Nokia to reveal Ngage 2 at E3!

[TEXAN]

Hey fafalada, do you know the real world in-game sustained polygon count figure for the PSP?

IGN said that it was 3 million. Is that true?

 

[Lazy8s]

The geometry levels of PSP titles and Dreamcast titles look very similar, so 3M-tri/sec as an estimate of sustained game performance capability for PSP seems good.

It's unclear what the OMAP2420 can sustain overall versus the ratings for just the HR-S, the R-S, and their different clock speeds. The specifications of the various products rate something like:

200-MHz.......HR-S w/VGP.........(90-nm).......870K-gates.......6.25M-tri/sec.......625M-pix/sec
200-MHz.......HR-S....................(90-nm)......660K-gates........6.25M-tri/sec.......625M-pix/sec
120-MHz.......HR-S w/VGP.......(130-nm)......870K-gates........3.75M-tri/sec.......375M-pix/sec
120-MHz.......HR-S..................(130-nm).......660K-gates.......3.75M-tri/sec.......375M-pix/sec
..80-MHz.......HR-S w/VGP.......(180-nm)......870K-gates.........2.5M-tri/sec.......250M-pix/sec
..80-MHz.......HR-S..................(180-nm)......660K-gates.........2.5M-tri/sec.......250M-pix/sec

120-MHz.........R-S w/VGP.......(130-nm)......470K-gates.........1.5M-tri/sec.......240M-pix/sec
120-MHz.........R-S..................(130-nm)......365K-gates..........1.5M-tri/sec.......240M-pix/sec
..80-MHz.........R-S w/VGP.......(180-nm)......470K-gates............1M-tri/sec.......160M-pix/sec
..80-MHz.........R-S..................(180-nm)......365K-gates.............1M-tri/sec.......160M-pix/sec

 

[Lazy8s]

Could the 90-nm MBX core in the OMAP2420 be an R-S w/VGP clocked at about 220-MHz? That would deliver that 'over 2.5M-tri/sec' (2.75 actually) and give it 440M-pix/sec fillrate.

 

[PC-Engine]

It's now clear that Fox5 was making the point that the DC might still become transform limited in the scenarios where too much CPU power was being devoted to other tasks beside simple transform. We hadn't been arguing that.

Yes, that's why I was wondering what Fox's point was seeing as it's very obvious that in-game it's very likely that the limitation would be triangle or transform limited. It's similar to the PS2 situation except totally opposite where on the DC the raw transform capability is higher than the triangle setup (10M>7M) while on PS2 the triangle setup is higher than the raw transform (66M<75M).

BTW are the fillrates listed above raw or is it factoring in overdraw? IIRC the HR-S has two pixel pipes.

 

[Lazy8s]

I'm not sure what level of efficiency is assumed for either the triangle or pixel rates. I suspect the triangle performance isn't quite peak, and the pixel speed isn't quite raw. I just derived the stats from indescriptive documentation.

If the pixel fillrate is meant to be an IMR-comparative figure, I'd say it's selling itself a little short. An MBX-based system would handle games of roughly a Quake3-level of complexity, and that game tested to have an average overdraw above 3.

 

[Lazy8s]

On to Renesas.

Specifications for Sega Sammy Mobile Gaming Platform:
SuperH Full Compatibility

90-nm System-on-a-Chip SH3707

300-MHz Central Core Digital Signal Processor SH4AL-DSP
........2.1-GFLOPS
........540-MIPS

8-Channel PCM/ADCPM Audio

MPEG-1, -2, and -4 Video

200-MHz 2D/3D Graphics Accelerator Core with SIMD Co-Processor MBX with VGP
........Vertex Shader 1.0+ Programmability with Skinning
........Curved Surface Processing with Fractional Tesselation and Support for Differing Levels-of-Detail on Neighboring Patch Edges
........32-Bit Blending and Z-Test Precision Independent from Buffer Depth
........Supersampling Full-Scene Anti-Aliasing Standard
........1024x768 Image Resolution Maximum
........DOT3 Per-Pixel Lighting Support
........Anisotropic Texture Filtering Support
........6M-tri/sec Sustained

 

[TEXAN]

Was the SH3707 made exclusively for Sega Sammy?

I wonder what their plans are?

 

[TEXAN]

The geometry levels of PSP titles and Dreamcast titles look very similar, so 3M-tri/sec as an estimate of sustained game performance capability for PSP seems good.

It's unclear what the OMAP2420 can sustain overall versus the ratings for just the HR-S, the R-S, and their different clock speeds. The specifications of the various products rate something like:

200-MHz.......HR-S w/VGP.........(90-nm).......870K-gates.......6.25M-tri/sec.......625M-pix/sec
200-MHz.......HR-S....................(90-nm)......660K-gates........6.25M-tri/sec.......625M-pix/sec
120-MHz.......HR-S w/VGP.......(130-nm)......870K-gates........3.75M-tri/sec.......375M-pix/sec
120-MHz.......HR-S..................(130-nm).......660K-gates.......3.75M-tri/sec.......375M-pix/sec
..80-MHz.......HR-S w/VGP.......(180-nm)......870K-gates.........2.5M-tri/sec.......250M-pix/sec
..80-MHz.......HR-S..................(180-nm)......660K-gates.........2.5M-tri/sec.......250M-pix/sec

120-MHz.........R-S w/VGP.......(130-nm)......470K-gates.........1.5M-tri/sec.......240M-pix/sec
120-MHz.........R-S..................(130-nm)......365K-gates..........1.5M-tri/sec.......240M-pix/sec
..80-MHz.........R-S w/VGP.......(180-nm)......470K-gates............1M-tri/sec.......160M-pix/sec
..80-MHz.........R-S..................(180-nm)......365K-gates.............1M-tri/sec.......160M-pix/sec

Thats exactly what I was looking for, thanks Lazy!

