ReA4091: SCSI2 on Zorro III for everyone! - Page 5

grelbfarlk · 22 June 2022, 05:06

I'm wondering if it's possible to actually use a 53C720 for SCSI-III on the A4091. These chips are still pretty available and as long as the driver can be updated a bit, it should be possible. Seems like it should be possible for Toni to adapt the driver as with the Warpengine ROM for PP&S Zeus, A4091 and CSA Magnum.

cdh · 22 June 2022, 09:51

Sure, it would be possible if you were designing a new card and implement a custom driver. The 53C720 (wide SCSI) isn't pin-compatible with the 53C710. Stefan and I are considering a wide SCSI controller if we ever do a less exact clone of the original card.

alexh · 22 June 2022, 09:53

You also need to consider the bandwidth of Zorro III. Does having SCSI-III have any performance advantage over SCSI-II when using a Zorro III bus?

reinauer · 24 June 2022, 04:43

Quote:

Originally Posted by alexh

You also need to consider the bandwidth of Zorro III. Does having SCSI-III have any performance advantage over SCSI-II when using a Zorro III bus?

AlexH is making a very valid point here. Except in some rare edge cases or once we get a super magical Buster 12, the 710 is pretty much fully saturating the Zorro bus (or, could fully saturate, if multi cycle transfers were implemented)

grelbfarlk · 25 June 2022, 05:22

I don't see how you can say that the A4091 is fully saturating the Zorro bus when people are only getting 5MB/s out of it.
The 53C710 is getting sure, 9MB/s and more on an accelerator.

reinauer · 25 June 2022, 18:51

Ah, that is where the confusion is. Sorry for the unfortunate wording. Well, several things are going on here:

The A4091 does not (currently) support multi transfer cycles. This is unrelated to the NCR53C710 and has to be implemented in the GAL logic. Without this, every data transfer will require address and data on the bus, and unlike Zorro II they share the same signals on the bus. With multi transfer cycles ("Zorro III burst") you put an address once and then keep pushing data onto the bus. I appreciate help on this topic, if you have expertise.
The NCR53C710 supports SCSI-II which maxes out at 10 MB/s. Unless we get that speed onto the bus, there is not really a point in upgrading to a faster chip/newer SCSI.
As in theory Zorro III is an asynchronous bus with up to 150 MB/s theoretical bandwidth. In practice, Super Buster 11 limits you to about 13.5MB/s (and yes if we get there, a 720 would be a reasonable option for getting that 3.5 MB/s extra out of the door.
I have been doing all these tests with an A3640 and an A3660. with either of these accelerators your on-board fast memory speed is limited to about 11MB reading / 6MB writing, chip memory obviously being less. Zorro III memory will also be a good bit slower than that. You are not going to see transfer speeds faster than memory access.
As you observed, the 710 will do 9MB/s on an accelerator. This should be a good indication to you that it is not the part that needs to be fixed here. If you get Zorro III out of the way and attach that same SCSI chip to a 68040/68060 CPU bus and memory with 40+MB/s you can get a very different story of course.

So, yes, you can design a new board and write a new driver (scsi.device) that will work with the NCR720, but that will not make any of the other bottle necks go away and you end up with the same speed. Focusing on fixing GAL logic on the A4091, improving the accelerator (possibly even redoing the GALs on the 3660, although other alternatives are on the horizon for a similar BOM price point) are the gating bottle necks. If those are out of the way, it makes sense looking at a 720/725 or even 770 based storage system and potentially fixing Buster limits with a Buster12 implementation (although here the surrounding logic contributes to some of the bottle necks)

Jope · 25 June 2022, 19:09

Based on info Jens has posted, his Buster 12 will not really improve throughput, as that would break compatibility with many cards.

reinauer · 25 June 2022, 19:17

Quote:

Originally Posted by Jope

Based on info Jens has posted, his Buster 12 will not really improve throughput, as that would break compatibility with many cards.

