FPGA based A500 accelerator

[NovaCoder]

Is this promising project MIA?

Another one bites the dust?

Or just taking a break maybe...

[Promilus]

I believe I saw Mike taking a part in discussion about Buffee on discord, I do not think it's dead project though. I might be repeating myself but I think using mass produced dev-boards like qmtech's would be much faster than designing and assembling custom boards. All you really need with dev boards are voltage level translators. With Spartan6, Artix7, Cyclone IV and Cyclone V you get the same pin headers (well, not exactly the same wiring - since FPGAs are different, but supply pins are in the same place and fpga signals takes the rest). So there could be universal "turbo motherboard" with any of mentioned above as a daughterboard. Of course that also means you would have to make 4 different projects. But with Altera that's most likely only pin assignment and synthesis for different model. With Xilinx that's also different environment (spartan6 is webpack ise only, artix7 35k is iirc vivado-only, 100k should work with webpack). That's the general idea behind what I am pursuing atm. And if it works out there might be even Zynq7000 option (board has different pinout though).

[Gorf]

I hope this project is not abandoned ... is still would be a nice alternative to Buffee
(and we don't know how long this will really take and how fast it will be in the end...)

The big advantage is the mass storage, as Buffee will not provide one...

Also a slimmed down version with only Gayle emulation and card-reader and support for a real (50MHz CMOS?) 68000 or later Buffee would be a nice sandwich solution!

[mkstr]

Not abandoned, but I have been too busy to think about it with lockdowns and home-schooling. I am also watching PiStorm (and Buffee) closely in case it renders this obsolete. I've felt for a while that software emulation would offer the most cost effective accelerator solution, but I also note that there are a lot of "purists" in the scene that might prefer an FPGA solution over software, because it's considered to be less like emulation. I was surprised to see (just now) that PiStorm is still a way behind my prototype in terms of SysInfo Dhrystones.

The A1200 version, by the way, hit a roadblock because the FPGA hasn't got enough pins to be able to use the same one.

Regarding a dev board solution - if I was to look at this option it would probably be with the DE0 nano (with a corresponding reduction in performance) because it's a better form-factor. The Arty board has barely got enough pins and they are split across different headers sitting in different orientations, so the prototype has all kinds of bodge wires to make it work.

[mkstr]

Oh, and the idea of doing a standalone Gayle+SD had crossed my mind already. This part would fit in a small MAX10 or maybe even something smaller.

[Higgy]

This board is still very relevant.
A1200 users are desperate (my words ) for something more than 030 performance and (a lot) cheaper than the upcoming 060 cards. (I hate to see what Brexit is going to do to the shipping/importing cost of the ACA1240/1260).
I agree on the purist comment.

[Gorf]

Originally Posted by mkstr

Oh, and the idea of doing a standalone Gayle+SD had crossed my mind already. This part would fit in a small MAX10 or maybe even something smaller.

That plus maybe some Clockport pins would probably be a great thingy for A500/A1000/A2000 users ...

Would it make sense to spit the project anyways and to make it modular?
Just FPGA-CPU for the ones that want one.
Just Gayle+SD(+Clockport) for others.
Combine both if you want ...

[Gorf]

Originally Posted by mkstr

The A1200 version, by the way, hit a roadblock because the FPGA hasn't got enough pins to be able to use the same one.

could a second smaller FPGA, maybe with some additional multiplexers solve this problem?

Here again it would be great to have a modular solution:
Best would be adapter board for the A1200 with a DIP 68k socket for your FPGA-68k.
Bus translation and expansion to 32bit would be done by the chips on the adapter board.
This would also be great for Buffee or others ...

(I think we have already enough solutions that try to do all in one)

[utri007]

Originally Posted by mkstr

Not abandoned, but I have been too busy to think about it with lockdowns and home-schooling. I am also watching PiStorm (and Buffee) closely in case it renders this obsolete. I've felt for a while that software emulation would offer the most cost effective accelerator solution, but I also note that there are a lot of "purists" in the scene that might prefer an FPGA solution over software, because it's considered to be less like emulation. I was surprised to see (just now) that PiStorm is still a way behind my prototype in terms of SysInfo Dhrystones.

The A1200 version, by the way, hit a roadblock because the FPGA hasn't got enough pins to be able to use the same one.

Regarding a dev board solution - if I was to look at this option it would probably be with the DE0 nano (with a corresponding reduction in performance) because it's a better form-factor. The Arty board has barely got enough pins and they are split across different headers sitting in different orientations, so the prototype has all kinds of bodge wires to make it work.

Just make small patch to see how much you can sell. This supports DMA, so it would allow many things, wich are not possible with pufee.

[mkstr]

Originally Posted by utri007

Just make small patch to see how much you can sell. This supports DMA, so it would allow many things, wich are not possible with pufee.

Actually it doesn't support external DMA for the same reason Buffee doesn't. As an FPGA solution I could theoretically support it, but it complicates the level shifter circuit because the entire bus ends up having to be bidirectional.

[mkstr]

Originally Posted by mkstr

Actually it doesn't support external DMA for the same reason Buffee doesn't. As an FPGA solution I could theoretically support it, but it complicates the level shifter circuit because the entire bus ends up having to be bidirectional.

Also the fastmem uses the Z3 address space which would be inaccessible from the outside due to the 24-bit address bus on 68000 and EC020 systems. Might be a feature worth considering on a future KEL card version, but not before.

[Promilus]

Also the fastmem uses the Z3 address space which would be inaccessible from the outside due to the 24-bit address bus on 68000 and EC020 systems.

