Thoughts on Sega's "clean" hardware design aspects . . . and the Saturn.

Quote:

---

Originally Posted by Stef

Actually the MAC instruction is really handy for matrix / vector multiplication and so for me it's part of the 3D transformation stuff.
The only missing instruction to get complete and powerful enough 3D transformation was a fast division...
A fast 3D transformation code will usually require 9 or 12 multiplications (depending your matrix) per vertex, so given the number of cycles, even taking the worst case (4 cycles) you can achieve about 600000 vertex transformed per second (10000 per frame at 60 fps rate). Ok that would eat all the SH2 time, and this is without any 2D projection (and the famous division)...
A real SIMD instruction would have be more powerful but at this time it was just out of question.

---

Yes, MAC is handy for 3D... when you don't have actual matrix operations. That doesn't mean it was specifically made for 3D, it just happens to have an instruction meant for something else entirely that makes it SLIGHTLY better at 3D than bog generic instructions. That was my point.

Anyway, your calculations on the vertex processing is part of why there were two SH2s - Sega expected that folks needing the best 3D throughput could have the other SH2 grinding through the vertexes in parallel to the other SH2. I need to break out my old Yeti3D example and see if I can't do some more experiments on that.

Chilly Willy, I've just watched your Yeti3D 32X demo and it's vey impressive.
Now I'm even more conviced that dropping the VDP1 in favor of a third SH2 in the Saturn would have helped a lot the console in its early days (at least in 3D). Devs could have chosen to render triangles if they wanted to.

Maybe a Saturn built with the same philosophy in mind than the 32X (except this one should never have been launched), that's to say everything done in software wouldn't have been dumb in theses early days of 3D when 3D thech was quickly evolving...

What do you think of a Saturn with one 14.3 mhz SH1 as central processor handling CD drive stuffs, sound contolling (I believe SH1 had embedded A/D and D/A converters) and some game logic, and 3 28.6 mhz SH2 doing everything else, plus a 32X Super VDP like and a single pool of 2 to 4 MD of SDRAM?

SH processors weren't very expensive and having 4 of them would have easily allowed later consolidation (as well as the single pol of ram)...

Quote:

---

Originally Posted by LastBronx

Chilly Willy, I've just watched your Yeti3D 32X demo and it's vey impressive.
Now I'm even more conviced that dropping the VDP1 in favor of a third SH2 in the Saturn would have helped a lot the console in its early days (at least in 3D). Devs could have chosen to render triangles if they wanted to.

Maybe a Saturn built with the same philosophy in mind than the 32X (except this one should never have been launched), that's to say everything done in software wouldn't have been dumb in theses early days of 3D when 3D thech was quickly evolving...

What do you think of a Saturn with one 14.3 mhz SH1 as central processor handling CD drive stuffs, sound contolling (I believe SH1 had embedded A/D and D/A converters) and some game logic, and 3 28.6 mhz SH2 doing everything else, plus a 32X Super VDP like and a single pool of 2 to 4 MD of SDRAM?

SH processors weren't very expensive and having 4 of them would have easily allowed later consolidation (as well as the single pol of ram)...

---

It wouldn't have been as fast at pushing polys. VDP1 might have had its problems, but it was no slouch at rendering quads. Good Saturn engines pushed the PSX for poly drawing power. The SH2 on its own wouldn't come close. It's easily an order of magnitude slower.

No, the VDP to get rid of would have been the 2D chip, VDP2, not VDP1. Also get rid of the SH1 and have the sound processor handle the CD as well as the sound. Make VDP2 just a simple frame buffer handler + palette, much like the 32X VDP.