I read that statement. I have not seen any technical detail on these claims, so it is a bit hand wavy. But it's not unreasonable to assume that trying to optimize there will get you to compatibility nightmares, just from what I have seen in implicit timing assumptions and differences in the Zorro III implementations on the Amiga 3000 and 4000 alone. A lot of stuff out there works because it was tested in those two machines, not because it formally complies with the Zorro III spec.

Of course, in theory if Buster 12 were to become an open source project, users could decide which optimizations to turn on based on the cards they intend to use. But let's not go there, it's on the wishlist somewhere between world peace and the big rock candy mountain.

Jope · 25 June 2022, 19:22

It is what it is yes. I get the idea that a lot of Amiga HW was made based on observed behaviour. Then again the info in the RKM set is not 100% error free either, so definitely some amount of debugging in real systems must have happened.

reinauer · 28 June 2022, 06:34

Some experiments with a ReA4091 somewhere in the 'verse.

alexh · 28 June 2022, 09:46

Is that a big improvement? 50%?

Toni Wilen · 28 June 2022, 21:36

Quote:

Originally Posted by grelbfarlk

I'm wondering if it's possible to actually use a 53C720 for SCSI-III on the A4091. These chips are still pretty available and as long as the driver can be updated a bit, it should be possible. Seems like it should be possible for Toni to adapt the driver as with the Warpengine ROM for PP&S Zeus, A4091 and CSA Magnum.

720 had many incompatible register level changes compared to 710. 720 and later models are almost fully compatible. 710 is different.

Multiple registers have changed and SCSI selection has also changed, 710 had bit mask (can technically select multiple devices at the same time which does not make much sense), 720+ (including 8xx) uses SCSI ID binary number (which is more sane)

When I added 710 emulation long time ago, I originally took 8xx emulation from QEMU (710 emulation didn't exists) and it needed lots of changes to make it 710 compatible. 8xx emulation was almost fully 720 compatible.

reinauer · 01 July 2022, 04:31

Quote:

Originally Posted by alexh

Is that a big improvement? 50%?

It's a Zorro to Zorro transfer to a Zorro III memory card with my personal, overclocked A4091.

Base line on an A3000 things looked like this without overclocking the 53C710:

Code:

Test a4091.device 1 with Chip RAM
read  256 KB xfers          3128 KB/sec
read   64 KB xfers          2888 KB/sec
read   16 KB xfers          2222 KB/sec
write 256 KB xfers          3224 KB/sec
write  64 KB xfers          2689 KB/sec
write  16 KB xfers          1739 KB/sec
Test a4091.device 1 with Fast RAM
read  512 KB xfers          5050 KB/sec
read  128 KB xfers          4912 KB/sec
read   32 KB xfers          4000 KB/sec
write 512 KB xfers          5742 KB/sec
write 128 KB xfers          4990 KB/sec
write  32 KB xfers          3333 KB/sec
Test a4091.device 1 with Zorro III RAM
read  512 KB xfers          6164 KB/sec
read  128 KB xfers          5954 KB/sec
read   32 KB xfers          4705 KB/sec
write 512 KB xfers          7065 KB/sec
write 128 KB xfers          5822 KB/sec
write  32 KB xfers          3478 KB/sec

This is interesting because clearly Zorro III multi cycle transfers are not the whole story.

ahandyman59 · 14 July 2022, 18:39

Are these boards available yet, assembled and/or circuitboard?

reinauer · 16 July 2022, 23:47

Hey folks, since you didn't hear much from us in the last while, here's a small update:

Burn-in Testing

During our tests we have found an Amiga 3000 with timings a lot more borderline than the three Amiga 3000 machines we had been testing on before. This has led us to go back to doing out homework and testing the boards with different components with regards to long term stability. Side note: The Amiga 3000 is a very different beast than the 4000, it is a small miracle how signals hold up (or don't, on occasion).
Unfortunately, older Lattice GALs and the Atmel ATF GALs that are still avaliable today, have profound differences in timing behavior, even if choosing chips with the same timing specs (-10 / -15 / ...) It was very important for us to understand the behavior better and make sure this is not causing problems. In our current configuration our cards run super solid everywhere, over hundreds of test cycles and hours of load on end.

