New Universal Agnus 8372, 75 etc

[nonarkitten]

Originally Posted by NorthWay

So riddle me this; I thought the OCS/ECS chip memory was 16 bits wide at 280ns, so is AGA really slower than 140ns to need to use double-CAS? Shouldn't there be time to simply do two accesses in 280ns? And to follow on from that: Shouldn't AGA be able to practically do quad-CAS in 280ns?

Effing EAB and their session length, I lost my reply. Dammit. Have to remember to always CTRL-C everything first...

Okay, so in a nutshell because I don't feel like retyping it all, FPM needs 70ns plus 50ns per word after the read, so 120ns is the minimum cycle time for one word. Lisa works on 140ns cycles (70ns edges) and that aligns nicely with 70ns and 140ns edges to read -- the second read is "late" but that's okay with DRAM.

Data is read into a FIFO which converts any read in an up-to 4-long, 16-bit wide stream for the bitplane shifters to serialize. If the shifters are only running at lowres, then we get idle cycles which kindly gives the CPU more time. When the last word is latched from the FIFO into the shifters, then Alice reads then next chunk.

To do quad-CAS two things would need to happen. First, you'd have to clock the data bus up to 35ns cycles (four times as fast!) to get a 52.5ns delay from the first 70ns on both Alice and Lisa. And then you can just barely get four reads in.

Secondly, you'd also have to double up the FIFOs and that would add several thousand transistors to a chip budget that Commodore already didn't want to spend anything on. Every DMA channel would now need 8-deep, 16-bit FIFOs and that's getting really expensive.

Remember, Commdore hated AGA. It was a "stop gap" and they fought mercilessly to suppress information on it insisting AAA was "right around the corner."

So sum-up, AGA as it was couldn't. Lisa and Alice would need to operate at a higher clock and Lisa would need significantly larger FIFOs to handle the data rate and both of these exceeded the willingness of Commodore to invest anymore in the original Amiga design.

[nonarkitten]

Originally Posted by Promilus

@roondar - nope. AGA is 32b chipset EXCEPT blitter which is inside Alice and still 16bit wide (just like Alice memory interface). That's also a reason why with even slightly improved A1200 blitter itself was becoming obsolete as it was pretty hard to use it efficiently in conjunction with fast, fully 32b CPU. It was just easier to leave it out and save those DMA slots. As the blitter was a workhorse for OCS/ECS effects and amazing performance and was in decline in AGA it just meant good ideas by using "co-processors" in chipset were running out at that time.

Chip RAM is a hybrid 16/32 bus. The CPU and Lisa access 32-bits at a time and Agnus and Paula are both 16-bits. If the Blitter supported double-CAS (though not 32-bit) it would have been as fast as the CPU. Using "co-processors" wasn't in decline, it was just that Commodore left the Amiga chipset to languish and wasn't willing to invest more into it. I think when AGA was "kicked off" they had the budget for ONE new chip. That they snuck in a "tweaked" Agnus was already pushing things.

Today we all have "co-processors," our GPUs, and that trend started very soon after the demise of the Amiga. From S3 and 3Dfx to our modern NVIDIA and Radeon cars, they've been "co-processors" the whole way. In fact, 3D was the first to really define a "standard" for accessing them (via OpenGL and DirectX) which up until then had limited the acceleration you could get on SVGA.

[Promilus]

@nonarkitten - blitter coprocessor wasn't changed at all since ICS. And surely THAT KIND of coprocessor was in decline (in gaming consoles) as their functionality was replaced by fully programmable coprocessors offering more functions (and even some basic pseudo 3d effects). So trying to sell in 92 the very same thing which did work in 85 was plainly stupid.

[nonarkitten]

Originally Posted by Promilus

@nonarkitten - blitter coprocessor wasn't changed at all since ICS. And surely THAT KIND of coprocessor was in decline (in gaming consoles) as their functionality was replaced by fully programmable coprocessors offering more functions (and even some basic pseudo 3d effects). So trying to sell in 92 the very same thing which did work in 85 was plainly stupid.

It was more than PC's had in 92. The S3 Virge didn't come out until 1995 and was $199 US when it did, just for the card. We didn't have standards yet either, so games were very hit-or-miss. This didn't happen until miniGL and OpenGL (c. 1996) and later Direct3D (c 1998), and if memory serves, OpenGL was really slow at first. $10,000+ Silicon Graphics workstations not included.

AGA was not so bad when you fairly compare it to contemporary hardware. Could it have been better? Sure, but Commodore was split on AAA, AGA and Hombre on an already underfunded engineering team. They cut corners and engineers made bad decisions in a rush to push anything to market. They didn't WANT AGA to succeed. They did everything they could to delay AGA in the great hope of AAA. So I'm sometimes surprised it was as good as it was.

