• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

PSP Specs from Sony at Hot Chips: 35 million polygons per second

doncale

Banned
http://www.extremetech.com/article2/0,1558,1639233,00.asp?kc=ETRSS02129TX1K0000532


The PSP's graphics engine will feature a 512-bit interface, Okabe said, pushing 664 million pixels or 35 million polygons per second. Freed from the need to conform to any other graphics API besides its own, Sony decided to support some basic graphics primitives as well as directional lighting, clipping, environment projection and texture mapping, fogging, alpha blending, depth and stencil tests, and dithering, all using either 16- or 32-bit color. The 166-MHz graphics core will include 2-Mbytes of embedded graphics memory.


Sony apparently will support a graphics model based on surfaces, rather than polygons. Okabe displayed an illustration of a cartoon character that looked more realistic than a polygon-based model, which he said contained the same amount of data. The graphics block will also be capable of vertex blending, a morphing technology that can interpolate changes made between objects.



"Small data size is advantageous to mobile data software," Okabe said.
 

Phoenix

Member
I also heard that its got 4MB of integrated memory instead of 8MB. So I have to ask again - does Sony shop at a memory store where the prices never go down?
 

Phoenix

Member
Mejilan said:
I didn't bother to read that, but is that in game or just raw numbers?

raw theoreticals. Polygon count in a game is going to be different in the game - and all its various scenes. Whenever someone gives you a figure like that - slap them and say 'tell me what its like in the average game utilizing all features of the architecture and exposed to the limitations of memory and the bus'
 

Lord Error

Insane For Sony
I also heard that its got 4MB of integrated memory instead of 8MB. So I have to ask again - does Sony shop at a memory store where the prices never go down?
They upgraded the main RAM from 8 to 32 MB.

What was the count for the DC and N64?
DC was limited to something like 6M P/s. Most games were lucky to reach even 1-2M, but some (well one, that LeMans game made by Melbourne house) were peaked even up to 4-5M by developers words.

N64 was much less than that. I'm not sure exactly, but the maximum was in the neighbourhood of 200K
 

tenchir

Member
Phoenix said:
I also heard that its got 4MB of integrated memory instead of 8MB. So I have to ask again - does Sony shop at a memory store where the prices never go down?

Two questions. 1st) when you say "integrated memory," do you mean embedded memory or main memory? Big difference. 2nd) If it's main memory, what kind of memory/speed are they using?

I tried looking for it at psp.ign.com, but it seems conflicting.

Initial Specs:

PSP CPU Core
MIPS R4000 32-bit core
128-bit bus
1-333MHz (1.2V)
8MB eDRAM main memory (*Upgraded to 32MB in May 2004)

PSP Graphics Core
1-166MHz (1.2V)
256-bit bus
2MB eDRAM (VRAM)

Updated Specs:

PSP CPU (System clock frequency 1~333MHz)
32MB Main Memory
4MD Embedded DRAM


It doesn't seem that it is using DDR memory chips that are mass produced for PCs and instead is using very costly embedded memory.
 

doncale

Banned
PSP's 35 million polygons/sec is equivalent to PS2's 66 million. re: realworld in-game sustained polygon performance with texture mapping, lighting, effects, will be much lower. probably under 10 million pps for most games.

Dreamcast could do 3-7 million polygons/sec depending on the situation. most games were in the 1-2 million polygon range. many games were under 1 million.

Nintendo 64 could manage only about 160,000 polygons/sec with everything on, in games.

I expect PSP games to look somewhere inbetween Dreamcast and Gamecube. better than PS2 in most areas (like Gamecube) with less polys than PS2.
 

btrboyev

Member
"It's not official but basicly they say that Sony has told them the price would be $150, and say that this price will be official announced (therefore confirmed) by Sony at TGS in September."

riiiiiiiiiggght
 
D

Deleted member 284

Unconfirmed Member
Drexon said:
"It's not official but basicly they say that Sony has told them the price would be $150, and say that this price will be official announced (therefore confirmed) by Sony at TGS in September."

http://forum.teamxbox.com/showthread.php?t=287821


Sony as a company cannot afford to release PSP @ that price point. Shit, I don't think MS could (from a shareholder pov)
 

Chittagong

Gold Member
Okabe's presentation of the I/O also contained some unexpected surprises. Early disclosures of the PSP indicated that the player would be capable of communicating via 802.11b WiFi. The only I/O functions Okabe described were USB 2.0 and Memory Stick, Sony's small-form-factor flash memory format.

