NatAmi 68070 design draft

[Plaz]

@ alexh

How's this for my own theory. Dual Coldfire setup. CFv4/266 to start. It's primary job, to run the JIT for backward code compatability. The JIT feeds the second CPU ( CFv2/83 or perhaps another CFv4 @90-100mhz) that takes the place of the 68K.

I've made this suggestion before, but it didn't get much milage.

Plaz

[Plaz]

@bloodline

Probably because your idea defeats the point of a JIT...

How so? Just because the function is on a second core? How would a dual or quad core handle JIT?

Plaz

[Plaz]

@ alexh

I imagined they were trying to do it all in one CPU, and that can only handle so much.

Plaz

[Plaz]

@bloodline

The JIT executes only once per code block (in theory)...

In a perfect world, yes. But this is Amiga. :-) Trying to reproduce 100% compatibility with less than compatible hardware is going to be more costly than the original or custom part in mass. So yes, some cost might be wasted on a second CPU. However if the second CPU is relatively unused and over powered, you could either use a less expensive version, or use the extra horse power to drive peripherals like ethernet, DDR ram, USB.... of course this is all in theory. Unless we're going to port AOS to x86, fast AND cheap just isn't coming our way any time soon.

Plaz

[Plaz]

@bloodline

and missing addressing modes add lots of instructions to the translated code

Yes, I did a lot of reading while part of the amiga coldfire project mailing list. That's where the dual theory began to start spining in my head as solutions to the less than adaquate supervisor mode.

P.S. Oh and don't forget Ami started out with serveral "co-processors" in the first place.  (I know... that's stretching it)

Plaz

[Plaz]

@bloodline

like just how suited the Coldfire actually is to running 68k code

You're correct, it isn't. That's why you put in the JIT.

You are suggesting a symmetric design, which is expensive for this...

Not so sure. Not in CPU's anyway. 166/V2's about $20US each. 200/V4e's About $32US each. 68060 if you can find new stock....$500US. Coding bios an glue logic for a card is where the time would be spent though.

Plaz

[Plaz]

The 68k Addressing modes can require a huge amount of additional logic and arithmetic code to emulate and we haven't even covered Flag calculation

There's only about 12 instructions to deal with. You just never know when their going to hit the CPU though. The coldfire is a 68K core and about 99.2% compatible with 68x00. The trouble comes when one of thes dozen instructions does a diffent task on the CF than it does on the cold fire. Supervisor mode can capture and redirect invalid instructins to a lib where software takes care of it. However, these few instructions are not invalid and can't be trapped that way. You have to intercept them some how. If you have to intercept a few, you're just as well off intercepting them all with a fast JIT. IO is another problem. V2 is very pin compatible, but slow. V4 is much faster, but not IO compatible. So... to be more specific on my idea. V4 to do JIT to handle invalid instructions, and "feed" the V2 to handle buss IO and execute program code.

All good conversation, but obviouly going nowhere unless I win the lottery this week :-P

Just buy a cheap PC (£250 including everything including a monitor!) and run WinUAE

Not any cheaper than a new card if some one can come up with one. PC emulation is a perfect idea for some things, but I can't run all my hardware. For video and music studio work I might as well just stick with my XP and all the USB hardware. As long as emulation is just amiga "running in a vacuum" it's not useful to me. It all comes back to... no hardware. Ok I'm depressed now. Time to go work on the toaster/flyer box.

Plaz

[Plaz]

it will be a nightmare to develop and write software for unless you abstract the entire system... and if you are going to do that... why use Coldfire?

Yes, I mentioned earlier that the bios (firmware) and glue logic would be the biggest tax on such a design. Coldfire would have seemed like the logical successor to 68K for backward code compatibility with classics, but it's proving to be untrue. Too limited, too slow. A card with an embedded JIT and a 4ghz x86 chip wouldn't be easy either, but starts to make more sense.

Plaz

[Plaz]

@alexh

In order for them to be ridiculed, first they must write something ridiculous surely?

Surely you jest. One only need post the observed weather in some places to become a target for fodder. :-)

Plaz

[Plaz]

@bloodline

The ARM is low power, well supported, it's not going anywhere and with a nice JIT should reach standard 680x0 performance with the current generation...

Something else I need to look into then.

Plaz