A590 DMAC goes only that far anyway so it's Z2 fast ram option alone and there's really no benefit of trying it to reach further. That said... uwscsi of BlizzardPPC/CyberstormPPC hasn't been beaten yet, more than 20 years after those cards were released. Pretty shame the most sophisticated amiga accelerator nowadays reaches HALF of the UWSCSI level. And UWSCSI does less than half of UDMA5 (and less than 10% of SATA3).

[ex68k]

Originally Posted by Promilus

A590 DMAC goes only that far anyway so it's Z2 fast ram option alone and there's really no benefit of trying it to reach further. That said... uwscsi of BlizzardPPC/CyberstormPPC hasn't been beaten yet, .

How fast was/is that?

[Promilus]

DMA of A590 - probably tops at ~2MB/s, UW-SCSI of Blizzard has theoretical max at 40MB/s but usually it goes at around 30-35MB/s, FastATA has maximum of 16,6MB/s but usually it's 5-7MB/s, same with Cyberstorm MK2 SCSI (which has a max of 10MB/s). Vampire supposedly achieve near 20MB/s. I know there are many ppl which find that level of performance satisfying. But again - if you're satisfied with 30-year old performance level why trying to get something faster than 030@50MHz? If someone is trying to push things with way faster RAM and faster CPU core... it only makes sense to provide faster storage option as well.

[mkstr]

Originally Posted by Promilus

A590 DMAC goes only that far anyway so it's Z2 fast ram option alone and there's really no benefit of trying it to reach further.

The problem is that there is no fastmem in the Z2 area - it's in the Z3 area which, as you say, the A590 can't see.

I guess the fastmem could be split so 8MB goes in the Z2 space and the rest is declared as 32-bit memory, but you'd end up wasting nearly half the RAM because autoconfig can only declare powers of two in each region afaik.

[robinsonb5]

Originally Posted by mkstr

I guess the fastmem could be split so 8MB goes in the Z2 space and the rest is declared as 32-bit memory, but you'd end up wasting nearly half the RAM because autoconfig can only declare powers of two in each region afaik.

Autoconfig can only reserve regions in powers of two, but it's not mandatory that the region is entirely populated - have a look at http://amigadev.elowar.com/read/ADCD.../node02C8.html
According to that, if you set the subsize field of er_flags to "0001" the OS will automatically determine how much RAM is actually present, in 512k increments.

[Promilus]

Originally Posted by mkstr

The problem is that there is no fastmem in the Z2 area (..)

Oh well that depends how you define Zorro II area since ZII is basically 68k bus with buffering and additional control signals.
It doesn't matter much for me. If someone has antique and working A570 or A590 (or similar devices) those already covers at least some portion of original 68000@68EC020-based Amiga Fast Ram space (from 24bit addressing space). And it only matters that much with DMAC equipped - again- just a handful. I do not think covering DMA from such devices is worth it. While SCSI DMA on pure A500 is way better than PIO it still sux and doesn't go (much) better with faster CPU (while PIO seems to gain more).

[Valken]

I just want to add my comments but I think mkstr should continue the project.

Buffee is awesome but it is focused on 68000 market.

This is 68K + up to whatever plus FPGA has a lot of modern HW options - USB, WIFI, HDMI output, mini-jack audio in/out, possible MIDI emulation and much more.

I know times are hard world wide, but I think it is because of this and lack of spending power have pushed more people to look at bang for the buck AND good enough, not necessarily the latest and greatest.

I see a huge wave of interests in Retro even among my friends who have really good means. Instead of paying scalper prices for mass production products that basically are in essence "accelerators" to their current PC or Mac (CPU, MB, RAM and GPU), they want a new or new-retro classic experience.

Console sales, VR and retro are the new wave.

Amiga, Atari ST, Mac and even console fans (Megadrive, NeoGeo, Sharp 68K (yeh, I know it is a PC technically, CPS) are looking to enhance their experience and that drive is create a new, alb et small, market.

I seem room for Buffee, this, Vampire, re-Amigas and even WinUAE/Hatari on current x86/x64 hardware.

To validate my comment, I am (now) working back in IT, thank Jebus. The market is BOOMING like a fughamudda (TM) and the commercial aspects are there now. We build out custom hardware in the 100s of units, not 1000s as the market is opened to it but the business keeps growing.

There is room for all of these projects currently in the market and more as I believe for every 10K WinUAE user, there will be 100 HARDWARE users and among those HW USERS, they will want to own multiples of different projects.

The price to them is worth it compared to paying 2x-3x for JUST a barebones PC CPU or GPU today.

If you are unsure of the market potential, do a rough cost and toss it out there. Let the market tell you to proceed or not. As I told the Buffee team on discord, hit up the Mac68k group as well, our Atari and 68K console cousins. I want to see you guys succeed.

[utri007]

Originally Posted by mkstr

The problem is that there is no fastmem in the Z2 area - it's in the Z3 area which, as you say, the A590 can't see.

I guess the fastmem could be split so 8MB goes in the Z2 space and the rest is declared as 32-bit memory, but you'd end up wasting nearly half the RAM because autoconfig can only declare powers of two in each region afaik.

If you could make this work with CDTV and price would be something like 150€ or little less you could easily sell about 300 units in CDTV facebook group.

[Seiya]

i think that these new and open hardwares destroy a little Amiga productivity because it several times that i not read about news games or new apps (ok, that some time someone release a game, but very few things). Why in 2021 i have to buy a card to use the same software that i used happily with A500 or A1200?

Today i think that everyone know that Amiga is requested in very hard way in retrogaming market and not in hardware market.

The only software that i see very very most daily updated are Sysinfo and Aibb that are only software that support new hardwares :P, so i see so many benchmarks discussions and threads and stops.

Amiga needs strong advertise in retrogaming world much more than is today with its softwares because who really need extreme hardware when software is not good for it.