Was anyone else disappointed with the A1200?

[sandruzzo]

@Cyprian

That's why I've talked about more buses

[hammer]

Originally Posted by sandruzzo

@Cyprian

That's why I've talked about more buses

AGA needs math processing improvements e.g. 68EC020-25 and DSP3210.

For 1992 to Q3 1994 with a desktop computer use case and delivering a "new 32bit 2.5/3D gaming experience", the 486SX-25 / 68LC040 @ 25 Mhz compute level is sufficient.

The alternative is to battle SNES's strong 2D 16-bit gaming experience at a low price segment.

DSP3210 @ 50 Mhz (25 MFLOPS, FP32 only) is the $20 poor man's near 68060 @ 50Mhz (27.89 MFLOPS) alternative.

[sandruzzo]

@hammer

Have you ever been a professional software developer? We don't need MIPS per se, we need true actual easly performances..

[pixie]

Originally Posted by sandruzzo

@hammer

Have you ever been a professional software developer? We don't need MIPS per se, we need true actual easly performances..

Quake was brought to a 030 falcon through the use of DSP ...

[hammer]

Originally Posted by pixie

Quake was brought to a 030 falcon through the use of DSP ...

It wasn't PC's Quake since Falcon's DSP is a pure INT16/INT24 fixed point co-processor.

[pixie]

Originally Posted by hammer

It wasn't PC's Quake since Falcon's DSP is a pure INT16/INT24 fixed point co-processor.

Well, quake pc wasn't a pure "square root" quake as it only used "fast square root" algorithm... there's also that! :P

[Cyprian]

Originally Posted by hammer

https://www.amigawiki.org/doku.php?id=dearts:alice
Alice's shows 16-bit data pins i.e. DRD0 to DRD15. RGA has 8 pins. 32 bit Alice+ evolution wouldn't improve 2.5D/3D situation.

https://www.amigawiki.org/doku.php?id=dearts:lisa
Lisa shows 32-bit data pins i.e. D0 to D31. RGA has 8 pins.

mistyping, of course I meant the video chip - Lisa. It is 32bit to the memory bus, but from the CPU side it is still old poor 16bit.

[Bruce Abbott]

Originally Posted by pixie

Quake was brought to a 030 falcon through the use of DSP ...

It was?

[hammer]

Originally Posted by sandruzzo

@hammer
Have you ever been a professional software developer?

Yep.

Originally Posted by sandruzzo

We don't need MIPS per se, we need true actual easly performances.

[ Show youtube player ]
AA3000+ DSP3210 vs 68060 @ 50 Mhz on FP32 Mandelbrots.

DSP3210 supports the standard 32-bit integer and IEEE 32-bit floating point.

https://techmonitor.ai/technology/mo...0_next_quarter
In 1994,
68060 @ 50 Mhz has $263 (with FPU and MMU)
68LC060 @ 50 Mhz has $169. (no FPU)
68EC060 @ 50 Mhz has $150. (no FPU, no MMU)

http://archive.computerhistory.org/r...-05-01-acc.pdf
1994 CPU wholesale prices

80486DX2-50 has $231.75
Pentium-60 has $748.00.
R4000SC-50 has $546.75

68060 @ 50 Mhz had "power without the price" advantage but without economies of scale, the Amiga platform wasn't able to exploit this advantage.

68LC060's price was close to 386DX-25's 1992 price.


https://archive.computerhistory.org/...-05-01-acc.pdf
Page 119 of 981

For 1992
68000-12 = $5.5
68EC020-16 PQFP = $16.06,
68EC020-25 PQFP = $19.99,

68EC030-25 PQFP = $35.94
68030-25 CQFP = $108.75

68040-25 = $418.52
68EC040-25 = $112.50
---
Competition

AM386-40 = $102.50
386DX-25 PQFP = $103.00

486SX-20 PQFP = $157.75
486DX-33 = $376.75
486DX2-50 = $502.75


The X86 PC clone market competition kept the distribution channel's profit margin relatively low i.e. it minimized the "Phase 5" tax.

[hammer]

Originally Posted by pixie

Well, quake pc wasn't a pure "square root" quake as it only used "fast square root" algorithm... there's also that! :P

Are you forgetting PC Quake's FDIV usage? Quake FPU heavy workload culled many X86 cloners from the market.

