Was PCI for Amiga a good choice?

[matthey]

Using a PCI setup is the same as using a CSPPC card, they were not in Amiga's plans and are not "original hardware," but much of the appeal of the Amiga (and many other systems) is the ability to adapt to the user's needs; I call this, "Wouldn't it be cool to..."  This is the same statement that gets us in trouble as well -- usually with a significant other.

November 7, 1993:

There is a very good chance Commodore will endorse the PCI bus sometime in the future.  No official announcement has been made yet, of course. However, PCI is very much a solution to a problem I started working on over two years ago. It's very close, in fact, to my solution to this problem. And standard, as well.  Draw your own conclusions :-)

October 01, 2003:

Well, it’s hard to say everything for sure. But I can tell you this. In the fall of 1991, with Sydnes basically cancelling every project, I decided to sit down and design the next system architecture, the thing that would hopefully replace the A3000 design (used in all A3000/A4000 machines). This was called “Acutiator”, and fully modularized the architecture, so that graphics, for example, could be separate from sound and basic I/O. This originally used a custom bus I designed, called the AMI Bus (Amiga Modular Interconnect).

But then a funny thing happened: PCI came out. PCI was designed to solve the very same problem, and by the time Intel kicked it out to the PCI SIG and they improved it, it was way better than the AMI bus at a bunch of things. And also, it was likely to be this huge standard. That’s a good thing....

See, there’s this misconception about C=/Amiga engineering and standards. We LOVED to use standards – any standard – as long as they did not suck. So you see all these proprietary buses and such around the Amiga, and figure, these guys hate standards. Not at all. We liked the good ones. PCI was a very good one, even then.

So, with all of that said, the next generation Amiga would have had a PCI bus. Also, probably, a PCI to Zorro III bridge. Graphics would have been on PCI. I had speced out PCI interface chips for AA and AAA subsystems, so the graphics could go on a card. Not at all cloning The PC; but the functionality is correct, to make these pieces modular if possible. I’ll let you say I’m copying the Apple ][ here is you like – after all, that’s what IBM did anyway.

There was a feature in Acutiator most systems simply don’t have: the TPU, or Transfer Processing Unit. Any time you had a bus to bus interface, you would (ideally) have a TPU there, in the chip that did that bus to bus interface. This was a very simple 32-bit microprocessor (I designed the architecture) which would transfer data, efficiently, from bus to bus. It would so largely because it understood, perfectly, both of the buses at issue. So, no imposed wait states if there were synchronization issues, speed mismatch, etc. You could write directly to memory/IO on the far side of that bus, but better still, just drop a transfer instruction into the queue for a particular TPU, and it would run the transfer for you, then signal when done. The goal: every bus in the system could be busy, all at once.

Anyway, that’s the kind of things I had in mind for the system. For graphics, Hombre, as mentioned, and that was also PCI – Dr. Hepler also saw the wisdom in PCI, even as I did independently. Beyond that, it’s questionable if Commodore would have remained in the graphics business. Most of the PC markers used to make their own graphics chips, too. Today, it’s nVidia, ATi, Matrox, and few others. Like Intel, Motorola, and National Semiconductor, you only need so many different CPUs around.

[matthey]

Is video refresh and speed noticeable fast than a CV 64 for normal tasks, like using final writer or playing a game like nightlong?  

Probably. Higher resolutions and greater screen depths would definitely be noticeably faster as well as allowing higher resolutions and greater screen depths than the CV64 is capable of. The CV64 has a faster gfx bus (can transfer from the Amiga to gfx card faster) but most of the time the more modern gfx card performance and extra gfx memory more than make up for it. Using P96Speed in 640x480x8, I calculated that my 3000T+CSMKIII 68060@75MHz+Mediator 3000(T)/4000(T)+Voodoo 4 is on average 3.00 times faster for P96Speed gfx operations than the 4000T 68060@70MHz+PIV in the P96Speed database. The PIV was overall a little faster than the CV64 (although the CV64 has a little faster gfx bus and some specialized hardware which would make it faster in a few cases). Just for fun, by the same calculation of averaging the "Diff" for all operations in P96Speed, my Amiga is 118.42 times faster than the 4000 68060@50MHz using AGA in the database. I have a faster system with a few tweaks so the gfx performance is not completely isolated by these tests. Once the gfx are fast enough then what is noticeable becomes diminished also. The CV64 is a nice board especially for a 3000 because of the passthrough and hardware planar support. The Mediator offers better performance, more gfx memory, potentially newer gfx outputs like DVI and potentially 3D.