If there's a GAL / WinCUPL expert reading this, please message me.

Batch 2 and 3

If you made it to our "internal list" before the Google form[1] existed, you will be in batch 2 (25 boards). If you signed up in the form, you are in Batch 3 (100+ boards). In the beginning there was no email confirmation upon filling out the form, I added that later unfortunately.
Our batch 2 is almost completely assembled at this point and we are super busy testing. For batch 3 we have decided that we have to get help with the assembly because at our current assembly speed and even with significant time investment it would take a year to assemble these boards, with not doing much else. Not a good use of our and everybody else's time. For batch 3 we are therefore working on getting our fabrication files in order so we can let others help us scale while we finish batch 2 in my garage.

As soon as we have more definite answers about mass production and logistics, we will contact everyone who filled out the Google form. At this point you will have the opportunity to make changes to your order, bail out, etc...

AutoBoot/Bootrom/Driver

We (it was Chris, really, my impact was mostly motivational) have the storage part of the boot rom from scratch, based on the BSD driver for the NCR710. Reading and writing of blocks works great, but a few final pieces of integrating the a4091.device into the boot rom and Amiga OS are not complete yet, and so it is not possible to mount partitions during booting yet. The good news for everyone watching my talk at VCF east[2] is that we solved the speed issue that the driver had at the beginning of May (turns out there's a bug in the BSD driver). Of course the new open source driver also supports TD64 and NSD.

Speed and Overclocking

Haha, yes. My personal prototype card had to become a victim of overclocking (and successfully so!). The design of the A4091 was originally targeting 33MHz instead of the 25 that finally shipped. Initially we assumed that this needs changes in the boot rom, but it appears that that is not a hard requirement.
With an A3660 the A4091 reaches around 5MB/s with a 50MHz oscillator (25MHz clock rate for the GALs and NCR710). At 66MHz (33MHz for the chips) throughput goes up to 5.5MB/s. Not a linear increase, but if you think about it, the A3660 can only write to the A4000's fast ram at about 6MB/s, so this is pretty decent. Empowered by haughtiness and equipped with one of LIV2's GottaGoFaZ3ter ZorroIII memory boards and a 72MHz oscillator, I measured 7.2MB/s (see above) - Mind blowing if you ask me, and very promising.

More tests with other accelerator boards will follow, but for now back to building A4091 boards.

[1] Pre-order / interest form
[2] My talk at VCF East

trixster · 17 July 2022, 00:40

Interesting!

The timing issues between GALs and Atmel ATFs became very apparent when we were testing different combinations to get the DSP working on the AA3000+ with both motherboard 030 and a variety of different accelerators. There's a spreadsheet with dozens and dozens of different combinations and test runs to see what GALs / ATF timings are required and what work and dont work with each acclerator.

reinauer · 03 November 2022, 06:35

Halloween update:

The full project has been released at https://github.com/A4091/ (gerbers, schematic, EasyEDA files, AutoConfig ROM source code, etc)

jbbolcato · 01 December 2022, 18:28

Hey Stefan.

Very interested by your card, what a cool project; well done with completing it! I suspect I won't be the only one with an A3000 boosted by a BFG9060 - increasingly popular 68060 accelerator which sadly forces us to disable the onboard SCSI to avoid write errors/disk corruption. So we'll be all seeking for alternate ways to connect our SCSI gear! I have to settle for IDE & Buddha for the time being...

I declared interest on your website scsi.me - but I am a little unclear if you are still doing a larger batch yourself, or if you released all data for others to pick it up?

Cheers,
JBB

reinauer · 08 December 2022, 22:47

Hi JBB,

I am not planning to build a larger batch myself, hand crafting 30 of these cards has been about as many as I can do. I will probably build a few more over time, after I get some other stuff done, but my focus right now is more finding ways to get this (internationally distributed) mass produced.