Falcon DSP56 @ 33Mhz (16 MIPS INT24) "Quake" results weren't at frame-rate playable due to insufficient performance headroom.

My selection between CyberStorm 68060 @ 50 vs Pentium 150 includes performance headroom for smooth frame rates for similar spending budgets.

PS1's Quake is a refactored fixed-point integer port with performance headroom for good frame rates. PocketQuake is another fixed-point integer port for Windows CE ARM.
Windows CE ARM and PS1 support 32-bit integers.

[ Show youtube player ]
Falcon DSP has a playable (about 12 fps) frame-rate Doom port with a standard envelope screen size and low details.


Falcon DSP56K Quake port needs 14 MB RAM expansion.

https://content.invisioncic.com/r322...0644d42f09.png
Falcon with 14 MB RAM is 999 UKP.
Falcon with 4 MB RAM and 64 MB HDD is 699 UKP.
Falcon with 4 MB RAM is 599 UKP.
Falcon with 1 MB RAM is 399 UKP

DSP3210 @ 50 Mhz can do 12.5 MIPS INT32 and 25 MFLOPS FP32. DSP3210 supports standard INT32 and FP32 data formats like a normal 040/486 CPU. Atari's DSP56 selection wasn't good for general purpose.


https://dosdays.co.uk/topics/1993.php
Commodore DT486DX-25 is 760 UKP with 4 MB RAM, 52 MB HDD, MS-DOS 5.0, Win3.1, mouse, 14" color VGA monitor.

If my home country was the UK, I would have purchased a Commodore DT 486DX-25.

399 UKP is similar to $770 AUD. My budget is $1500 AUD range which is similar to $1000 USD. I'm a 799 UKP price range customer.

13000-14000 Falcons sold with an additional 4164 units in stock is the last official number, enough said. "16-bit" 68030 mockery didn't help Atari's case.

999 UKP is like $2000 in Australia.

I prefer the following
499 UKP A1200 with DSP3210 @ 50 Mhz, 68EC020-25 and 1 MB 32 bit Fast RAM.
599 UKP A1200 with DSP3210 @ 50 Mhz, 68EC020-25 and 2 MB 32 bit Fast RAM.
699 UKP A1200 with DSP3210 @ 56 Mhz, 68EC020-28 and 4 MB 32 bit Fast RAM.
Standard 3.5-inch HDDs.

A1200's dual 32-bit bus would be "64-bit" in Atari Jaguar marketing. LOL

[Cyprian]

Originally Posted by Bruce Abbott

It was?

I confirm, DML converted Quake to Atari Falcon and it uses DSP intensively.

[Cyprian]

Originally Posted by hammer

[ Show youtube player ]
AA3000+ DSP3210 vs 68060 @ 50 Mhz on FP32 Mandelbrots.

DSP3210 supports a 32-bit integer and IEEE 32-bit floating point.

DSP3210 is really nice but the downside is the RAM size. It has only 2kB 32bit words (8kbytes) of fast internal RAM, lack of byte/word access to the main RAM, and two CPU cycles per one instruction.
Big advantage is that it can run parallel to the main CPU.


68060 has a twice more fast internal RAM - 8kB instruction cache plus 8kB data cache and have byte/word access. It is superscalar (more than one instruction per clock cycle) and it has also faster main clock.

[Bruce Abbott]

Originally Posted by Cyprian

mistyping, of course I meant the video chip - Lisa. It is 32bit to the memory bus, but from the CPU side it is still old poor 16bit.

The only thing the CPU has to access in Lisa is its control registers. For compatibility most of them are the same as OCS/ECS. The benefit of making them 32 bit would be small.

You show an example of 'changing colour zero to the colour $123456', which the CPU wouldn't normally be doing in real time so the extra overhead of doing it 16 bits at a time is not important (it might be for the Copper, but that's another matter).

For compatibility the color registers are split into 8 banks of 32 colors, mapped to the 32 OCS color register addresses. When the upper 12 bits of a color register are written they are automatically copied to the lower 12 bits to produce the correct color. So unless you need an exact 24 bit color the CPU (or Copper) only needs to do one write per color register. There is also an 8 bit XOR color index mask register that enables swapping a bunch of color registers with a single write.

