better score with HDDs, especially late/fast HDD models.

---

25Mhz base frequency will hurt disk i/o for sure.
Looking at some of the metrics i shared in previous posts with 386 VLB motherboards, the results are just not great at all - mostly around ~3-5Mb/s. Still pretty ok for 386 class hardware.
The best is Alaric Cougar that comes with on-board Adaptec VLB EIDE controller - ~6Mb/s on average, reported by SpeedSys and Coretest, but that is with BL3 processor.
I doubt it can be outdone with third party controllers, but that's just a presumption.

---

I remember seeing the Holtek "A" models for sale but passed on them since the "B" variant here.
If you can give it a spin and share the numbers that will clarify if it is the same silicon with trimmed-down PCB, or an early design with different perf characteristics.

While the B model is very fast, it exhibits a write access issue with CF cards.
From time to time it will fail to write something to the storage device and needs to be retried (via prompt that the OS/tools show at the Write Access failure).
That's why i cannot recommend it for CF cards really.
Solid with mechanical HDDs.

Since i use CF cards only - Promise EIDE2300Plus is the best/fastest thing for that medium.
Also, second best after Holtek with HDDs.
It is the better place overall in my opinion.

---
Don't have consistent data about partition sizes.
I think most controllers see storage devices above 8Gb, but not sure if they can allocate partitions past that threshold.

---

Attached you the drivers.
Holtek HT6560B v2.41 (DEVICEHIGH=C:\PATH\TO\HTIDE2.SYS /D0=2,2)
Promise EIDE2300Plus v3.31 (DEVICEHIGH=C:\PATH\TO\EIDE2300.SYS /T /M0:8)
Winbond W83759F v1.26 (put C:\PATH\TO\WBIDE.EXE in AUTOEXEC.BAT)

---

The whole VLB EIDE/SCSI thing needs better investigation than my brief tests with SpeedSys/Coretest and few CF/HDDs only. I feel that the controllers need to be checked more thoroughly with different classes of mechanical HDDs.
Also, SpeedSys and Coretest are anything but reliable to start with, so more tests need to be included.
Finally, i test only in DOS. Picture may be different in Windows 3.1 / 95 based on how good the drivers are.

OK, so it is the L1:WB issue that you are referencing. If that is the case, I've no issue whatsoever. I've installed W95c, NT4, various software, tested games, benchmarks, etc. No problemo armano. Do you recall what BIOS version is on your 2842? I'd have to double check, but I think mine is v2.0.

Thanks for the driver pack. The Holtek doc mentions specifically the *B variant of the HT6560, so I may need to hunt down the *A version. Or the driver may be able to differentiate, who knows. I'm pretty far out from getting to test this. The BL3 will come after my VLSI 286 build, which is waiting for some 5V SXLC-40, SLC2-25/50, and 87SL-33. I'll be using CF-ACARD-SCSI on this system. Also waiting on S3 801 and true colour ET4000AX, sitting at AVICC still.

I don't own any Alaris product. However, it is pretty rare to find a combination of 386 motherboard and a BL3 upgrade adapter that actually work with L1 enabled. I've only come across one such pair with a Buffalo adapter. While those IO Data modules are nice looking, they are intended for a PC-98 system and there's something preventing them from letting the L1 work without a cache flush every I/O. That flush kills performance to the point that it lags behind even an SXL2-50. The Buffalo unit, for whatever reason, didn't work at 2x33 (IIRC) on my Daewoo board, but works at 3x25. I'll have to reinvestigate this. Maybe swap the BL3 CPU from my IODATA and put it on the Buffalo. I know one of the BL3 chips on an IODATA module I have can operate up to 110 MHz.

The only catch with my VLB+386 build plan is that the Buffalo upgrade board blocks access to both VLB slots. The solution to this dilemma is somewhat embarrassing, but I'm sure you can guess at my temp fix for it. Let's say that shorter VLB boards are preferred for this build. PGA-132 rotator boards would be preferred.