[Promilus]

It was more than PC's had in 92.

Oh? Except copper and sprites PCs in 92 had blitter in their graphic cards (but ... since there was no standard solution it wasn't widely used in games, but it was used - through dedicated drivers - in Windows and some software). And also those cards already supported modes like 800x600x8bpp or 640x480x16bpp. Way beyond AGA capabilities. So... no. Not really.

[nonarkitten]

Originally Posted by Promilus

Oh? Except copper and sprites PCs in 92 had blitter in their graphic cards (but ... since there was no standard solution it wasn't widely used in games, but it was used - through dedicated drivers - in Windows and some software). And also those cards already supported modes like 800x600x8bpp or 640x480x16bpp. Way beyond AGA capabilities. So... no. Not really.

Liar liar, pants on fire.

In 1992, your average "high end" PC was a 386SX20 with 1MB of RAM and SVGA that could do 640x480 8bpp and 1024x768 16-colour, not 16bpp. And that beast would set you back $1799 from Radio Shack USD.


SVGA then was over ISA which was cripplingly slow and the 386SX20 which had to do basically all the work was MUCH slower than a 14MHz 68020. SVGA was little more than a dumb frame buffer. Even the "blitter" on a few high-end cards (which alone cost more than a whole Amiga) offered little more than simple character blitting or line-drawing for CAD work and outside of a few CAD tools had no support.

By 1994 when Commodore was going bankrupt, the best Radio Shack had to offer was a 486SX33 with 4MB and VLB based SVGA pushing 1024x768 in an earth-shattering 8bpp (and 16bpp in 320x240) ... all for the low-low price of $2,199.

But don't take my word for it.
[ Show youtube player ]

The 486SX was about as fast as an 020 clock-for-clock and by then a lot of us already had 030's and/or Fast RAM in our 1200's. My first, and almost immediate upgrade, was an FPU and 4MB SIMM. I don't remember who made that, but it was nice. The next was a GVP "Jaws" 68EC030 at 40MHz (which a year or so back I maxed it out to 32MB and swapped the CPU for a regular 68030).

[nonarkitten]

Originally Posted by Promilus

@nonarkitten - blitter coprocessor wasn't changed at all since ICS. And surely THAT KIND of coprocessor was in decline (in gaming consoles) as their functionality was replaced by fully programmable coprocessors offering more functions (and even some basic pseudo 3d effects). So trying to sell in 92 the very same thing which did work in 85 was plainly stupid.

Also, "OCS." That was bugging me.

Also, consoles never had a Blitter. Atari used Display-List-Lists. Sega and Nintendo were both sprite or tile-based as was the Neo Geo as was most arcade hardware. The Atari DLL was as close to a COPPER as you can get, but neither have ANYTHING to do with the Blitter.

Also, early 3D cards were NOT fully programmable. While early GPU's could pull off some cool tricks with multi-pass rendering, we did not get programmable GPU's until the GeForce 256 in 1999.

Clearly age has not been kind to your memory.

[nonarkitten]

My bad, the T&L on the GeForce 256 was half-programmable, the predecessor to vertex shaders. But we didn't get pixel shaders until the GeForce 3 in 2001. Then the whole pipeline was "mostly programmable." This coincided with the launch of DirectX 8.1 which also provided general programmability and the dawn of the "GPGPU."

That's *almost* a decade since the last Amiga model was released and seven years after Commodore's demise in a decade of unprecedented computer evolution that we'll never see again. In 1990 we got the SNES. In 2000 we had the Dreamcast and PS2. In 1990 we could play Flight Simulator 2 on our Tandy 1000 in EGA mode at 2-3 FPS. In 2000 we could play Flight Simulator 2000 with texture mapped cities, buildings, all at a decent 10-20fps.

In 1991, this passed as a great "3D racer".
[ Show youtube player ]

In 2001, this was a great 3D racer.
[ Show youtube player ]

[grelbfarlk]

There was also the Glide era of 3D acceleration, at about 1997.
[ Show youtube player ]

[nonarkitten]

Originally Posted by grelbfarlk

There was also the Glide era of 3D acceleration, at about 1997.
[ Show youtube player ]

Yes, Glide was pretty impressive, but this is getting off topic.

[grelbfarlk]

Originally Posted by nonarkitten

Yes, Glide was pretty impressive, but this is getting off topic.

I'm trying to subconsciously get you thinking of tacking on a chip for Warp3D, with not a whole lot of memory, say 16MB.

[nonarkitten]

So there's two basic methods of rasterizing polygons. Most classic games and demos use the triangle rasterizer that uses two Bresenham interpolators to determine the start end end points and then draws the horizontal line between them. If they're using texture mapping, they'll have additional coordinates that "walks" across the texture in the same way. This method is good for a single core CPU but scales poorly.