The "ideal" Saturn that would have been easier for programmers and cheaper for Sega would have been:

One SH2 at 28.6MHZ with 2MB of SDRAM
VDP1 with 1MB VRAM and two 256KB frame buffers
68000 with 512KB of DRAM and sound chip, and CD controller (maybe with extra DRAM for CD buffering)
VDP2 changed to only fetch and display frame buffer, with palette ram
SCU DSP for handling 3D math operations (better docs and tools or a library so devs could actually use it)

As it was, most devs ignored the extra SH2, so it went to waste, and used VDP2 as little as possible. As it was, the SH1 controlling the CD was an utter waste of money. The SCU DSP was a good idea, but too little docs and no real example code meant many devs ignored it as well. Splitting the system ram into 1MB of high ram and 1MB of low ram made memory management harder than it needed to be, and I doubt it saved much if any money. Sure, SDRAM was more expensive, but spiltting your ram into two banks made the memory interface and PCB more complex and expensive as a tradeoff. I would bet decent money it wound up MORE expensive than one block of 2MB of SDRAM, which could be directly handled by the SH2 alone without any extra hardware.

Same kind of thing with the VRAM - the Saturn has 512KB of VRAM for VDP1 (along with two banks of 256KB for the frame buffer), and 512KB of VRAM for VDP2. Having one bank of VRAM for VDP1 and making VDP2 just a display handler would have probably made things much cheaper, and easier to handle. It would have been very similar to most GPU cards, especially as a few at the time also did quads. It hadn't quite been settled at that time whether GPUs would go with triangles or quads. That was pretty much settled by the NEXT generation of consoles and PC video cards.

Ok, thank you.
With your setup, don't you think the Saturn would have lacked a bit of polygonal capabilities compared to the 66 mips Playstation GPE with no second SH2 and just the 14.3 mhz SCU DSP (one cycle MAC so 14 mips?) to help?
Even at 28.6 mhz, the DSP would have been 2x less powerful than the GPE wouldn't it?
Also, the actual VDP1 had a 16 bit data bus; with your Saturn and its simpified VDP2, wouldn't a 32 bit (or even 64 bit) maybe with a 2x latch (that kitty kool89 talked in another thread) have been better?
I also think support for hardware bilinear filtering would have killed the Playstation as it would have allowed les pixellated graphics...

Quote:

---

Originally Posted by LastBronx

Ok, thank you.
With your setup, don't you think the Saturn would have lacked a bit of polygonal capabilities compared to the 66 mips Playstation GPE with no second SH2 and just the 14.3 mhz SCU DSP (one cycle MAC so 14 mips?) to help?
Even at 28.6 mhz, the DSP would have been 2x less powerful than the GPE wouldn't it?
Also, the actual VDP1 had a 16 bit data bus; with your Saturn and its simpified VDP2, wouldn't a 32 bit (or even 64 bit) maybe with a 2x latch (that kitty kool89 talked in another thread) have been better?
I also think support for hardware bilinear filtering would have killed the Playstation as it would have allowed les pixellated graphics...

---

100,000 quads are equal to 200,000 tri polygons. I agree with chilly except with the 68K, the SH1 was about on par. How about letting any CPU access the CD controller, but recommending the the SH1 for that purpose, the only 1 CPU (1 master at 28.6 MHZ, 1 slave at 20 mhz), and the un-buffed VDP1 (buffed VDP1 + Stripped down VDP2 on 1 chip or 2 separate chips) and DSP (@ 28.6 mhz). Also, more polygonal power > better than bilinear filtering. Everything else is good.
EDIT: Oh, and a PCM only audio solution, because the SCSP is overkill for a CD based system.

Quote:

---

Originally Posted by alextheman

100,000 quads are equal to 200,000 tri polygons. I agree with chilly except with the 68K, the SH1 was about on par. How about letting any CPU access the CD controller, but recommending the the SH1 for that purpose, the only 1 CPU (1 master at 28.6 MHZ, 1 slave at 20 mhz), and the un-buffed VDP1 (buffed VDP1 + Stripped down VDP2 on 1 chip or 2 separate chips) and DSP (@ 28.6 mhz). Also, more polygonal power > better than bilinear filtering. Everything else is good.
EDIT: Oh, and a PCM only audio solution, because the SCSP is overkill for a CD based system.

