Insights on how accelerator cards work on Amigas

[jman]

Hello eveyone,

I'd like to better understand how these cards generally work on the Amiga platform. I probably have a lot of of wrong ideas in my head, so I'd appreciate someone who can shed some light.
I tried to have a random look at some devices on amiga-hardware.com but descriptions do not asnwer my questions.

So, if I'm not wrong, these are not "daughter" boards as it was a 3DFX Voodoo, they do not accelerate just some specific task: they are plugged tipically to an expansion slot and *replace* completely the main CPU. They do have they own RAM, possibly a FPU chip. Latest Cybervision also had a dedicated GPU unit, though.

Q1) Does the stock CPU and its bus/dma is excluded and replaced by the new one? Or is it still used for "light" tasks such as input interrupt, device access and alike? In this case does the stock CPU (and its bus) is itself a bottleneck?
Q2) How much of the original host system is not excluded? Maybe the custom OCS/ECS/AGA chipset?
Q3) Is the expansion slot a bottleneck for these external cards?

Well, I think that's all I need to know for the moment :-)

Thank you!

[Jope]

Originally Posted by jman

Q1) Does the stock CPU and its bus/dma is excluded and replaced by the new one? Or is it still used for "light" tasks such as input interrupt, device access and alike? In this case does the stock CPU (and its bus) is itself a bottleneck?

The stock CPU is disabled, but the accelerator is linked to its bus.

Q2) How much of the original host system is not excluded? Maybe the custom OCS/ECS/AGA chipset?

Everything is still there. Access to the motherboard components is limited by the capabilities of the connection to the motherboard. The motherboard keeps running at the same clock speed as it used to.

The accelerator's CPU needs to wait until the right time to access anything that is not on the accelerator board's local bus to keep things in sync.

Q3) Is the expansion slot a bottleneck for these external cards?

Yes.

[AndyLandy]

Originally Posted by jman

Q3) Is the expansion slot a bottleneck for these external cards?

Rather more fundamentally, the Amiga itself is the bottleneck in most cases. That's the reason why FastRAM really needs to be local to the card for any real performance. When you get to something like the Cyberstorm PPC, with on-board Fast RAM, SCSI and connect up a G-REX busboard, it's almost as if the CSPPC is the mainboard and the A4000 motherboard is just an expansion board that has the Amiga custom chipset on it.

[jman]

Originally Posted by Jope

The accelerator's CPU needs to wait until the right time to access anything that is not on the accelerator board's local bus to keep things in sync.

As an example, we can say that an application/videogame on a floppy would load at the same speed of an unexpanded machine but then would run at lightspeed (if not properly syncronized with the Amiga video signal).

Same thing if this application is loaded from HD? In this case the bottleneck would be the "stock" bus connection between the CPU on the accelerated card and the system HD, correct?

[Claw22000]

Too bad they didn't make the processors replaceable.

[Jope]

Originally Posted by jman

As an example, we can say that an application/videogame on a floppy would load at the same speed of an unexpanded machine but then would run at lightspeed (if not properly syncronized with the Amiga video signal).

Theoretically yes, but in reality it will never work as the CPU would be accessing the custom chip registers at the totally wrong moment of the clock pulse. You'd just end up with a crashed machine.

Same thing if this application is loaded from HD? In this case the bottleneck would be the "stock" bus connection between the CPU on the accelerated card and the system HD, correct?

Well, same thing meaning that if it's not clocked right, it won't work.

However with a HD controller, transfers will speed up if it is not a DMA controller. This is because non-DMA controllers occupy the CPU to move the data from the hard drive to memory. A faster CPU has more time to run your applications and handle the data transfer, hence the non-DMA HD controller can potentially go at maximum speed, if your CPU is fast enough and you don't have too many processes running.

You might want to read the Lucas documentation that is linked on support.hardware.mods for some ideas on what kind of problems manifest when you connect the accelerator to the 68000 bus.

The A3000 and A4000 are different because the accelerator slot is connected to a 68030 style bus..

[jaycee]

Hmm, just wondering if anyone here knows how exactly the onboard CPU is inhibited ? I had thought XCLKEN would be used but thats not on the expansion slot. It'd help me solve a mystery with an A1200 i have - it wont work unless an accelerator is plugged in!

[jman]

Originally Posted by Jope

Theoretically yes, but in reality it will never work as the CPU would be accessing the custom chip registers at the totally wrong moment of the clock pulse. You'd just end up with a crashed machine.

Thank you, that was interesting. I will give a glance at the documentation you suggested.

[rkauer]

Originally Posted by jaycee

Hmm, just wondering if anyone here knows how exactly the onboard CPU is inhibited ? I had thought XCLKEN would be used but thats not on the expansion slot. It'd help me solve a mystery with an A1200 i have - it wont work unless an accelerator is plugged in!

Probably you have a dead EC020 CPU on the board. Using an accelerator replaces the broken CPU.

All accelerators send a BUS_REQUEST ( _BR signal) to original CPU, then the original CPU send a BUS_REQUEST_ACKNOWLEDGED followed by BUS_REQUEST_Granted signals, then the acc will start its business.

Some acc even don't make all this fuss (and cut the CPU chit-chat), they send a HALT or BR signal to the CPU and take over the bus.

Notice that on A500 the 68000 still gives the E-clock to the system.

[jaycee]

Originally Posted by rkauer

Probably you have a dead EC020 CPU on the board. Using an accelerator replaces the broken CPU.

All accelerators send a BUS_REQUEST ( _BR signal) to original CPU, then the original CPU send a BUS_REQUEST_ACKNOWLEDGED followed by BUS_REQUEST_Granted signals, then the acc will start its business.

Some acc even don't make all this fuss (and cut the CPU chit-chat), they send a HALT or BR signal to the CPU and take over the bus.

Notice that on A500 the 68000 still gives the E-clock to the system.

I had figured that if the 020 was dead, it would most likely misbehave on the bus and screw up the system... need to get my 'scope on that board I think!


I Think I see what you're saying though, i dont really know the 68k bus cycle very well, but you're saying the accelerator says "hey, can i have the bus?", and the 020 says "sure, tell me when youre done". The accelerator never answers, so the 020 sits there. Pretty neat.

[Stedy]

@jman

Take a look at the A1200 functional spec:
http://www.ianstedman.co.uk/downloads/A1200FuncSpec.txt

And the A500/A2000 technical reference manual:
http://www.shiftreload.com.au/users/4x4/schematics/

Ian

[jaycee]

Aha thanks Ian, that answers my questions too


Hmm, I was hoping it would give me some info on the iffy A1200 board I have. I was hoping maybe the onboard oscillator was dead and that an accelerator was supplying a suitable system clock. I'd still expect a damaged 020 to interfere with the bus and prevent proper functioning.


Does anyone have any other ideas other than "dead 020"? I've got an oscilloscope though it is only 60MHz and not digital or storage.

[rkauer]

Nope, a dead 020 will not interfere with the accelerator. A guy (here or on Amibay, don't really remember) had a destroyed 020 CPU inside his A1200 (the CPU even had a hole + crack on the top!). Zetr0 replaced the dead CPU for a new one and all was fine again.

The Amiga worked fine with an accelerator hooked in.