Best AGA implementation on FPGA?

[JimDrew]

Anything 040 specific can be made to work with the 020, with the exception of the MMU. The FPU can be completely emulated in software - the original MAC II machines had either a 68882 or nothing, in which case FPU calls are trapped and done in software. Likewise, even some of Apple's 040 machines used LC040's (MMU only, no FPU). There are some differences in the exception handling for the 040, but all in all, it's not a stretch to make the 020 core 040 compatible for anything that does not need the MMU. From a CPU core standpoint, implementing the 040's MMU is hands down easier than the 030, because the 030 has multiple translation table sizes and other oddities that came from the 68851. With the 040 its either 4k or 8K page, and that's it.

[lurch]

Originally Posted by eXeler0

Lurch, JimDrew is actually working on the FPGA arcade, Im just following the project (and got to see the actual board a year ago).
Anyhoo, the FPGA is NOT (yet) the ultimate Amiga setup that gives a 060 equipped 1200 a kicking.
It is an attempt to to a very exact reimplementation of the original A1200, improving speed only where it doesnt break compatibility.

Hmm, really looking at something that will replace my A1200T. Maybe with the daughter board it might replace it, depends on how fast the daughter board is.

A fast 020 is nice, but for demos and some games I play I need an 060.

@JimDrew How fast will the daughter board run at?

[JimDrew]

I think the daughter board uses a 75MHz 060, but I guess that people don't understand that the FPGA's 020 core is much faster than the real 060 daughter board setup.

[robinsonb5]

Originally Posted by JimDrew

I think the daughter board uses a 75MHz 060, but I guess that people don't understand that the FPGA's 020 core is much faster than the real 060 daughter board setup.

Originally Posted by Lurch

A fast 020 is nice, but for demos and some games I play I need an 060.

It's kind of hard to spell it out much clearer than that.

A mega-fast '020 is a very interesting processor but I remain to be convinced that FPU emulation won't come with a severe speed penalty - I'd love to be proved wrong though.

I remember playing with a Mac SoftFPU in ShapeShifter and being delighted to see that the emulated FPU leaked over into the Amiga side too - yet I don't think anyone created an equivalent for the Amiga, did they?

[matthey]

Originally Posted by JimDrew

I think the daughter board uses a 75MHz 060, but I guess that people don't understand that the FPGA's 020 core is much faster than the real 060 daughter board setup.

There is a video on YouTube showing MAC emulation using the daughter board with a revision 6 68060 clocked over 100MHz. Other 68060 revisions would be lucky to clock to 66MHz with most being limited to about 60MHz.

Originally Posted by robinsonb5

A mega-fast '020 is a very interesting processor but I remain to be convinced that FPU emulation won't come with a severe speed penalty - I'd love to be proved wrong though.

I expect from fastest to slowest:

1) hard FPU like 68060
2) FPGA FPU emulation
3) software floating point

Floating point for the 68k requires fast wide shifting which is slow in an FPGA. Of course speed is relative as 6888x performance should be possible in FPGA but even 68060@50MHz FPU performance would likely be difficult.

Originally Posted by robinsonb5

I remember playing with a Mac SoftFPU in ShapeShifter and being delighted to see that the emulated FPU leaked over into the Amiga side too - yet I don't think anyone created an equivalent for the Amiga, did they?

Any software FPU emulation would have been slower than:

1) IEEE math fp libraries (accelerated with FPU)
2) FFP math fp libraries (faster than IEEE but no FPU acceleration)
3) compiling with software fp

SAS/C (Lattice) allowed all 3 options above from very early and other compilers had similar options as well. FPU emulation would likely involve trapping which would be slower than all the options above.

[JimDrew]

With Mac emulation on the 060 you need superscaler off, and most often need the instruction cache off as well (depending on the application). A 33MHz 040 beats a 50MHz 060 in Mac benchmarks.

What makes you think shifting of any width is "slow" on an FPGA?

[matthey]

Originally Posted by JimDrew

With Mac emulation on the 060 you need superscaler off, and most often need the instruction cache as well (depending on the application). A 33MHz 040 beats a 50MHz 060 in Mac benchmarks.

I found the video with the 68060@106MHz.

[ Show youtube player ]

Do you think a 68040 can give this performance? The Mac benchmarks show the 68060@106MHz to outperform the 68040 and even early PPC processors.

Originally Posted by JimDrew

What makes you think shifting of any width is "slow" on an FPGA?

Most FPGAs have multipliers which can be used for shifts but a 64 bit or even 53 bit shift is too wide for this. Larger shifts are not a problem but take more logic and several cycles. Even the 6888x could do a significand (fraction) shift in 1 cycle. Most FPU operations require the shift as part of the normalization process.

[NovaCoder]

Originally Posted by JimDrew

You would need to chat with Mike about the cost of these items. I only sell to North America (that is our agreement). The daughter board is getting a few small changes (from what I have heard), but the basic board is done and works. I am not sure when they will be available. To me, an 060 makes no sense with the speed we can get now with FPGA.

