https://www.ardent-tool.com/tech/ASIC_Info.html#71G0438

The Lacuna and Reply use BIUs, The Reply 60/80 also actually has an SSC.

_MAYBE_ the inclusion of a BIU turbo-diddles the performance of the 9-K?

A weg to investigate this angle would be to use a Type 3 ["M"] or a Type
4 ["N"] complex, both of which do not have a BIU.

Kevin Bowling wrote:
>
>> That the Lacuna and Reply perform so low is disappointing, but maybe
>> there's a reason why IBM did not sell this adapter to PS/2 users - or
>> maybe we will find a magic switch to make the card operate faster.

Another factoid, the Lacuna and Reply boards use IDE.

Even _IF_ the BIU / IDE thing is a lead weight, I don't expect a non-IDE
/ non-BIU system to hit 10MB/s...

So even though Albert Einstein breathed air like I do, that correlation
does not mean that we are both geniuses...

Louis Ohland wrote:
> https://www.ardent-tool.com/tech/ASIC_Info.html#71G0438
>
> The Lacuna and Reply use BIUs, The Reply 60/80 also actually has an SSC.
>
> _MAYBE_ the inclusion of a BIU turbo-diddles the performance of the 9-K?
>
> A weg to investigate this angle would be to use a Type 3 ["M"] or a Type
> 4 ["N"] complex, both of which do not have a BIU.
>
> Kevin Bowling wrote:
>>
>>> That the Lacuna and Reply perform so low is disappointing, but maybe
>>> there's a reason why IBM did not sell this adapter to PS/2 users - or
>>> maybe we will find a magic switch to make the card operate faster.

Am 21.11.23 um 22:00 schrieb Wolfgang Gehl:
> Am 21.11.23 um 17:38 schrieb Christian Holzapfel:
>> Hear ye, Hear ye!
>> There's a new dream team on the block.
>> That may be you, your favorite PS/2 and your IBM 10/100 Mbps Ethernet
>> Adapter (9-K)!
>> After weeks and months of reverse and forward engineering, we now have
>> drivers for this very late manufactured, 32-bit Micro Channel 100 Mbps
>> capable Ethernet Adapter type 9-K, code name "San Remo", that IBM
>> released only for the RS/6000 series as one of the very last MCA cards.
>> For 30 years there were no drivers outside of AIX, IBM's proprietary
>> Unix operating system.
>> Today you may start running it in Windows 95, probably also 98, and
>> Linux 2.2!
>> Get your copy here:
>> http://www.holzapfel.biz/8F62/sanremo-win9x.zip
>> https://github.com/holzachr/sanremo-linux
>> Please note that those are considered beta drivers, and may not be
>> perfect yet.
>> They were tested on Windows 95 B OSR2 and Debian 2.2 on a souped-up PC
>> 750. I would love to hear your personal feedback on a true Micro
>> Channel system.
>> This was a joint effort of Ryan Alswede and me.
>> Enjoy! 😀
>
>
> Cool stuff, will definitely try it out on a 9595 or server 520 and
> report back. If only I had a little more time to play around with MCA ...
>
> The problem with using a 9-K in an Intel system is the processor load.
> On a 7030-3BT I have about 4.5MB throughput with a comparatively
> sluggish AIXwindows. A 9595 will not achieve this. As soon as a GUI
> comes into play, the 90MHz Pentium will go to its knees.
>
> In any case, thank you for your work, which I greatly appreciate!
>
> Wolfgang

Sorry for coming late to the show. Here are the results for a 9595. The
test runs were performed five times and averaged.

Model 9595 (Server 95)
Pentium 90 MHz
256 MB RAM
Windows 95C
sanremo driver first release
Netio 1.32

Packet size 1k bytes: 2003.55 KByte/s Tx, 2474,76 KByte/s Rx.
Packet size 2k bytes: 2842,76 KByte/s Tx, 2988,59 KByte/s Rx.
Packet size 4k bytes: 3416,27 KByte/s Tx, 3265,17 KByte/s Rx.
Packet size 8k bytes: 4255,07 KByte/s Tx, 3328,77 KByte/s Rx.
Packet size 16k bytes: 4730,52 KByte/s Tx, 3286,24 KByte/s Rx.
Packet size 32k bytes: 4570,87 KByte/s Tx, 3137,11 KByte/s Rx.