There's another and it's based on barycentric coordinates. This is how most modern fragment shaders work and it basically allows unlimited number of parallel shaders with pretty minimal up-front computation. It has built-in perspective correction as well.

So this is fun and all, where am I going?

So I had this terrible, terrible idea of making the fragment shaders out of a minimal core or some sort -- say like a J68. It only needs about 1500 LUTs, so you could have say sixteen on even a modest FPGA all running in parallel as little rendering cores. Yeah, they're integer. Yeah, they're slow. But you have sixteen of them and all they need to do is return an 8-bit RGB based on a 8-bit UVW coordinate (where 0 to 255 represents 0 to 1.0 in barycentric coordinates).

It would be awesome. Evil, but awesome.

[QuikSanz]

Abstract, out of the box,, impressive!

[Eaglef00t]

Originally Posted by nonarkitten

Liar liar, pants on fire.

In 1992, your average "high end" PC was a 386SX20 with 1MB of RAM and SVGA that could do 640x480 8bpp and 1024x768 16-colour, not 16bpp. And that beast would set you back $1799 from Radio Shack USD.

And this is without monitor!

[Promilus]

Originally Posted by nonarkitten

Liar liar, pants on fire.

By 1994 when Commodore was going bankrupt, the best Radio Shack had to offer was a 486SX33 with 4MB and VLB based SVGA pushing 1024x768 in an earth-shattering 8bpp (and 16bpp in 320x240) ... all for the low-low price of $2,199.

It's easy to verify. There's something called PC Mag - that was pretty damn big magazine and released monthly. There are scans on google books. It takes one look to verify what kind of configurations WERE available BACK IN 1992/93.
You don't believe me? Here you go
https://books.google.pl/books?id=LIy...page&q&f=false
Page 165 and downwards might, just might enlighten some about what PC world actually had to offer back then when it comes to VGA card. And magazine itself is filled with ads of 486 based systems.
Who's the liar now?

[Mr.Flibble]

Yeah, 486's were certainly around in '92, anywhere from the sx-25 up to the dx2-50, and they were still all around the $2000 mark, but back then the video hardware was a but ropey. 1 meg was the norm but there were a few 2meg cards, which usually meant a max of 256 colours at 1024x768, although there were a few which could raise that to 65,536.

By the time '94 rolled around , 2-4meg cards were available, ATi had the Mach64 based cards out which could handle 24bit colour up to 1280x1024, and Matrox were spitting out some monsters (although they were very expensive).

Calling folks liars all over the shop isn't really helping anyone though. 30 years ago is enough to make anyone's memory a little fuzzy here and there.

[Promilus]

And to think it begun when I told AGA blitter wasn't workhorse (because it wasn't). Blitter did work well with 7MHz 68000 - made things much faster without actually hurting 68k memory access to chip ram (which is the only ram for stock A500, A600, A1200 etc.) In A1200 there's 68EC020 running at 14MHz and accessing 32bit chunks of memory. Blitter was still fairly useful to bit shifting and logic operations.
VGA blitter wasn't nearly as useful, as nonarkitten wrote it was mostly drawing lines and eventually filling polygons. But the functionality of bit shifting and logic operations was on VGA as well, just in different functional block. And each manufacturer implemented slightly different functionality in much different way. So it wasn't standard. That's why hardly any game use those except things like Lotus, CAD or windows itself.

But still - my point wasn't about PC market but gaming consoles of which whole gen5 (starting with 3DO and Jaguar in 93) had at least one DSP, vector processor or something similar to support fixed-function hardware with things they couldn't do. That's why AAA (which was initially in-development around '88) was intended to have 32bit blitter already running faster and with much greater and faster chipram. Why this piece of hardware didn't change at all in AGA eludes me.

As for Willoe I guess biggest seller is 2MB of chipram and maybe new features available when combines with other custom chips replacement versions.

[nonarkitten]

Originally Posted by Promilus

It's easy to verify. There's something called PC Mag - that was pretty damn big magazine and released monthly. There are scans on google books. It takes one look to verify what kind of configurations WERE available BACK IN 1992/93.
You don't believe me? Here you go
https://books.google.pl/books?id=LIy...page&q&f=false
Page 165 and downwards might, just might enlighten some about what PC world actually had to offer back then when it comes to VGA card. And magazine itself is filled with ads of 486 based systems.
Who's the liar now?

1. This is 1993. In 1993, a 386SX20 was still a bit more than a 1200 (though slower) and the 486's were being advertised in the $2000-range. Remember, 20MHz 386SX is about the speed of a 20MHz 68000 or 10MHz 68020. A 20MHz 486 is about the speed of a 20MHz 68020 or 030.