Our original intent was to gather interest so we can see how many boards we should order, but interest was 10x more than what I expected, so I had to shift to look for help in the endeavor. Of course the Gerbers are out there, so anyone can make their own as well.

(So if anybody out there is interested in making a bunch of ReA4091 boards, get in contact with me!)

jbbolcato · 10 December 2022, 13:05

Thanks for taking the time to reply!
Well, let’s hope someone picks it up indeed. These revival projects are successful (like the Buddha anniversary or the a3660) so yours should too.
I spotted Chris Edwards unboxing his purple one recently it: should boost coverage and help gets things rolling hopefully.
In the meantime please let me know if you ever end up with a spare one, I’ll snap it up!
Best,
Jbb

25 June 2022, 18:51	#86
reinauer Registered User Join Date: Jan 2020 Location: Mountain View, California Posts: 56	Ah, that is where the confusion is. Sorry for the unfortunate wording. Well, several things are going on here: The A4091 does not (currently) support multi transfer cycles. This is unrelated to the NCR53C710 and has to be implemented in the GAL logic. Without this, every data transfer will require address and data on the bus, and unlike Zorro II they share the same signals on the bus. With multi transfer cycles ("Zorro III burst") you put an address once and then keep pushing data onto the bus. I appreciate help on this topic, if you have expertise. The NCR53C710 supports SCSI-II which maxes out at 10 MB/s. Unless we get that speed onto the bus, there is not really a point in upgrading to a faster chip/newer SCSI. As in theory Zorro III is an asynchronous bus with up to 150 MB/s theoretical bandwidth. In practice, Super Buster 11 limits you to about 13.5MB/s (and yes if we get there, a 720 would be a reasonable option for getting that 3.5 MB/s extra out of the door. I have been doing all these tests with an A3640 and an A3660. with either of these accelerators your on-board fast memory speed is limited to about 11MB reading / 6MB writing, chip memory obviously being less. Zorro III memory will also be a good bit slower than that. You are not going to see transfer speeds faster than memory access. As you observed, the 710 will do 9MB/s on an accelerator. This should be a good indication to you that it is not the part that needs to be fixed here. If you get Zorro III out of the way and attach that same SCSI chip to a 68040/68060 CPU bus and memory with 40+MB/s you can get a very different story of course. So, yes, you can design a new board and write a new driver (scsi.device) that will work with the NCR720, but that will not make any of the other bottle necks go away and you end up with the same speed. Focusing on fixing GAL logic on the A4091, improving the accelerator (possibly even redoing the GALs on the 3660, although other alternatives are on the horizon for a similar BOM price point) are the gating bottle necks. If those are out of the way, it makes sense looking at a 720/725 or even 770 based storage system and potentially fixing Buster limits with a Buster12 implementation (although here the surrounding logic contributes to some of the bottle necks) Last edited by reinauer; 25 June 2022 at 19:09. Reason: add more info

28 June 2022, 06:34	#90
reinauer Registered User Join Date: Jan 2020 Location: Mountain View, California Posts: 56	Some experiments with a ReA4091 somewhere in the 'verse. Attached Thumbnails

