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

Quote:

---

Originally Posted by Stef

Too bad the scu dsp does not have fast division... and indeed at 14.3 Mhz it's not really faster than the SH2 even for the specialized MAC instruction.
Honestly the saturn design is such a mess, such a waste in many parts :-/

---

I'm not really sure the dsp hadn't a hardware divisor; I'm just pretty sure about its one cycle MAC though. Futhermore, the dsp was doing fixed point computations (16 bit) which is sufficient for triangles but not so convenient for quads.

I was watching a video of 32X game Metal Head recently; it's not a very good game and it was rushed but it shows that dual SH2 can do pretty good texture mapped 3D. Of course, it's nowhere in the league of what Saturn and Playstation could do but the 32X had a very constraint 16 bit wide 256 KB sdram bank and pretty much nothing else to help (its "Super VDP" was just a bitmap displayer with a fill function).

That's why I talked about replacing the VDP1 by a third SH2 : maybe nearly as powerful but much more flexible. Sure it was difficult for developpers to get the max out of 2 SH2 (and consequently 3 ones) but some managed so it was up to Sega to provide good tools...

My dream saturn would be
Sega CD based CD controller
Single SH2 clock-doubled
Single VDP (VDP1 + Framebuffer)
Faster DSP (@ 28.6 MHZ.)
and a cheaper PCM only audio solution.
Anyone agree with me?

Quote:

---

Originally Posted by LastBronx

The SH2 had 3d specialized instructions and a 4 KB cache (good for texture mapping knowing that this cache can be split in 2 KB cache and 2 KB scratchpad ram).

---

Actually, I can't think of any SH2 instructions that are particularly specialized for 3D. The primary usage for the SuperH before Sega put them in consoles and arcades were as motor controllers, and audio filters. It has a fast multiplier, and multiply-and-accumulate instructions, which are what you need for digital filters.

The main reasons Sega used them were: fixed 16-bit instruction width meant it kept rom sizes small; high speed risc core meant high instruction throughput; and it had a pretty damn big/good cache for the price. It was like using a 486 at a much lower price. It didn't have many of the kernel features the 486 had that you would want for a computer, but it was ideal for consoles where you didn't need things like kernel level operations and memory protection.

Sorry, I should have written multimedia instructions?
Do you think it was good (for the time) at texture mapping?

Quote:

---

Originally Posted by alextheman

My dream saturn would be
Sega CD based CD controller
Single SH2 clock-doubled
Single VDP (VDP1 + Framebuffer)
Faster DSP (@ 28.6 MHZ.)
and a cheaper PCM only audio solution.
Anyone agree with me?

---

Not bad but definitively : a double clocked SH2 (57.2 mhz) was not possible in 1993/94; 28.6 mhz was the fastest speed available at this time for the SH2.
At first, I also thought the dual SH2 setup was shitty but it was not so. Getting the max out of it was difficult but not impossible and was nearly achived very early in Saturn life cycle (see Sega Rally or Virtua Fighter 2. Even Virtua Fighter Remix and Panzer Dragoon made good use of both the SH2). Sega should have provided their tools (even scaled back) to third parties to help them to program the 2 SH2.

The SH2 was a very good cpu for its time and is almost the only good piece in the Saturn. In 1993, there were not numerous options for Sega for cpu to choose. A MIPS R4000 based cpu would have been more expensive than 2 SH2 (even 3 of them I think) and not more powerful. A MIPS R3000 based one would have been too weak (remember than the Playstation R3000A has on-die embedded coprocessos like the 66 mips gpe that give it its polygonal capabilities).

What Saturn needed is a faster dsp. At 57.2 mhz instead of 14.3 mghz, the Saturn SCU one would have been very powerful. And Sega could have kept both the SH2. With such a dsp, developpers wouldn't have had to rely such on the second SH2 so that it could have been used for sound controlling and processing (along with a simple 16 bit DAC) by lazy programmers and for further 3d effects by good devs...

There were also too many buses in the Saturn and too many memory pools of memory of different types. That's why Saturn was so expensive to produce and was to remain so during all its life cycle because it was almost impossible to consolidate. One unique pool of 2 to 4 MB of sdram on a 64 bit bus (with why not a 2x speed latch mechanism) would have been very good. It wouldn't have been a problem for the 2 SH2 thanks to their 4 KB cache (as it wasn't a problem for the R4300i in the N64). The VDP1 had to be tweaked to create a framebuffer in this pool. It would have been more flexible as you could choose the framebuffer size or it to be double buffered or not on demand. The VDP1 also desperately needed to be able to render triangles. Allowing quads (mainly for 2D) and triangles (for 3D) would have been a dream. As they were working on the Model 2 board with Lockeed Martin by this time, they could have tried to include functionalities that even the Playstation lacked : perspective correction (even partial) texture mapping and bilinear filtering of textures. To summerize, my dream saturn would be

- Sega CD based CD controller and dram (Motorola 68000 and 128 KB of cheap dram).
- Dual SH2 setup (same as actual Saturn).
- Single VDP : VDP1 on a 64 bit bus (128 bit 2x latch?). banwidth : 228.8 (457.6 with 2x latch) MB/s against 135.2 MB/s for the Playstation (and more than 550 MB/s for the N64).
2 KB cache ram for texture lookup tables.
2 to 4 KB cache ram for texture buffers.
2 to 4 KB cache rom of generic textures.
Partial perspective correct texture mapping (subdivided texture mapping).
Bilinear filtering.
RGB and YCbCr (for Gouraud shading) colors.
- Simpler SCU : one main bus
Really fast dsp (57.2 mhz).
Embedded 16 bit stereo DAC in the SCU.