With your driver, the 9-K is three to four times faster than the
Etherstreamer MC32, which was up to now the fastest Ethernet adapter for
the IBM PS/2. The system remains responsive even if Sysinternal's
process explorer claims 100% CPU utilization.

Well done!

Will try your Linux driver with Slackware 11 and the second release of
your W95 driver as soon as possible.

Wolfgang

Well, there are different amounts of RAM.

What effect does adding RAM have to performance? Where is the sweet spot?

Wolfgang Gehl wrote:
> 256 MB RAM

Seems to be an imbalance in the Rogaine flow.

Especially 8K - 32K, 1K-1.5K difference.

Why?

Wolfgang Gehl wrote:
>
> Model 9595 (Server 95)
> Pentium 90 MHz
> 256 MB RAM
> Windows 95C
> sanremo driver first release
> Netio 1.32
>
> Packet size  1k bytes:  2003.55 KByte/s Tx,  2474,76 KByte/s Rx.
> Packet size  2k bytes:  2842,76 KByte/s Tx,  2988,59 KByte/s Rx.
> Packet size  4k bytes:  3416,27 KByte/s Tx,  3265,17 KByte/s Rx.
> Packet size  8k bytes:  4255,07 KByte/s Tx,  3328,77 KByte/s Rx.
> Packet size 16k bytes:  4730,52 KByte/s Tx,  3286,24 KByte/s Rx.
> Packet size 32k bytes:  4570,87 KByte/s Tx,  3137,11 KByte/s Rx.

Seems to be an imbalance in the Rogaine flow.

Especially 8K - 32K, 1MB - 1.5MB difference.

Why?

Wolfgang Gehl wrote:
>
> Model 9595 (Server 95)
> Pentium 90 MHz
> 256 MB RAM
> Windows 95C
> sanremo driver first release
> Netio 1.32
>
> Packet size 1k bytes: 2003.55 KByte/s Tx, 2474,76 KByte/s Rx.
> Packet size 2k bytes: 2842,76 KByte/s Tx, 2988,59 KByte/s Rx.
> Packet size 4k bytes: 3416,27 KByte/s Tx, 3265,17 KByte/s Rx.
> Packet size 8k bytes: 4255,07 KByte/s Tx, 3328,77 KByte/s Rx.
> Packet size 16k bytes: 4730,52 KByte/s Tx, 3286,24 KByte/s Rx.
> Packet size 32k bytes: 4570,87 KByte/s Tx, 3137,11 KByte/s Rx.
> Model 9595 (Server 95)
> Pentium 90 MHz
> 256 MB RAM
> Windows 95C
> sanremo driver first release
> Netio 1.32
>
> Packet size  1k bytes:  2003.55 KByte/s Tx,  2474,76 KByte/s Rx.
> Packet size  2k bytes:  2842,76 KByte/s Tx,  2988,59 KByte/s Rx.
> Packet size  4k bytes:  3416,27 KByte/s Tx,  3265,17 KByte/s Rx.
> Packet size  8k bytes:  4255,07 KByte/s Tx,  3328,77 KByte/s Rx.
> Packet size 16k bytes:  4730,52 KByte/s Tx,  3286,24 KByte/s Rx.
> Packet size 32k bytes:  4570,87 KByte/s Tx,  3137,11 KByte/s Rx.
>
> With your driver, the 9-K is three to four times faster than the
> Etherstreamer MC32, which was up to now the fastest Ethernet adapter for
> the IBM PS/2. The system remains responsive even if Sysinternal's
> process explorer claims 100% CPU utilization.
>
> Well done!
>
> Will try your Linux driver with Slackware 11 and the second release of
> your W95 driver as soon as possible.
>
> Wolfgang
>
>

Louis Ohland wrote:
> Seems to be an imbalance in the Rogaine flow.
> Especially 8K - 32K, 1MB - 1.5MB difference.
>
> Why?

That must be Kevin's interrupt rate.
For the same total data rate, many small packets are more difficult than a few larger ones.