I agree

At this point in time, I'd be more interested in purchasing an new FPGA based accelerator than a real 060 based board.

[lurch]

Originally Posted by JimDrew

I think the daughter board uses a 75MHz 060, but I guess that people don't understand that the FPGA's 020 core is much faster than the real 060 daughter board setup.

So I go with a fast 020, from what I understand I would be unable to run 060 demos or games? I'm not worried if it's a real 060 or not, I just want to be able to run 060 software.

I don't know about other ppl but 95% of the demos and games I want to run need an 060. WHDLoad is another story though. So the only option to get 100% compatibility would be to have either 060 emulation or an 060.

So it is not about speed but compatibility, unless there is a trick to fool software into running on the 020...

Originally Posted by robinsonb5

It's kind of hard to spell it out much clearer than that.

A mega-fast '020 is a very interesting processor but I remain to be convinced that FPU emulation won't come with a severe speed penalty - I'd love to be proved wrong though.

I remember playing with a Mac SoftFPU in ShapeShifter and being delighted to see that the emulated FPU leaked over into the Amiga side too - yet I don't think anyone created an equivalent for the Amiga, did they?

It's not just the FPU needed, MMU is too. Plus most demos and games I run have been based on the 060 and check for that.

[clebin]

Originally Posted by JimDrew

I think the daughter board uses a 75MHz 060, but I guess that people don't understand that the FPGA's 020 core is much faster than the real 060 daughter board setup.

Where you getting that from? I'm getting 030 speeds here, and I thought the cache stuff Mike is working on would bring it up to 040 level. Much faster than an 060?

Chris

[JimDrew]

Originally Posted by matthey

I found the video with the 68060@106MHz.

[ Show youtube player ]

Do you think a 68040 can give this performance? The Mac benchmarks show the 68060@106MHz to outperform the 68040 and even early PPC processors.

LOL! Yeah, FUSION with a 33MHz 040 has way faster benchmarks than that - especially the FPU tests. We wrote complete Pack 4 and Pack 5 replacements, so the Mac FPU speed was 2-5 times faster than the real equiv. speed Mac.

This is why Hollywood production studios used EMPLANT equipped Amigas to replace their Macs for doing rendering and using Avid video editing software.

Originally Posted by matthey

Most FPGAs have multipliers which can be used for shifts but a 64 bit or even 53 bit shift is too wide for this. Larger shifts are not a problem but take more logic and several cycles. Even the 6888x could do a significand (fraction) shift in 1 cycle. Most FPU operations require the shift as part of the normalization process.

I am not an FPGA guru, but the solution for having to use multiple cycles (if that is really necessary) is using a faster base clock for you logic.

[JimDrew]

Originally Posted by clebin

Where you getting that from? I'm getting 030 speeds here, and I thought the cache stuff Mike is working on would bring it up to 040 level. Much faster than an 060?

Chris

Mike has stated that with the caches enabled you will have more than a 4x speed boost to what you are seeing now. That will be improved upon as well.

[clebin]

Originally Posted by JimDrew

Mike has stated that with the caches enabled you will have more than a 4x speed boost to what you are seeing now. That will be improved upon as well.

That's what I basing my crude calculations on - that 4x the speed of a 25mhz 030 would be something like a 25-33mhz 040. I'm happy for others to correct me on that. Still, I don't see how that's "much faster" than a 75mhz (or more) 060 as you said.

A more minor point - Mike has said "close to 4x the speed" but you've translated that as "more than 4x speed boost". Is this based on a separate conversation or is it more exaggeration?

Obviously improvements can happen beyond that, but then we're getting very hypothetical when not even the caching work has been finished yet. You say "the FPGA's 020 core is much faster than the real 060 daughter board setup" without any kind of caveats.

So with all due respect, I think you're being a little misleading and as a reseller you should be clear about the facts if someone is considering putting in a preorder with you. The FPGA Arcade is a great machine with even greater potential and it doesn't need any extra push. Again, if I'm wrong on any of this please correct me.

Chris

[matthey]

Originally Posted by JimDrew

LOL! Yeah, FUSION with a 33MHz 040 has way faster benchmarks than that - especially the FPU tests. We wrote complete Pack 4 and Pack 5 replacements, so the Mac FPU speed was 2-5 times faster than the real equiv. speed Mac.

The 68060 FPU is clocked higher and the instructions are generally less cycles than the 68040 FPU although the 68040 FPU can do a parallel FMOVE and has some advantages in memory/caches. The 68060 FPU should be able to outperform the 68040 FPU with 68040 optimized FPU code most of the time. Minor changes to take advantage of the 68060 FINT/FINTRZ instruction would give the 68060 FPU an even larger advantage. It would be interesting to compare Fusion results to the ShapeShifter results on the same hardware as well as see how well it runs. The FPGA Arcade 68060 daughter board is likely to be the first new 68k hardware capable and compatible enough to run Fusion even if you are biased against the 68060.

Originally Posted by clebin

That's what I basing my crude calculations on - that 4x the speed of a 25mhz 030 would be something like a 25-33mhz 040. I'm happy for others to correct me on that. Still, I don't see how that's "much faster" than a 75mhz (or more) 060 as you said.