All of this would have been nothing without good dev tools, clever marketing and solid managing. Sega should have launched the Saturn on september 1995 in USA like originally planned. I don't think they needed to sell it at a 299 dollars price point. IMHO, early adopters don't really care about price but they do really care about value for the price. A399 dollars Saturn with 2 package games (let's say Bug and Virtua Fighter Remix) would have offset the Playstation 299 dollars price point for one year until N64 launch and prevent Sega to lose too much money on each sold console.

Quote:

---

Originally Posted by LastBronx

Not bad but definitively : a double clocked SH2 (57.2 mhz) was not possible in 1993/94; 28.6 mhz was the fastest speed available at this time for the SH2.
At first, I also thought the dual SH2 setup was shitty but it was not so. Getting the max out of it was difficult but not impossible and was nearly achived very early in Saturn life cycle (see Sega Rally or Virtua Fighter 2. Even Virtua Fighter Remix and Panzer Dragoon made good use of both the SH2). Sega should have provided their tools (even scaled back) to third parties to help them to program the 2 SH2.

The SH2 was a very good cpu for its time and is almost the only good piece in the Saturn. In 1993, there were not numerous options for Sega for cpu to choose. A MIPS R4000 based cpu would have been more expensive than 2 SH2 (even 3 of them I think) and not more powerful. A MIPS R3000 based one would have been too weak (remember than the Playstation R3000A has on-die embedded coprocessos like the 66 mips gpe that give it its polygonal capabilities).

What Saturn needed is a faster dsp. At 57.2 mhz instead of 14.3 mghz, the Saturn SCU one would have been very powerful. And Sega could have kept both the SH2. With such a dsp, developpers wouldn't have had to rely such on the second SH2 so that it could have been used for sound controlling and processing (along with a simple 16 bit DAC) by lazy programmers and for further 3d effects by good devs...

There were also too many buses in the Saturn and too many memory pools of memory of different types. That's why Saturn was so expensive to produce ans was to remain so during all its life cycle because it was almost impossible to consolidate. One unique pool of 2 to 4 MB of sdram on a 64 bit bus (with why not a 2x speed latch mechanism) would have been very good. It wouldn't have been a problem for the 2 SH2 thanks to their 4 KB cache (as it wasn't a problem for the R4300i in the N64). The VDP1 had to be tweaked to create a framebuffer in this pool. It would have been more flexible as you could choose the framebuffer size or it to be double buffered or not on demand. The VDP1 also desperately needed to be able to render triangles. Allowing quads (mainly for 2D) and triangles (for 3D) would have been a dream. As they were working on the Model 2 board with Lockeed Martin by this time, they could have tried to include functionalities that even the Playstation lacked : perspective correction (even partial) texture mapping and bilinear filtering of textures. To summerize, my dream saturn would be

- Sega CD based CD controller and dram (Motorola 68000 and 128 KB of cheap dram).
- Dual SH2 setup (same as actual Saturn).
- Single VDP : VDP1 on a 64 bit bus (128 bit 2x latch?). banwidth : 228.8 (457.6 with 2x latch) MB/s against 135.2 MB/s for the Playstation (and more than 550 MB/s for the N64).
2 KB cache ram for texture lookup tables.
2 to 4 KB cache ram for texture buffers.
2 to 4 KB cache rom of generic textures.
Partial perspective correct texture mapping (subdivided texture mapping).
Bilinear filtering.
RGB and YCbCr (for Gouraud shading) colors.
- Simpler SCU : one main bus
Really fast dsp (57.2 mhz).
Embedded 16 bit stereo DAC in the SCU.

All of this would have been nothing without good dev tools, clever marketing and solid managing. Sega should have launched the Saturn on september 1995 in USA like originally planned. I don't think they needed to sell it at a 299 dollars price point. IMHO, early adopters don't really care about price but they do really care about value for the price. A399 dollars Saturn with 2 package games (let's say Bug and Virtua Fighter Remix) would have offset the Playstation 299 dollars price point for one year until N64 launch and prevent Sega to lose too much money on each sold console).

---

Clock doubling means a multiplier of 2. :daze: The 486DX2, released in 1992, had a clock multiplier of 2. Also, DSPs clocked faster than the system are a big nono. Though your plans for launch are pretty good.