What's also interesting is Lionel's 486-era Pentium Overdrive 83 versus Wolfang's "real" Pentium 90 - what a huge gap.
Pretty much like comparing the RS6k workstation and server.
So MHz is not everything (but it surely helps...).

Thank you guys for testing. I'm glad it works out for you.

lharr...@gmail.com wrote:
> What are the chances of a DOS driver?

The leaked sources for the original PCnet DOS packet driver are around.
They are plain x86 Assembly, processed by

# MAKE Version 3.6
# TASM Version 3.1
# TLINK Version 5.1

Generally, the places such a driver needs modification to work with our 9-K are well known to Ryan and me and properly documented now - but I'm not that fluent in Assembly (yet).
Furthermore, the DOS driver is working in 16-bit mode only, while the 9-K ASIC and also the PCnet chip in our case need some 32 bit addressing.
So it's a little more to it, but generally doable.
Maybe in the boring, gray start of next year I could look into it.
If someone else is willing to pick that up, I'm happy to help :-)

Seems the secret is using a non-IBM SCSI controller with the RP2040
ZuluSCSI.

BusLogic BT-646 / SDC3211F
ZuluSCSI RP2040

lharr...@gmail.com wrote:
> On Sunday, December 3, 2023 at 7:12:51 AM UTC-5, Christian Holzapfel wrote:
>> lharr...@gmail.com wrote:
>>> What are the chances of a DOS driver?
>> The leaked sources for the original PCnet DOS packet driver are around.
>> They are plain x86 Assembly, processed by
>>
>> # MAKE Version 3.6
>> # TASM Version 3.1
>> # TLINK Version 5.1
>>
>> Generally, the places such a driver needs modification to work with our 9-K are well known to Ryan and me and properly documented now - but I'm not that fluent in Assembly (yet).
>> Furthermore, the DOS driver is working in 16-bit mode only, while the 9-K ASIC and also the PCnet chip in our case need some 32 bit addressing.
>> So it's a little more to it, but generally doable.
>> Maybe in the boring, gray start of next year I could look into it.
>> If someone else is willing to pick that up, I'm happy to help :-)
>
> I've been pretty decent writing code for c# and such, but not sure if I could pick this up and be useful. I'd have to start reading about it and see.
>
>
> As far as the 9x driver, I have been unable to get the updated driver working at all. Even with a fresh install of Windows 95. I'll get an IP but can't ping the gateway, the reply times out. I didn't change any of the default settings of the driver... I think the buffer or whatever was set to RX + TX. This weekend has been bit busier than I expected and I have to run so.. not sure how much I can play with it more this weekend. Also... with the fresh install, the 32bit driver for my BusLogic card is intermittently loading... which what the hell?!?! So I want to nuke the install again and find a more trusted ISO of Windows 95 C. Downloaded some fresh copies last night and had one fail that it couldn't find a file in a cab in the middle of the install... turns out people commented on the download page that the ISO was wonky. I dont burn the ISO's to CD-R as ZuluSCSI will emulate them as a CDROM if you put the ISO on the SDCard and name it properly.
>
>
> Anyway, with the fresh install, performance was a tad better. I truly think the reply board is slow in general. The 486 Overdrive 100mhz and Kingston TurboChip also both score about the same as the POD83... which is 20% slower than any rando board out there that can take these chips. I've had a few others with the Reply board get the same results. Not sure why.
>
>
> TCP connection established ...
> Receiving from client, packet size 1k ... 1267.69 KByte/s
> Sending to client, packet size 1k ... 1425.52 KByte/s
> Receiving from client, packet size 2k ... 1583.88 KByte/s
> Sending to client, packet size 2k ... 1693.53 KByte/s
> Receiving from client, packet size 4k ... 2123.17 KByte/s
> Sending to client, packet size 4k ... 1846.58 KByte/s
> Receiving from client, packet size 8k ... 2132.16 KByte/s
> Sending to client, packet size 8k ... 1903.23 KByte/s
> Receiving from client, packet size 16k ... 2304.99 KByte/s
> Sending to client, packet size 16k ... 1941.89 KByte/s
> Receiving from client, packet size 32k ... 2434.78 KByte/s
> Sending to client, packet size 32k ... 1929.03 KByte/s
>