As much as some people want the FPGA CPU emulation to compete with the 68060 (it is the only future as the 68060 is dead end), the 68060 is potent, especially when overclocked. I can't see the TG68 FPGA CPU or its variants reaching even 68060@50MHz performance in the FPGA Arcade FPGA. The Phoenix/Apollo CPU in the FPGA Arcade may be able to outperform a 68060@100MHz but it is unlikely to appear in the FPGA Arcade. The 68060 daughter board would offer good performance and excellent value considering the features are more than performance.

1) easily overclockable 68060 performance
2) 100% 68060 compatibility including FPU and MMU
3) 128MB of memory
4) ethernet
5) USB

That's not bad for 100-150 Euros although I wish there was a PCI slot. I hope it gets produced as the Amiga would benefit from all kinds of high spec hardware.

[JimDrew]

You keep forgetting that a real Mac does not run at all with an 060. FUSION works with the 060 after patching the crap out of the OS and disabling superscaler, and some applications also need the instruction cache off or they will crash. So, the 060 is severely crippled, and thus it should not be surprising that it is a lot slower than 33MHz 040 setup... but hey, what would I know.

People have seen what the Apollo core does on a smaller FPGA with a slower memory interface. You would be rather foolish to be believe that substantially better performance on the Replay is far off.

[matthey]

Originally Posted by JimDrew

You keep forgetting that a real Mac does not run at all with an 060. FUSION works with the 060 after patching the crap out of the OS and disabling superscaler, and some applications also need the instruction cache off or they will crash. So, the 060 is severely crippled, and thus it should not be surprising that it is a lot slower than 33MHz 040 setup... but hey, what would I know.

I understand that the 68060 needs some patching to run MacOS but the ShapeShifter video above shows that the 68060 can outperform the 68040 while being stable. Fusion is not as stable as ShapeShifter on the 68060 but it also works. The AmigaOS needs some patching for the 68060 but it is worthwhile and stable. The 68060 daughter board could be available before any FPGA 68k CPU+FPU and an FPGA implementation would likely have bugs making it less stable than a patched 68060 MacOS. The only other possibility near term would be to have a different voltage regulator for the FPGA arcade daughter board to run a 68040 but most people would still choose a 68060 because it is faster for the Amiga and produces much less heat.

Originally Posted by JimDrew

People have seen what the Apollo core does on a smaller FPGA with a slower memory interface. You would be rather foolish to be believe that substantially better performance on the Replay is far off.

The Apollo/Phoenix CPU core has a deeper pipeline, a more parallel advanced design and is highly optimized for the FPGA it is in. The TG68 would need major changes to achieve 68060@50MHz performance. A more advanced design with FPU and MMU may not fit with a chipset. Realistically, this level of performance is unlikely for years while the 68060 daughter board could be out in 6 months or less.

[lurch]

Anyway, just need to know when the Reply and the 060 board is out in the wild. Thinking strongly about buying both.

Is what yaqube was saying in another thread on here still the case?

"The Replay board is able to host Minimig AGA core (which can be switched into OCS/ECS compatibility modes) and allows all AGA stuff to work. This core comes with a lot of additional goodies: 8MB of FAST RAM, 1.5MB of SLOW RAM, 48MB of extra CHIP RAM, support for 2MB ROMs, 24-bit AHI sound, Picasso96 compatible RTG board capable of 1920x1080 resolution and full colour display (although not at the same time).

The Replay board can be easily expanded with different sorts of daughterboards. For example the one which I have made has a real 68060 CPU running at 106MHz, 128MB of local SDRAM, High Speed USB host controller, Fast Ethernet controller, RTC, micro SD card socket, TOSLINK optical audio output."

[Megol]

Originally Posted by JimDrew

I am not an FPGA guru, but the solution for having to use multiple cycles (if that is really necessary) is using a faster base clock for you logic.

What? The reason one use multiple cycles (in a pipelined design) is to enable a higher clock rate.
One can't separate the logic latency from the clock rate.

[JimDrew]

Originally Posted by matthey

I understand that the 68060 needs some patching to run MacOS but the ShapeShifter video above shows that the 68060 can outperform the 68040 while being stable. Fusion is not as stable as ShapeShifter on the 68060 but it also works.

That's not the case. First of all the benchmarks shown are slower than a 33MHz 040. Perhaps you are not familiar with FUSION. It is far more stable, more compatible, and hands down faster, than Shapeshifter.

The Apollo/Phoenix CPU core has a deeper pipeline, a more parallel advanced design and is highly optimized for the FPGA it is in. The TG68 would need major changes to achieve 68060@50MHz performance.

Perhaps, if TG68 was always going to be used.

[JimDrew]

Originally Posted by Megol

What? The reason one use multiple cycles (in a pipelined design) is to enable a higher clock rate.
One can't separate the logic latency from the clock rate.

If you have a base clock of 50MHz and change that base clock to 250MHz, you would see a 5x speed increase, no? Faster part = faster speed.