Personally I would rather have compatibility than some complicated register scheme just to get a tiny bit more speed. If there was one thing I would add it is giving the Copper the ability to load a register with a different value on every color clock. That could give you a 160x200 or 320x200 chunky pixel display with zero bitplanes!

[pixie]

Originally Posted by hammer

Falcon DSP has a playable (about 12 fps) frame-rate Doom port with a standard envelope screen size and low details.

Interestingly I've tried on similar settings that of a 50Mhz 68030 (1.38x of 608030 25mhz in sysinfo) on adoom in winuae, but on rtg, and it got me about the same fps.

(and even on aga the results aren't too shabby either)

[Bruce Abbott]

Originally Posted by Cyprian

I confirm, DML converted Quake to Atari Falcon and it uses DSP intensively.

Got a link to that?

The best video of DML's work I could find was this...

[ Show youtube player ]

...which isn't Quake.

[hammer]

Originally Posted by Cyprian

mistyping, of course I meant the video chip - Lisa. It is 32bit to the memory bus, but from the CPU side it is still old poor 16bit.

I prefer discrete 32-bit Fast RAM for the CPU/DSP32.

[hammer]

Originally Posted by pixie

Interestingly I've tried on similar settings that of a 50Mhz 68030 (1.38x of 608030 25mhz in sysinfo) on adoom in winuae, but on rtg, and it got me about the same fps.

DSP56 @ 32 Mhz delivers a similar MIPS magnitude as 68030 @ 50 Mhz. PC Doom requires 4 MB RAM.

Can't avoid the SuperFX2 addons method.

[hammer]

Originally Posted by Bruce Abbott

Got a link to that?

The best video of DML's work I could find was this...

[ Show youtube player ]

...which isn't Quake.

There's not enough performance headroom for the actual game, let alone Hollywood's 24 fps standard.

[hammer]

Originally Posted by Cyprian

DSP3210 is really nice but the downside is the RAM size. It has only 2kB 32bit words (8kbytes) of fast internal RAM, lack of byte/word access to the main RAM, and two CPU cycles per one instruction.
Big advantage is that it can run parallel to the main CPU.


68060 has a twice more fast internal RAM - 8kB instruction cache plus 8kB data cache and have byte/word access. It is superscalar (more than one instruction per clock cycle) and it has also faster main clock.

https://techmonitor.ai/technology/mo...0_next_quarter
In 1994,
68060 @ 50 Mhz has $263 (with FPU and MMU)
68LC060 @ 50 Mhz has $169. (no FPU)
68EC060 @ 50 Mhz has $150. (no FPU, no MMU)

It comes down to compute power per dollar.

Modern GPU's proposition is compute power per dollar and performance per watt.

For 68EC060's and 68LC060's wholesale prices are similar to 1992 386DX-25 prices, hence it should land about $1500 AUD range or 799 UKP range.

A500 has $699 USD in 1987 which is about $825 USD in 1992.

68EC060's $150 is very difficult for the CD32 price range.

The Amiga didn't have 68040 socket mass production like PC's 486 and Pentium sockets or Apple's 68040 socket economies of scale.

I can buy low-cost POWER8 CPUs, but the problem is the motherboard.

PS1's development meetings are dominated by price vs performance trade-off debates.

If Commodore was alive and healthy in 1994, the 68LC060+A1200 combo could have undercut the Pentium 60s.

[hammer]

Originally Posted by Cyprian

DSP3210 is really nice but the downside is the RAM size. It has only 2kB 32bit words (8kbytes) of fast internal RAM, lack of byte/word access to the main RAM, and two CPU cycles per one instruction.
Big advantage is that it can run parallel to the main CPU.

68060 has a twice more fast internal RAM - 8kB instruction cache plus 8kB data cache and have byte/word access. It is superscalar (more than one instruction per clock cycle) and it has also faster main clock.

https://ftp.fau.de/aminet/docs/misc/dsp3210info.txt
DSP3210 has 8KB on-chip memory

DSP3210's 8KB is closer to 68040's 4KB+4KB L1 cache.

DSP3210's supported datatypes are DSP32, IEEE FP32, INT16, INT32, 8bit unsigned, mu-law and A-law.

Good luck building 68LC060-based Amiga 1200 for 3DO's $699 target.