16 July 2022, 23:47	#95
reinauer Registered User Join Date: Jan 2020 Location: Mountain View, California Posts: 56	Updates Hey folks, since you didn't hear much from us in the last while, here's a small update: Burn-in Testing During our tests we have found an Amiga 3000 with timings a lot more borderline than the three Amiga 3000 machines we had been testing on before. This has led us to go back to doing out homework and testing the boards with different components with regards to long term stability. Side note: The Amiga 3000 is a very different beast than the 4000, it is a small miracle how signals hold up (or don't, on occasion). Unfortunately, older Lattice GALs and the Atmel ATF GALs that are still avaliable today, have profound differences in timing behavior, even if choosing chips with the same timing specs (-10 / -15 / ...) It was very important for us to understand the behavior better and make sure this is not causing problems. In our current configuration our cards run super solid everywhere, over hundreds of test cycles and hours of load on end. If there's a GAL / WinCUPL expert reading this, please message me. Batch 2 and 3 If you made it to our "internal list" before the Google form[1] existed, you will be in batch 2 (25 boards). If you signed up in the form, you are in Batch 3 (100+ boards). In the beginning there was no email confirmation upon filling out the form, I added that later unfortunately. Our batch 2 is almost completely assembled at this point and we are super busy testing. For batch 3 we have decided that we have to get help with the assembly because at our current assembly speed and even with significant time investment it would take a year to assemble these boards, with not doing much else. Not a good use of our and everybody else's time. For batch 3 we are therefore working on getting our fabrication files in order so we can let others help us scale while we finish batch 2 in my garage. As soon as we have more definite answers about mass production and logistics, we will contact everyone who filled out the Google form. At this point you will have the opportunity to make changes to your order, bail out, etc... AutoBoot/Bootrom/Driver We (it was Chris, really, my impact was mostly motivational) have the storage part of the boot rom from scratch, based on the BSD driver for the NCR710. Reading and writing of blocks works great, but a few final pieces of integrating the a4091.device into the boot rom and Amiga OS are not complete yet, and so it is not possible to mount partitions during booting yet. The good news for everyone watching my talk at VCF east[2] is that we solved the speed issue that the driver had at the beginning of May (turns out there's a bug in the BSD driver). Of course the new open source driver also supports TD64 and NSD. Speed and Overclocking Haha, yes. My personal prototype card had to become a victim of overclocking (and successfully so!). The design of the A4091 was originally targeting 33MHz instead of the 25 that finally shipped. Initially we assumed that this needs changes in the boot rom, but it appears that that is not a hard requirement. With an A3660 the A4091 reaches around 5MB/s with a 50MHz oscillator (25MHz clock rate for the GALs and NCR710). At 66MHz (33MHz for the chips) throughput goes up to 5.5MB/s. Not a linear increase, but if you think about it, the A3660 can only write to the A4000's fast ram at about 6MB/s, so this is pretty decent. Empowered by haughtiness and equipped with one of LIV2's GottaGoFaZ3ter ZorroIII memory boards and a 72MHz oscillator, I measured 7.2MB/s (see above) - Mind blowing if you ask me, and very promising. More tests with other accelerator boards will follow, but for now back to building A4091 boards. [1] Pre-order / interest form [2] My talk at VCF East Last edited by reinauer; 20 July 2022 at 17:33.

03 November 2022, 06:35	#97
reinauer Registered User Join Date: Jan 2020 Location: Mountain View, California Posts: 56	Halloween update: The full project has been released at https://github.com/A4091/ (gerbers, schematic, EasyEDA files, AutoConfig ROM source code, etc) Last edited by reinauer; 09 November 2022 at 08:18.

01 December 2022, 18:28	#98
jbbolcato Registered User Join Date: Mar 2022 Location: Abingdon Posts: 12	Hey Stefan. Very interested by your card, what a cool project; well done with completing it! I suspect I won't be the only one with an A3000 boosted by a BFG9060 - increasingly popular 68060 accelerator which sadly forces us to disable the onboard SCSI to avoid write errors/disk corruption. So we'll be all seeking for alternate ways to connect our SCSI gear! I have to settle for IDE & Buddha for the time being... I declared interest on your website scsi.me - but I am a little unclear if you are still doing a larger batch yourself, or if you released all data for others to pick it up? Cheers, JBB Last edited by jbbolcato; 01 December 2022 at 18:35.

22 June 2022, 05:06	#81
grelbfarlk Registered User Join Date: Dec 2015 Location: USA Posts: 2,902	I'm wondering if it's possible to actually use a 53C720 for SCSI-III on the A4091. These chips are still pretty available and as long as the driver can be updated a bit, it should be possible. Seems like it should be possible for Toni to adapt the driver as with the Warpengine ROM for PP&S Zeus, A4091 and CSA Magnum.