tiistai 21. marraskuuta 2023 klo 18.40.23 UTC+2 Christian Holzapfel kirjoitti:
> And as our friend Shane pointed out:
>
> FYI, there’s a seller on eBay that has >200 of these in “new, bulk” condition for ~$60 shipped and shipping costs drop if you buy multiple cards.
>
> https://www.ebay.com/itm/193693679438
Best 10/100 NIC for MCA PS/2 machines! And Christian has made drivers for it, thanks!
I would love to get it for my 9577 but sending costs from USA are too much for me.
Has someone in EU to sell these NICs! Because selling costs inside EU are more resonable
and without customs that collect VAT from outside EU sent items.

I did two more benchmarks.
This time I'm on my PC 750 (PCI/MCA) @ 400 MHz again, running Linux 2.2.17, using Netio 1.11 (because I don't have GLIBC3 there) and running the RX direction, so I'm sending data from a modern multi-core, multi-GHz system to the 9-K:

NETIO - Network Throughput Benchmark, Version 1.11
(C) 1997-1999 Kai Uwe Rommel

TCP/IP connection established.
1k packets: 6513 k/sec
2k packets: 5564 k/sec
4k packets: 5805 k/sec
8k packets: 5256 k/sec
16k packets: 6782 k/sec
32k packets: 7308 k/sec

As a comparison, I am using the exact same setup, but sending data to a 9-P card, which is the PCI variant of our MCA 9-K.
It has the same PCnet ethernet chip and chip revision, accompanied by the same amount and speed rating of on-board memory.
The Linux drivers of the PCnet (9-P) and San Remo (9-K) are 99.5 % identical with the exception that the 9-K driver tunnels all control I/O accesses through the ASIC, which is not speed relevant.
All driver parameters like RX/TX queue lengths, interrupt rate, burst settings, error handling and all other configuration of the PCnet chip's registers are exactly the same.
So from the hardware point of view, the only difference is that in the 9-K case, the hardware access is tunneled through the PC 750's PCI-to-MCA bridge, and then through the 9-K's MCA-to-PCI ASIC bridge.

This is how the PCI 9-P performs:

NETIO - Network Throughput Benchmark, Version 1.11
(C) 1997-1999 Kai Uwe Rommel

TCP/IP connection established.
1k packets: 11433 k/sec
2k packets: 11422 k/sec
4k packets: 11533 k/sec
8k packets: 11543 k/sec
16k packets: 11478 k/sec
32k packets: 11404 k/sec

So I dare to conclude that the system itself, Linux, the general driver structure (no matter if 9-P or 9-K), the interrupt rate, the data rate, DMA and CPU usage and of course the PCnet chip itself are generally capable of saturating a 100 Mbit Ethernet link under the exact same operating parameters..
The only difference is that the 9-K is attached to the system through the two bridge chips in my case.

There might be some parameters in IBM's original AIX driver that they adjusted inside the ASIC or PCnet to make those two work better together to squeeze a little more out of it - but generally (see 9-P), there is no obvious misconfiguration of the PCnet chip.

>> There might be some parameters in IBM's original AIX driver that they adjusted inside the ASIC
Would be cool to see what the ASIC parameters are on your AIX server for the fields we aren't able to infer their meanings. Maybe a project for next year when your time allows.

0x1D 1 WO Init 0x00 Written to on Init only
0x1E 1 WO Init 0x4F Written to on Init only
0x1F 1 WO Init 0x04 Written to on Init only
0x20 2 WO Init 0x03FF Written to on Init only
0x22 1 WO Init 0x7F Written to on Init only

I keep having the suspicion that these are passed to the PC-Net
registers. Are these values passed to the PC-Net registers?

0x1D 1 WO Init 0x00 Written to on Init only
0x1E 1 WO Init 0x4F Written to on Init only
0x1F 1 WO Init 0x04 Written to on Init only
0x20 2 WO Init 0x03FF Written to on Init only
0x22 1 WO Init 0x7F Written to on Init only

Ryan Alswede wrote:
>>> There might be some parameters in IBM's original AIX driver that they adjusted inside the ASIC
> Would be cool to see what the ASIC parameters are on your AIX server for the fields we aren't able to infer their meanings. Maybe a project for next year when your time allows.