I may consider swapping out the 2842A on the Asus VL/I-486SV2GX4 board for the Promise 2300 if it has BIOS support for >8.4 GB HDDs. But on the other hand, there is a certain elegance inherent to using SCSI, esp. w/working L1:WB.

feipoa wrote on 2021-12-27, 03:06:

OK, so it is the L1:WB issue that you are referencing. If that is the case, I've no issue whatsoever. I've installed W95c, NT4, various software, tested games, benchmarks, etc. No problemo armano. Do you recall what BIOS version is on your 2842? I'd have to double check, but I think mine is v2.0.

Good to know.
There is some conflicting information about this L1 WB issue with SCSI bus mastering on Asus VLI.
As i mentioned early - this is just not an issue with other VLB boards. For example - yesterday i tried it on PcChips 919 and it was just fine.
But at least 1 more person (user PcEngineer) had the problem with VLI+AHA-2840.
Will look at this tomorrow to see if i missed something.

feipoa wrote on 2021-12-27, 03:06:

Thanks for the driver pack. The Holtek doc mentions specifically the *B variant of the HT6560, so I may need to hunt down the *A version. Or the driver may be able to differentiate, who knows. I'm pretty far out from getting to test this. The BL3 will come after my VLSI 286 build, which is waiting for some 5V SXLC-40, SLC2-25/50, and 87SL-33. I'll be using CF-ACARD-SCSI on this system. Also waiting on S3 801 and true colour ET4000AX, sitting at AVICC still.

Will be following your progress on the 286 hybrid.
You are taking a different direction than what i did before, so interest is high here.
Why not S3 928 ?

feipoa wrote on 2021-12-27, 03:06:

I don't own any Alaris product. However, it is pretty rare to find a combination of 386 motherboard and a BL3 upgrade adapter that actually work with L1 enabled. I've only come across one such pair with a Buffalo adapter. While those IO Data modules are nice looking, they are intended for a PC-98 system and there's something preventing them from letting the L1 work without a cache flush every I/O. That flush kills performance to the point that it lags behind even an SXL2-50. The Buffalo unit, for whatever reason, didn't work at 2x33 (IIRC) on my Daewoo board, but works at 3x25. I'll have to reinvestigate this. Maybe swap the BL3 CPU from my IODATA and put it on the Buffalo. I know one of the BL3 chips on an IODATA module I have can operate up to 110 MHz.

The only catch with my VLB+386 build plan is that the Buffalo upgrade board blocks access to both VLB slots. The solution to this dilemma is somewhat embarrassing, but I'm sure you can guess at my temp fix for it. Let's say that shorter VLB boards are preferred for this build. PGA-132 rotator boards would be preferred.

My success with CPU upgrade modules is very humble.
Drop-in Ti486SXL-2 for the 386 beast running at 50MHz natively is the high-mark. But there were SCSI reliability issues with CF cards, so had to switch to IDE and loose 1-2 Mb/s disk speed.
Then the 286 upgrade module with the same SXL2 CPU from above. The FPU never worked properly. Throw-in and the unreliable CF write access.
A BL3 386 upgrade module that never lit up on any 386 board. Previous owner mentioned that the only time it worked was some IBM machine.
And that's about it.

Even they work sometimes - there are always gotchas.

feipoa wrote on 2021-12-27, 03:06:

I may consider swapping out the 2842A on the Asus VL/I-486SV2GX4 board for the Promise 2300 if it has BIOS support for >8.4 GB HDDs. But on the other hand, there is a certain elegance inherent to using SCSI, esp. w/working L1:WB.

I would say - go for it.
It is the better place.
Simpler setup and logistics.
Faster.

Posted a while ago about what was supposed to be the final chapter of peak 286 overclocking to 27.5MHz and 0 wait states + other performance improvements.
Link here.
But ever since then i felt that the VLSI-200 system can handle 30MHz at 0 wait-states.
Achieved it today while looking at some other issues with Feipoa.
Snowballed there somehow.

This was hard.
From the small set of highly-curated 1Mb 30-pin SIMM memory modules only 2 worked reliably. They use Hynix/VT branded chips.
The rest were almost there but always led to some instabilities in different situations.
Then had to curate the VRAM of the Diamond SpeedSTAR 24 video card (picture in the first post of this thread).
Originally it used 60ns Siemens chips, but at 0-ws (JP1 on the card) some graphics were corrupted or not displayed at all.
Once swapped with VT branded ones (did you notice the pattern ?) things started working again.