---

You stated that the second SH2 was a waste and now you want a SH1 instead...

Also, clean filtered textures (thanks to bilinear filtering) > more polygonal power.

Quote:

---

Originally Posted by LastBronx

You stated that the second SH2 was a waste and now you want a SH1 instead...

Also, clean filtered textures (thanks to bilinear filtering) > more polygonal power.

---

Mainly because the SH2 can't be clock doubled without halving the bus speed. I think that most people didn't care (about bilinear filtering), and the PS1 owners were willing to ignore pixelated and warping textures, I agree with your statement.

Quote:

---

Originally Posted by alextheman

Mainly because the SH2 can't be clock doubled without halving the bus speed. I think that most people didn't care (about bilinear filtering), and the PS1 owners were willing to ignore pixelated and warping textures, I agree with your statement.

---

Maybe PS1 owners didn't care about bilinear filtering because Saturn didn't offer it and most N64 games looked blurry when in use (but not because of bilinear filtering but rather because of limited cart storage capacity resulting in smaller textures). See how Sega Model 2 games and Namco System 22 ones looked clean and beautiful back then (ok not only thanks to texture filtering but also thanks to higher resolution textures and display...). Could have been a good feature for Saturn imho...

I still wonder why they didn't license Lockeed Martin's technology to include a scaled back version of the Model 2 chip in the Saturn, knowing they were working together on the arcade board in the beginning of 1993.

Also, you were talking about a PCM chip instead of the SCSP and I agree with you. Furthermore, Sega already had a custom 28 channel PCM chip in System Multi32 and Model 1 boards (the latter has actually 2 of this chip : one for music and one for sound effects)...

Quote:

---

Originally Posted by LastBronx

Maybe PS1 owners didn't care about bilinear filtering because Saturn didn't offer it and most N64 games looked blurry when in use (but not because of bilinear filtering but rather because of limited cart storage capacity resulting in smaller textures). See how Sega Model 2 games and Namco System 22 ones looked clean and beautiful back then (ok not only thanks to texture filtering but also thanks to higher resolution textures and display...). Could have been a good feature for Saturn imho...

I still wonder why they didn't license Lockeed Martin's technology to include a scaled back version of the Model 2 chip in the Saturn, knowing they were working together on the arcade board in the beginning of 1993.

Also, you were talking about a PCM chip instead of the SCSP and I agree with you. Furthermore, Sega already had a custom 28 channel PCM chip in System Multi32 and Model 1 boards (the latter has actually 2 of this chip : one for music and one for sound effects)...

---

I agree with most of what you say. Lockheed Martin used the technology they used in the Model 1 & 2 originally in military training simulators, and Sega had to do some work scaling down the Model 1's price, and even then it was pretty expensive. Also, the arcade hardware in the Model 2 cost quite a bit of money, and combined with Sega's profit margin, couldn't be put to use in home consoles in the 300 to 500 range without sega losing some money in the process.

The SCU DSP can use DMA to fetch and store vertex data, while the GTE needs to use the CPU to load/store the data. VDP1 was about as fast at drawing (in some modes at least) as the GPU. Given those two points, I don't think that the Saturn would have suffered any loss of polygon drawing power for only having a single SH2 as long as Sega made good SCU DSP support code available.

Quote:

---

Mainly because the SH2 can't be clock doubled without halving the bus speed.

---

You still don't understand the clock situation. The SH1 is inferior to the SH2 in EVERY respect. You kept mentioning using clock doubling inside the CPU as a means to go faster than the maximum clock rate, and that's where the other guy said you would need to cut the clock in half to keep inside the max clock rate. It would be EXACTLY the same for the SH1. That 28.6 MHz max (or 21 MHz max for the SH1) is INVIOLATE, no matter how it's generated, be it a direct 28.6 MHz clock, or doubling a 14.3 MHz clock. You cannot give it a 28.6 MHz clock and then use the clock doubling mode as the final clock is then higher than 28.6 MHz. It would be the same for the SH1 - you could not feed it a 21 MHz clock and use the clock doubling mode as the final clock would be higher than 21 MHz.