 

[TEXAN]

Could the 90-nm MBX core in the OMAP2420 be an R-S w/VGP

Simon confirmed that it's the HR-S in this thread -

http://www.beyond3d.com/forum/viewtopic.php?t=20500&start=20

SiBoy wrote:

The MBX in the OMAP2420 (silicon was shown at ISSCC) was in 90nm and clocked at 200 MHz. Area is around 10mm^2.

TEXAN wrote:

Was that the MBX or MBX Lite?

Because TI have licensed both.

Simon F

Given the date of the following press release, I'm sure you could deduce the answer: (....unless they can spin chips in super duper record time!!!!)

Do you think TI are using the .18 figure by mistake? because if you go to PVR's site they only give numbers for the 0.18 models. That's what TI are wrongly using?

 

[Kristof]

Do you think TI are using the .18 figure by mistake? because if you go to PVR's site they only give numbers for the 0.18 models. That's what TI are wrongly using?

I am confused here but are you linking silicon process to performance ? The same chip can be created at different processes and thus result in different area usage but basic performance characteristics will not change, clockspeeds on the other hand...

K-

 

[Lazy8s]

The fact that Texas Instruments only more recently licensed MBX Lite doesn't necessarily suggest that OMAP2420 is based on HR-S, actually. What it seems to suggest is that R-S and HR-S are both non-Lite MBX, one implementation a high powered MBX and the other a low powered MBX.

The only way I can reconcile OMAP2420's "more than 2.5M-tri/sec" claim with ARM's performance guidelines is to figure that the 2420's MBX part is only the R-S, unless the MBX core is clocked really, really low comparatively (which is certainly possible). Then, there's always the matter of how the numbers are being reported, too.

Off topic: I think MBXs that are not of the Lite variety are just termed under the Pro label more specifically to avoid confusion. And there's a Lite and Pro version of the VGP to respectively accompany the two types of MBX cores.

The reason that there's both an HR-S line and R-S line made of the MBX Pro seems to be for targeting a wider range of markets. The R-S line appears to use only about half the gate count, half the embedded RAM, and two-thirds of the battery power.

 

[Lazy8s]

...Err, upon more review, I'm again thinking HR-S is MBX Pro and R-S is MBX Lite. Which begs the issue why OMAP2420's polygon throughput is so low or why it's at least claimed to be so low. Maybe the HR-S is only clocked around 55-MHz even though the rest of the chip is much higher, like the SH-Mobile3 SH73182.

 

[TEXAN]

Lazy, thats the exact same thing that I've been trying to find out to no avail.

 

[Kristof]

Sigh, let me help you guys out...

Download OpenGL ES Development for OMAP 2 under Linux - Tom Olson, TI from :

http://www.khronos.org/devu/library/

And then check page 19 of the presentation.

K-

 

[TEXAN]

oh, so it's running at 50Mhz.

Thanks for the link kristof.

 

[PC-Engine]

So that means the 50MHz HR-S has roughly the same performance as a 120MHz R-S? That's pretty sweet.

 

[amk]

oh, so it's running at 50Mhz.

Thanks for the link kristof.

I wanted a 200MHz PSP killer!

 

[amk]

...Err, upon more review, I'm again thinking HR-S is MBX Pro and R-S is MBX Lite. Which begs the issue why OMAP2420's polygon throughput is so low or why it's at least claimed to be so low. Maybe the HR-S is only clocked around 55-MHz even though the rest of the chip is much higher, like the SH-Mobile3 SH73182.

IIRC the descriptions go like this:

PowerVR       ARM
MBX Pro       n/a
MBX           HR-S
MBX Lite      R-S

(Code tag forces monospace fonts)

Are there any lists being kept of MBX equipped chips? So far I know of MBX Lite versions from Intel and Philips and MBX from TI and Renesas. All ARM except for Renesas' SH. Any others?

 

[Ailuros]

R-S/HR-S are codenames for ARM SoCs; unless any OMAP2 has the same CPU/components you can't really come to reliable estimates when it comes to throughput from the entire SoC (if it's CPU bound, poly rate is going to be more conservative from the geometry processor).

In any case why don't you folks just calculate stuff in MFLOPs? Since the VGP is a 4-way SIMD, you'll get 4 MFLOPs/MHz. While the clockspeed can vary from 50-200MHz, you'll have from 200-800MFLOPs/sec.

 

[Ailuros]

So that means the 50MHz HR-S has roughly the same performance as a 120MHz R-S? That's pretty sweet.

Albeit not sure at all I'd guess that the Lite version has only 1TMU, while =/>MBX could have 2TMUs.

50MHz * 2 = 100MPixels/sec
120MHz *1 = 120MPixels/sec

(add overdraw if you want to; those are estimated raw and not effective fill-rates. Actually PowerVR has a point in quoting effective fill-rates only for the time being, as long as competing hardware doesn't include yet culling techniques).