Quote:

---

Originally Posted by alextheman

Clock doubling means a multiplier of 2. :daze: The 486DX2, released in 1992, had a clock multiplier of 2. Also, DSPs clocked faster than the system are a big nono. Though your plans for launch are pretty good.

---

Sorry but I don't understand what you mean. The 486DX2 having a clock multiplier of 2 only means its core runs at 2 times its bus speed. The Hitachi SH2 runs at the same speed as its bus, that's to say 28.6 mhz max (in 1993). If it had a clock multiplier of 2, that would mean its bus would run at 14.3 mhz max as its max speed was 28.6 mhz in 1993/94. So less powerful than the actual SH2 and much less powerful than 2 SH2s even in master/slave configuration...

Also, why shouldn't the dsp run at 57.2 mhz, 2x the system bus speed. The SH2s in the 32X ran at 23 mhz, 3 times the Genesis Motorola 68000 7.67 mhz frequency. If it was a matter of master clock, then... Put the Saturn master clock at 57.2 mhz (16x ntsc).

Quote:

---

Originally Posted by LastBronx

Sorry but I don't understand what you mean. The 486DX2 having a clock multiplier of 2 only means its core runs at 2 times its bus speed. The Hitachi SH2 runs at the same speed as its bus, that's to say 28.6 mhz max (in 1993). If it had a clock multiplier of 2, that would mean its bus would run at 14.3 mhz max as its max speed was 28.6 mhz in 1993/94. So less powerful than the actual SH2 and much less powerful than 2 SH2s even in master/slave configuration...

Also, why shouldn't the dsp run at 57.2 mhz, 2x the system bus speed. The SH2s in the 32X ran at 23 mhz, 3 times the Genesis Motorola 68000 7.67 mhz frequency. If it was a matter of master clock, then... Put the Saturn master clock at 57.2 mhz (16x ntsc).

---

How does the cpu clock magically decrease when you have a clock multiplier? :daze: Also, that is different, as the 32X 2 SH2's were its main cpu, and the 68k was the side processor.

Quote:

---

Originally Posted by alextheman

How does the cpu clock magically decrease when you have a clock multiplier? :daze: Also, that is different, as the 32X 2 SH2's were its main cpu, and the 68k was the side processor.

---

It seems you don't know what a clock multiplier is. It's simple : in 1993 the SH2 had a max speed of 28.6 mhz and it ran at the same speed as its bus (clock multiplier of 1 if you will). If it had had a clock multiplier of 2, that would have mean it would have run at 57.2 mhz (2 times its bus speed) and that was impossible because its max speed was 28.6 mhz. The only way for the SH2 to reach 28.6 mhz with a clock multiplier of 2 was its bus running at 14.3 mhz, so less powerful than one actual SH2.

And I still don't get why the dsp couldn't run at 57.2 mhz. The gpe in the Playstation ran at 66 mhz, 2x the R3000A speed and they were on the same die. But we're going to say I'm wrong and let's say 28.6 mhz for the dsp...

Quote:

---

Originally Posted by alextheman

Also, that is different, as the 32X 2 SH2's were its main cpu, and the 68k was the side processor.

---

The SVP in the Virtua Racing cart ran at 23 mhz (some say it was only doing 12-14 mips so maybe rather 15.3 mhz) and I don't think it was the main cpu there and the 68000 a coprocessor...

Quote:

---

Originally Posted by LastBronx

Sorry, I should have written multimedia instructions?
Do you think it was good (for the time) at texture mapping?

---

The SuperH didn't have any multimedia instructions until late in the SH3 line. Hitachi worked some of those back into the SH2 line, but that was more than a decade after the 32X/Saturn. As mentioned, at the time Sega was buying SH2s, the only thing they had specific to signal processing was the multiply-and-accumulate.

This goes back to the question of clock rate - some people clearly have looked at the latest catalogue of SuperH processors and noticed 50MHz SH2s with multimedia instruction extensions and assumed those were available at the time. They were not. Maximum frequency of 28.6 MHz, and no multimedia instructions. Period.

Ok thank you, no multimedia instructions. I knew its maximum clock speed was 28.6 mhz...

Don't get me wrong, it's an awesome little processor, especially for the price. It's just wasn't MEGA awesome. ;) I like programming on the 32X precisely because of the SH2s. It's amazing what you can do with them.

Quote:

---

Originally Posted by alextheman

My dream saturn would be
Sega CD based CD controller
Single SH2 clock-doubled
Single VDP (VDP1 + Framebuffer)
Faster DSP (@ 28.6 MHZ.)
and a cheaper PCM only audio solution.
Anyone agree with me?

---

My dream Sega Saturn would be:
Model 2 at home for 400 dollars, Christmas 1995 :mrgreen:.

Quote:

---

Originally Posted by Chilly Willy

Actually, I can't think of any SH2 instructions that are particularly specialized for 3D. The primary usage for the SuperH before Sega put them in consoles and arcades were as motor controllers, and audio filters. It has a fast multiplier, and multiply-and-accumulate instructions, which are what you need for digital filters.

---

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.