The usual stats - numbers are great.
Notice the ~4Mb/s IDE readings.

And some performance tests.
The same Dunson Electronics motherboard, same everything, just higher frequency.
benchmark results

Captured video showing the PC rig with full gear, Wolf3D, Superscape, Landmark, NSSI, Windows 3.0.

Will need to find 2 more of these Hynix/VT modules to max-out the RAM.
Other than that - satisfied with the outcome. Bar moved further up.
Feel confident enough to claim "fastest 286 PC".
It will require CPU upgrade module with 486 DLC chip to improve on these metrics.
Did that before. The outcome was less than ideal. Left potential improvements for later.
Currently watching this thread - some good signs there for interesting results down the line.

---

EDIT (~1 year after this post was made)

The fastest ISA graphics adapters are based on Cirrus Logic GD-5434, but they require 386 or later class hardware, so the best option for 286 is ET4000 silicon.
Initially the PC was equipped with Diamond SpeedSTAR24, but Micro Labs's Ultimate VGA/TrueColor BIOS from year 1998 offers additional optimizations that make their card perform better than any other ET4000AX one. Unfortunately the BIOS does not work on other cards, so one non-period correct change had to be made to the rig.

Micro Labs Ultimate VGA/TrueColor

24 fps in Wolf3D look great for a 286 machine.

Wooohooo, well done. Yup fastest 286 I've ever heard of.

Shame you haven't got working 4MB SIMMs, then you could run the 386 instruction emulator and see how many frames per minute you get on Quake 🤣

wow!

Thanks.

Vanilla Quake 1 needs minimum of 6Mb RAM, but don't loose faith - Feipoa aims at 2x20MHz with SXL2 CPU and RAM extension card in addition to the 4Mb on-board.
There is a slim chance it works. That's were we get to see these frames per minute.
For now i will let him take the flack on that one. If he comes through i will follow.

In the meantime i splurged for promising looking SIMM modules. Hope to find the other two magical megabytes.

I am at the mercy of the postal service. COVID+WINTER+FLOODING+XMAS slowed everything down massively. And these miscellaneous IC sellers on seekic, alibaba, and the link aren't always easy to work with. However, I can confirm that Quake won't run on 4 MB.

Following Pshipkov’s and Ph4ntom’s lead I Installed these EliteMT 8ns cache modules in hopes of improving my cache timings on the LS486-D board.

The following was all tested at 180Mhz(60x3).

I was able to lower “L2 Cache/DRAM Cycle WS” from 3 to 2 and complete Quake Timedemos in DOS, where I wasn’t before. Unfortunately it’s not completely stable and GLQuake will crash after a couple minutes. see update below

I also noticed some interesting behavior with the “L2 Cache Policy” setting, with “L2 Cache/DRAM Cycle WS” set to 3 I can’t boot with L2 in WB. But if I lower this setting to 2 it works. And my fastest Quake timedemo is achieved with L2 in WB. Beating L2 WT by 0.2fps. This also only seems to work with a 64MB 50ns EDO module. Anything smaller and I take a performance hit with L2 in WB. Can anyone else confirm this behavior?

I’ll try some additional ram modules as pshipkov has suggested to see if I can find stability with these faster timings. Was really hoping to drop the “Cache Write Cycle” settings to 2, but I still can’t get past the memory test at post if I do.

*Update* I’ve found stability with a 16MB 50ns EDO module, losing 0.2fps in quake, as it runs significantly better with L2 in WT with the smaller 16MB module.

Your findings with the L2 Cache/DRAM Cycle WS are not uncommon.

Would you be willing to PM me your source for the LP61256GS-8 chips?

feipoa wrote on 2021-12-31, 12:26:

Your findings with the L2 Cache/DRAM Cycle WS are not uncommon.

Would you be willing to PM me your source for the LP61256GS-8 chips?

PM sent