>>Are these values passed to the PC-Net registers?
Negative, Captain. ASIC only.

I smell DMA settings but IBM won't tell us anything so much as a bit.

You smell that? It's DMA. Nothing else in the world that smells like it.
DMA... it smells like... Victory!

Ryan Alswede wrote:
> I smell DMA settings but IBM won't tell us anything so much as a bit.

Am 04.12.23 um 14:39 schrieb Christian Holzapfel:
> So from the hardware point of view, the only difference is that in the 9-K case, the hardware access is tunneled through the PC 750's PCI-to-MCA bridge, and then through the 9-K's MCA-to-PCI ASIC bridge.

My guess is a timing problem in the PCI-to-MCA bridge. May I ask you for
another test run? The PC 750 supports a PCI bus clock of 50MHz. I could
imagine that the MCA bus would cope much better with this than with the
66MHz bus clock.

The AMD K6-III can handle an external clock of 50MHz according to the
PowerLeap upgrade manual:
https://ardent-tool.com/CPU/PL-K6-IIIv2.pdf
It then runs internally at 300 MHz. That should be enough to get
100Mbits out of the LAN.

Wolfgang Gehl schrieb am Mittwoch, 6. Dezember 2023 um 23:48:51 UTC+1:
> My guess is a timing problem in the PCI-to-MCA bridge. May I ask you for
> another test run? The PC 750 supports a PCI bus clock of 50MHz. I could
> imagine that the MCA bus would cope much better with this than with the
> 66MHz bus clock.

With a 50 MHz base clock instead of 66, the adapter tops out at ~6700 k/sec:

NETIO - Network Throughput Benchmark, Version 1.7
(C) 1997-1999 Kai Uwe Rommel

TCP/IP connection established.
1k packets: 5326 k/sec
2k packets: 6079 k/sec
4k packets: 6273 k/sec
8k packets: 6648 k/sec
16k packets: 6755 k/sec
32k packets: 6680 k/sec

I guess the PCI clock is always BaseClock/2 on my system.
Wish I had a datasheet for the PC 750 clock chip, IMI SC471...

Let's play the name game.

IMI is who? Perhaps we can findt who bought them.

Christian Holzapfel wrote:
> IMI SC471

International Microcircuits Inc. (IMI)

Louis Ohland wrote:
> Let's play the name game.
>
> IMI is who? Perhaps we can findt who bought them.
>
> Christian Holzapfel wrote:
>> IMI SC471

In 2000, Cypress acquired IMI. After the acquisition, Refioglu ran the
timing technology division for Cypress. Then, he left the company and
started SpectraLinear in 2006. The startup then acquired Cypress' PC
clock-chip division in October of 2006.

Louis Ohland wrote:
> International Microcircuits Inc. (IMI)
>
> Louis Ohland wrote:
>> Let's play the name game.
>>
>> IMI is who? Perhaps we can findt who bought them.
>>
>> Christian Holzapfel wrote:
>>> IMI SC471

https://www.ic72.com/pdf_file/i/26693.pdf

Still haven't got the website yet.

Not the SC471, but it's something

Louis Ohland wrote:
> In 2000, Cypress acquired IMI. After the acquisition, Refioglu ran the
> timing technology division for Cypress. Then, he left the company and
> started SpectraLinear in 2006. The startup then acquired Cypress' PC
> clock-chip division in October of 2006.
>
>
>
> Louis Ohland wrote:
>> International Microcircuits Inc. (IMI)
>>
>> Louis Ohland wrote:
>>> Let's play the name game.
>>>
>>> IMI is who? Perhaps we can findt who bought them.
>>>
>>> Christian Holzapfel wrote:
>>>> IMI SC471