2. Yes, a few cards had some acceleration. And outside of AutoCAD these weren't used and cost more than an Amiga alone. Everyone tried to make a standard (VLB/AF anyone) but it wasn't until cards adopted OpenGL or Microsoft's DirectX that we had this.

3. Since I know you love geometric mean, the ISA cards averaged 2.44MB/s and the VLB 7.09MB/s. This is still less than AGA and anyone who USED PC's in the day remembers how temperamental VLB was.

4. And audio, even by 1994, getting a PC with a Sound Blaster 16 was not common and not cheap. An SB16 and CDROM alone were, again, more than a whole Amiga and certainly more than the CD32.

Sure. There were lots of machines that could beat the Amiga, even in 1985. How much money are you willing to spend? A 4MB UNIX Workstation that could compete with the Amiga in the 80's didn't cost thousands of dollars, they cost tens of thousands of dollars, and for the PC at this point, they were still more than four times the price.

In 1993 you could pick up a 1200 for $379, add a GVP Turbo+ for $299 (40MHz 68EC030) and 4MB SIMM for $150 for 6MB total and drop in a 270MB 3.5" HDD with an adaptor cable for $259, and a 14" Commodore 1084S for $199 (Total: $1,296). All 4MB is useable by the OS all the time and there's nothing to configure. You have a great desktop UI and aside from some OCS/ECS compatibility issues, everything just works.

In 1993 you could pick up a 486SX33 for $2,200 with 4MB RAM and 210MB and VLB SVGA video supporting 1024x768x256 colours. You'd need another $349 to get a Sound Blaster 16 and kill 2-3 weeks of your life allocating the IRQ and DMA channels so it doesn't conflict with the random IO ports the clone PC decided to include. Oh and have fun with those games that needed every byte of 640KB, because squeezing out that last little TSR was painful.

Twice the price and it would have still been slower than the Amiga.

Okay, maybe comparing the "console" Amiga to a "desktop" PC is unfair, so lets compare an Amiga 4000.

In 1993 a top-of-the-line Amiga 4000-040/25 was $2299 and came with a 120MB HDD. You could pop-in 4MB for about $80 for 6MB total RAM. So with rounding that's about $2400.

Now, the 040 is about three times the speed of the 030 and if the 030 is about the speed of a 486, this means we need about a 75MHz 486. I cannot even find anything that fast in your magazine for 1993. Ignoring the fact that PC manufacturers were notorious at this time for running ads based on speculative availability and still bubkiss on the Pentiums. The 486DX2/66 was it and those seem to run around $3000 to $3000.

The Amiga was very competitive.

[nonarkitten]

Originally Posted by Promilus

That's why AAA (which was initially in-development around '88) was intended to have 32bit blitter already running faster and with much greater and faster chipram. Why this piece of hardware didn't change at all in AGA eludes me.

Commodore budgeted for AGA to have ONE new chip. ONE. We got Lisa and a slightly tweaked Agnus. It was a bad decision. Like so many other bad decisions Commodore made.

Originally Posted by Promilus

As for Willoe I guess biggest seller is 2MB of chipram and maybe new features available when combines with other custom chips replacement versions.

Instead of coming here and posting about how far Commodore was falling behind, how about let's talk about what we can do now.

The ECS 16-bit bus can support a 2x and 4x FPM/EDO mode; the chips were the bottleneck here. This opens a lot of possibilities that can even exceed AGA without adding "bonus features" that no one will support. Their choice to only make one new chip and nerf the rest was stupid; they should have stuck with 16-bit and gone all-in with FPM. Because adding 2x and 4x FPM to all three chips probably would have been cheaper, more compatible with ECS and a lot faster.

What could be done with Willoe alone:
- 16MB chip RAM (2MB addressable from 68K); this could allow Workbench, for example, to have it's screen not take any of the critical base 2MB
- 2x and 4x Blitter speed and Copper prefetch (non blocking)
- Note that the 3000 will use real chip RAM for the base 2MB and will need upgrading to 70ns FPM to support 2x and 4x AGA fetch modes

What could be done with Willoe and Xander (A500/A2000):
- up to 8MB of chip RAM addressable from 68K
- non-blocking CPU cycles for up-to double the CPU bandwidth

What could be done with Willoe and Faith (A500/A2000/A3000):
- AGA 2x and 4x display modes, double playfield and HAM8
- possible side-banding up to 50% of bitplane DMA freeing more cycles
- integrated scan doubler for VGA output

What could be done with Willoe and Harmonie (A500/A2000/A3000):
- Paula 2x and 4x audio and disk speeds
- Support for HD floppy speeds (60kbps)
- Support for either 112kHz 8-bit or 56kHz 16-bit audio

[kipper2k]

i like pretty well all features for the a500 and 16 bit machines. We are getting to the point where the A500 can never die