22 June 2022, 09:51	#82
cdh Registered User Join Date: Dec 2017 Location: Los Altos, CA, USA Posts: 55	Sure, it would be possible if you were designing a new card and implement a custom driver. The 53C720 (wide SCSI) isn't pin-compatible with the 53C710. Stefan and I are considering a wide SCSI controller if we ever do a less exact clone of the original card.

22 June 2022, 09:53	#83
alexh Thalion Webshrine Join Date: Jan 2004 Location: Oxford Posts: 14,337	You also need to consider the bandwidth of Zorro III. Does having SCSI-III have any performance advantage over SCSI-II when using a Zorro III bus?

25 June 2022, 05:22	#85
grelbfarlk Registered User Join Date: Dec 2015 Location: USA Posts: 2,902	I don't see how you can say that the A4091 is fully saturating the Zorro bus when people are only getting 5MB/s out of it. The 53C710 is getting sure, 9MB/s and more on an accelerator.

25 June 2022, 19:09	#87
Jope - Join Date: Jul 2003 Location: Helsinki / Finland Age: 43 Posts: 9,861	Based on info Jens has posted, his Buster 12 will not really improve throughput, as that would break compatibility with many cards.

25 June 2022, 19:22	#89
Jope - Join Date: Jul 2003 Location: Helsinki / Finland Age: 43 Posts: 9,861	It is what it is yes. I get the idea that a lot of Amiga HW was made based on observed behaviour. Then again the info in the RKM set is not 100% error free either, so definitely some amount of debugging in real systems must have happened.

28 June 2022, 09:46	#91
alexh Thalion Webshrine Join Date: Jan 2004 Location: Oxford Posts: 14,337	Is that a big improvement? 50%?

14 July 2022, 18:39	#94
ahandyman59 Registered User Join Date: Mar 2017 Location: Tacoma, WA USA Posts: 94	Are these boards available yet, assembled and/or circuitboard?

17 July 2022, 00:40	#96
trixster Guru Meditating Join Date: Jun 2014 Location: England Posts: 2,337	Interesting! The timing issues between GALs and Atmel ATFs became very apparent when we were testing different combinations to get the DSP working on the AA3000+ with both motherboard 030 and a variety of different accelerators. There's a spreadsheet with dozens and dozens of different combinations and test runs to see what GALs / ATF timings are required and what work and dont work with each acclerator.

08 December 2022, 22:47	#99
reinauer Registered User Join Date: Jan 2020 Location: Mountain View, California Posts: 56	Hi JBB, I am not planning to build a larger batch myself, hand crafting 30 of these cards has been about as many as I can do. I will probably build a few more over time, after I get some other stuff done, but my focus right now is more finding ways to get this (internationally distributed) mass produced. Our original intent was to gather interest so we can see how many boards we should order, but interest was 10x more than what I expected, so I had to shift to look for help in the endeavor. Of course the Gerbers are out there, so anyone can make their own as well. (So if anybody out there is interested in making a bunch of ReA4091 boards, get in contact with me!)

10 December 2022, 13:05	#100
jbbolcato Registered User Join Date: Mar 2022 Location: Abingdon Posts: 12	Thanks for taking the time to reply! Well, let’s hope someone picks it up indeed. These revival projects are successful (like the Buddha anniversary or the a3660) so yours should too. I spotted Chris Edwards unboxing his purple one recently it: should boost coverage and help gets things rolling hopefully. In the meantime please let me know if you ever end up with a spare one, I’ll snap it up! Best, Jbb

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Zorro III Garbled Colors	CryptidWorks	support.WinUAE	11	14 January 2021 07:27
WinUAE OS3: Picasso IV Zorro III Vs UAE Zorro III	AMIGASYSTEM	support.WinUAE	3	17 October 2016 13:18
Zorro III not working... :(	DarrenHD	support.Hardware	19	27 January 2011 05:41
AD516 Zorro II \ III ?	THX1138	support.Hardware	6	16 February 2009 00:23
Zorro II Card - Zorro III Slot ?	THX1138	support.Hardware	4	03 May 2003 11:03