How stable is a rig using all PCI cards on the Mediator?

My "rig" is very stable although the driver installation is not for beginners and the drivers have some annoying but minor bugs that can mostly be worked around.

[matthey]

I don't like it one bit, because it turns awesome Amigas into nothing but glue logic. I don't have an Amiga for that.

ECS is definitely *not* awesome. AGA is *not* awesome either when my glue, some patches and a 20 year old PC gfx card are 118 times the graphics performance of AGA. This is even with a bottle necked slow gfx bus at a fraction of the performance potential of old legacy PCI. It should be possible to create new classic hardware with integrated graphics, avoiding the glue and bottle necks, which is 300-500 times the gfx performance of the old Amigas. Compatibility with ECS/AGA could be much better than an RTG system while providing up to 32 bit depth chunky through a single DVI/HDMI output and possibly 3D support. My high end system and the Natami prototype boards are fast enough to show that the extra graphics (and CPU) performance do not cause major problems for the Amiga. I want a modern Amiga without the "glue"!

[matthey]

@billt
Thanks. My Haynie quotes became garbled since this old thread started and I was too lazy to fix them. The links are nice too.

Who was it, DCE that had a PPC card for the A3000 that was directly connected via cable to the PCI backplane?  Rather than go through the system bus, that's the better way - let the card draw power from the system, do the data transfer directly with the CPU, and use the host system for native chipset, etc. (if we're talking about upgrading old gear).

I believe Phase5 first made the CSPPC/CSMKIII and then DCE bought the rights. There was still some hardware needed to turn the PCI signals into PCI slots or the G-Rex(s) would be simpler. This or a whole new motherboard is the way to go so the Zorro III to PCI bottleneck can be avoided. The later Boxer motherboard and Natami would have had active faster PCI slots as well.

Really? Then tell me, what's the point of an Amiga with a graphics card, sound card and PPC cpu? Where's the Amiga in that?

Want a practical machine? Use a peece, it's the only thing that makes sense. Amigas need to retain their retro value, and they don't retain that if you start using them as glue logic.

So, yes, I like to use the old chipset in the old computer I like. When I need something more powerful, I'll use my peecee.

So you prefer OCS over ECS or AGA and a 68000 over a 68030? Some enhancements are upgrades without losing much if any compatibility. Enhancements like more depth and chunky for the gfx and more bits/channels for the audio could be added to the custom chips with no significant compatibility problems. Likewise, the 68k CPU could be enhanced to provide more speed, more ease of use and better code density with better compatibility than the 68060. Going from 68k to PPC is a major change which has major disadvantages including incompatibility, 50% larger code and difficulty of low level programming and debugging compared to an enhanced 68k. I understand that the old classic Amigas are cool in the same way a vintage car is but they aren't as useful or convenient as a modern daily driver. Try using an unexpanded Amiga 1000 for awhile and let us know how enjoyable your experience is.

[matthey]

Imho, I think that approach should be used for any new fast FPGA accelerators! Best of both worlds, Zorro & PCI.

... or PCIe. FPGAs with SerDes for PCIe and SATA are more expensive but still relatively affordable. PCI is still the best for providing the cheapest commodity hardware and there are existing drivers for the Amiga but PCIe and SATA would be great for a little more cost and work .

[matthey]

AGA+50 mhz 68030 is my preferred system. 68040s and 68060 are too fast, while 68000s are too slow. The 68030 is just right.