https://ardent-tool.com/datasheets/IMI_SC425.pdf

Louis Ohland wrote:
> https://www.ic72.com/pdf_file/i/26693.pdf
>
> Still haven't got the website yet.
>
> Not the SC471, but it's something
>
> Louis Ohland wrote:
>> In 2000, Cypress acquired IMI. After the acquisition, Refioglu ran the
>> timing technology division for Cypress. Then, he left the company and
>> started SpectraLinear in 2006. The startup then acquired Cypress' PC
>> clock-chip division in October of 2006.
>>
>>
>>
>> Louis Ohland wrote:
>>> International Microcircuits Inc. (IMI)
>>>
>>> Louis Ohland wrote:
>>>> Let's play the name game.
>>>>
>>>> IMI is who? Perhaps we can findt who bought them.
>>>>
>>>> Christian Holzapfel wrote:
>>>>> IMI SC471

www.imicorp.com

Louis Ohland wrote:
> https://ardent-tool.com/datasheets/IMI_SC425.pdf
>
> Louis Ohland wrote:
>> https://www.ic72.com/pdf_file/i/26693.pdf
>>
>> Still haven't got the website yet.
>>
>> Not the SC471, but it's something
>>
>> Louis Ohland wrote:
>>> In 2000, Cypress acquired IMI. After the acquisition, Refioglu ran
>>> the timing technology division for Cypress. Then, he left the company
>>> and started SpectraLinear in 2006. The startup then acquired Cypress'
>>> PC clock-chip division in October of 2006.
>>>
>>>
>>>
>>> Louis Ohland wrote:
>>>> International Microcircuits Inc. (IMI)
>>>>
>>>> Louis Ohland wrote:
>>>>> Let's play the name game.
>>>>>
>>>>> IMI is who? Perhaps we can findt who bought them.
>>>>>
>>>>> Christian Holzapfel wrote:
>>>>>> IMI SC471

ftp://ftp.best.com/pub.i/imiweb8/ds/sc471.pdf

Louis Ohland wrote:
> www.imicorp.com
>
> Louis Ohland wrote:
>> https://ardent-tool.com/datasheets/IMI_SC425.pdf
>>
>> Louis Ohland wrote:
>>> https://www.ic72.com/pdf_file/i/26693.pdf
>>>
>>> Still haven't got the website yet.
>>>
>>> Not the SC471, but it's something
>>>
>>> Louis Ohland wrote:
>>>> In 2000, Cypress acquired IMI. After the acquisition, Refioglu ran
>>>> the timing technology division for Cypress. Then, he left the
>>>> company and started SpectraLinear in 2006. The startup then acquired
>>>> Cypress' PC clock-chip division in October of 2006.
>>>>
>>>>
>>>>
>>>> Louis Ohland wrote:
>>>>> International Microcircuits Inc. (IMI)
>>>>>
>>>>> Louis Ohland wrote:
>>>>>> Let's play the name game.
>>>>>>
>>>>>> IMI is who? Perhaps we can findt who bought them.
>>>>>>
>>>>>> Christian Holzapfel wrote:
>>>>>>> IMI SC471

https://web.archive.org/web/19970409051608/http://www.imicorp.com/products/prodlit/sfds/sc471.htm

crappy...

Louis Ohland wrote:
> ftp://ftp.best.com/pub.i/imiweb8/ds/sc471.pdf
>
> Louis Ohland wrote:
>> www.imicorp.com
>>
>> Louis Ohland wrote:
>>> https://ardent-tool.com/datasheets/IMI_SC425.pdf
>>>
>>> Louis Ohland wrote:
>>>> https://www.ic72.com/pdf_file/i/26693.pdf
>>>>
>>>> Still haven't got the website yet.
>>>>
>>>> Not the SC471, but it's something
>>>>
>>>> Louis Ohland wrote:
>>>>> In 2000, Cypress acquired IMI. After the acquisition, Refioglu ran
>>>>> the timing technology division for Cypress. Then, he left the
>>>>> company and started SpectraLinear in 2006. The startup then
>>>>> acquired Cypress' PC clock-chip division in October of 2006.
>>>>>
>>>>>
>>>>>
>>>>> Louis Ohland wrote:
>>>>>> International Microcircuits Inc. (IMI)
>>>>>>
>>>>>> Louis Ohland wrote:
>>>>>>> Let's play the name game.
>>>>>>>
>>>>>>> IMI is who? Perhaps we can findt who bought them.
>>>>>>>
>>>>>>> Christian Holzapfel wrote:
>>>>>>>> IMI SC471