Using an SH2 in single-clock mode with a 28.6 MHz clock is as good as you will get, and much better than an SH1 could ever be... 28.6/21 * 100 percent better, to be precise. :D

Quote:

---

Originally Posted by Chilly Willy

The SCU DSP can use DMA to fetch and store vertex data, while the GTE needs to use the CPU to load/store the data. VDP1 was about as fast at drawing (in some modes at least) as the GPU. Given those two points, I don't think that the Saturn would have suffered any loss of polygon drawing power for only having a single SH2 as long as Sega made good SCU DSP support code available.



You still don't understand the clock situation. The SH1 is inferior to the SH2 in EVERY respect. You kept mentioning using clock doubling inside the CPU as a means to go faster than the maximum clock rate, and that's where the other guy said you would need to cut the clock in half to keep inside the max clock rate. It would be EXACTLY the same for the SH1. That 28.6 MHz max (or 21 MHz max for the SH1) is INVIOLATE, no matter how it's generated, be it a direct 28.6 MHz clock, or doubling a 14.3 MHz clock. You cannot give it a 28.6 MHz clock and then use the clock doubling mode as the final clock is then higher than 28.6 MHz. It would be the same for the SH1 - you could not feed it a 21 MHz clock and use the clock doubling mode as the final clock would be higher than 21 MHz.

Using an SH2 in single-clock mode with a 28.6 MHz clock is as good as you will get, and much better than an SH1 could ever be... 28.6/21 * 100 percent better, to be precise. :D

---

I know you can't, and the SH1 is on par with the Motorola 68K.

Quote:

---

Originally Posted by Chilly Willy

The SCU DSP can use DMA to fetch and store vertex data, while the GTE needs to use the CPU to load/store the data. VDP1 was about as fast at drawing (in some modes at least) as the GPU. Given those two points, I don't think that the Saturn would have suffered any loss of polygon drawing power for only having a single SH2 as long as Sega made good SCU DSP support code available.

---

Thank you for the DSP explanation. I reckon VDP1 was about as fast at drawing as the Playstation's GPU in certain cases but it would haven't been enough in the case you drop the VDP2 for a simpler display controller (like you mentionned in your ideal Saturn setup) as it was very useful not only for 2D but also for 3D. Without VDP2, Virtua Fighter 2, Last Bronx or Radiant Silvergun would have been very inferior. To compensate the loss of the VDP2, you would have needed to at least double VDP1 bandwidth (either by doubling its clock speed or make its bus 32 bit) to simulate VDP2 backgrounds...

Quote:

---

Originally Posted by alextheman

I know you can't, and the SH1 is on par with the Motorola 68K.

---

What do you mean when you write "the SH1 is on par with the Motorola 68K"?
Performance wise, a 16.67 mhz 68K is rated for less than 3 mips and has no cache; at 20 mhz, a SH1 is 20 mips (15-16 mips real world I think), plus its has a 2 cycle MAC (for 16 bit calculations), 4 to 8 KB scratchpad ram (the one with 4 KB ram also has 4 KB of rom)... I don't think tey're in the same league at all.

Quote:

---

Originally Posted by LastBronx

What do you mean when you write "the SH1 is on par with the Motorola 68K"?
Performance wise, a 16.67 mhz 68K is rated for less than 3 mips and has no cache; at 20 mhz, a SH1 is 20 mips (15-16 mips real world I think), plus its has a 2 cycle MAC (for 16 bit calculations), 4 to 8 KB scratchpad ram (the one with 4 KB ram also has 4 KB of rom)... I don't think tey're in the same league at all.

---

I meant at least on par. A 68k at the same speed as a SH1 can only get 1/4 of the performance.

The SH1/2 would be kinda like a 68030/80386 with larger caches and slightly faster execution units (more efficient per clock). Not quite as good as either the 68040/80486, but a hell of a lot cheaper.