I prefer a slower clocked CPU which is easy to program and has consistent real world performance than a high clocked CPU with high theoretical performance but lots of pipeline bubbles and performance bottlenecks. IMO, the 68060 maintains good real world performance and is easy to program. Its weaknesses are primarily due to lack of resources (logic gates) as was common at the time, incomplete internal optimization, and an unchanged ISA and ABI. The lack of resources is not a problem today and a more compatible writethrough cache could provide better compatibility. Instruction scheduling takes a little effort to learn (the 68060 is relatively easy to schedule) but ignoring it would slow your code down if it is too fast. Besides, if the CPU is too fast then use a compiler .

Which is why I use a peecee. Very powerful, and lots of up to date, and often free, software. Lots of nice games, too.

Sure, but do you program your PC or power use your PC? Windows is the lazy man's brain dead OS and Linux is a pain in the butt geek's OS. AmigaOS is freedom in comparison.

Maybe one of each, put the SATA onboard where it can play nice with DMA and Fast ram!

It makes sense to provide PCIe and SATA on the accelerator if the FPGA supports SerDes and PCI (mini SD and PCI cards for drives) otherwise. Supporting both PCIe and PCI would make the boards unnecessarily more expensive.

[matthey]

That may be so, but apparently the  current complex CPUs are the only way forward performance wise, and  performance is the only thing anyone cares about these days.

Complexity adds cost and heat. Performance per clock is more important than maximum clock speed ever since the Pentium 4 days (perhaps Motorola's 68060 successors would have been able to overtake Intel at this point if they hadn't bet the farm on PPC). Die shrinks are responsible for higher clock speeds now but they are reaching practical limits because of cost. The 68060 has good performance per clock and similar technology could be clocked up substantially without causing problems. OoO and expensive die processes would be needed to compete in performance with modern desktop processors though.

What do you see as power use?

I suppose it is not well define. It is basically the OS getting out of the way to do more work. Using many programs at once while maintaining responsiveness, communicating between programs at my request (ARexx), advanced scripting without a book sized manual, etc.

Compared to AOS,  Windows and Linux are awesome powerhouses of might and magic AOS is thoroughly stuck in the past and it will always remain in the past. If only I could replace  it with a better, faster OS on my A1200.

Windows and Linux are powerful but Windows thinks for and limits the user and Linux requires reading bookshelf's equivalent of manuals to use the power. They are both adequate for running programs and PCs have enough processing power to make them fairly responsive even with modern bloated software. I can still do more in less time on my Amiga when I have the software and processing power to do it. Of course the 68k AmigaOS is 20 years out of date because it hasn't been updated. So the slow old Amiga hardware is still good but the fast old AmigaOS is bad today? You are strange indeed .

[matthey]

Yes, but isn't that where increased complexity comes from?

Some processors have better performance per clock, have less complexity and/or use less logic because of their design. CISC processors have a higher complexity and logic cost for a base CPU implementation but then they generally have good performance per clock and take less resources to improve this. RISC processors are cheaper to implement but require lots of resources to make powerful per clock and avoid bottlenecks (big caches, strong OoO, adding some CISC like features, etc.). RISC was originally designed to outperform CISC by out clocking it and by moving complexity from the CPU to the compiler but it lost both of these battles. Most modern powerful RISC processors have added some CISC features and are now RISC/CISC hybrids. Most modern CISC processors have adopted some RISC features and are now CISC/RISC hybrids.

You misunderstand. Amiga hardware is cool because it has a high retro coolness factor. AmigaOS isn't bad, it's just not as good and fast as it could be for 68k.

AmigaOS has a lot of good features and ideas which were not fully developed or implemented. There have been add-ons to the OS which should have been more carefully integrated. The AmigaOS is still fast and responsive by design even if compiler optimization has always been lacking. Most 68k processors are forgiving of poorly optimized code. I believe the 68k AmigaOS could be 20%-40% smaller with better optimization but this would probably only give 10%-20% better performance. Good algorithms are more important to performance.

@wawa
Don't make the AmigaOS 4 guys cry with your flawless logic. Modern in the Amiga world implies at least 10 year old technology. They knew that before they spent a lot of money for "modern" PPC Amiga hardware when a PC with the same performance would have cost a fraction of the price and provided modern OS features. Then again, some of them still hope to convince us 68k Amiga users to take the expensive leap of faith to a more modern out dated Amiga .