Eh?
 

Tekky

Member
Assuming those specs are anything like the specs for various chips they presented at ISSCC, then you can safely assume that:

1) the polygon figure applies to orthogonal projection only. Divide it by at least 2 to get a "real" (but still "peak theoretical") number.

2) the pixel figure is really just a count of the number of texels sampled from per clock. Divide it by 4 to get the real (but still "peak theoretical") number.

Never, ever count on Sony to give you viable performance figures. Do expect hype.

Do always remember: performance specs are pointless; only the actual games matter.
 

Lazy8s

The ghost of Dreamcast past
Assuming the balance of performance is more traditional on the PSP than the PS2, the PSP's polygon benchmarks should be more applicable. On PS2, it's at times faster to just throw more polygons at the scene which can't get used than to spend time on backface culling.
 

Lord Error

Insane For Sony
2) the pixel figure is really just a count of the number of texels sampled from per clock. Divide it by 4 to get the real (but still "peak theoretical") number.
So what you are saying is that PSP will have a theoretical pixel fillrate peak approx ten times less than the PS2? (twenty times if you count non textured fillrate on the PS2 as it's max)
 

Lazy8s

The ghost of Dreamcast past
If PSP's hidden surface removal is better than the PS2's, its fillrate will be more proportionally effective too.
 

KickyFast

Member
2) the pixel figure is really just a count of the number of texels sampled from per clock. Divide it by 4 to get the real (but still "peak theoretical") number.

I don't think this is correct. The PSP has 256-bit memory for texture sampling. That's enough for 16 16-bit texel reads per cycle. Based on the 664 Mpixel fillrate and the 166 MHz clockrate, we can assume that the PSP has 4 pixel pipelines. That gives 64-bits for each pixel pipeline. That's enough to sample 4 16-bit texels per pipeline for free bilinear filtering.

BTW, bilinear filtering is for free on PS2 also.
 

FightyF

Banned
I believe the SH-4 (Dreamcast) could punch out 11 million triangles, but you never saw anything near that in it's games. AFAIK, most of it's games ran at 700K to 1 mil per second.
 

Fafalada

Fafracer forever
Tenchir said:
It doesn't seem that it is using DDR memory chips that are mass produced for PCs and instead is using very costly embedded memory.
It's not using embeded memory for the 32MB block.

Doncale said:
Dreamcast could do 3-7 million polygons/sec depending on the situation
To be quite specific, Dreamcast peaked at 6.7MPP/s(limited by the GPU) - that's the number equivalent to 66/75 for the PS2.
N64 peaked around 500Kpoly/sec IIRC.

Tekky said:
1) the polygon figure applies to orthogonal projection only. Divide it by at least 2 to get a "real" (but still "peak theoretical") number.
Even if that were the case, you're making one helluva assumption about performance implications.
on certain hardware (ie. PS2) the difference between perspective and non-perspective transform is 1cycle (at peak) - or 60M/75M using a single VU.

As for the pixel figure, the chip native sampling is bilinear. So no, dividing by 4 wouldn't make any sense.

Lazy said:
On PS2, it's at times faster to just throw more polygons at the scene which can't get used than to spend time on backface culling.
Those polygons are always thrown at the scene Lazy, and the numbers you see quoted by devs Include them in the count also. Even on DC.
 

junkwaffle

In Front and Drawing Away
Tekky said:
Assuming those specs are anything like the specs for various chips they presented at ISSCC, then you can safely assume that:

1) the polygon figure applies to orthogonal projection only. Divide it by at least 2 to get a "real" (but still "peak theoretical") number.

2) the pixel figure is really just a count of the number of texels sampled from per clock. Divide it by 4 to get the real (but still "peak theoretical") number.

Never, ever count on Sony to give you viable performance figures. Do expect hype.

Do always remember: performance specs are pointless; only the actual games matter.


I sooooo believe this post.
 

Lazy8s

The ghost of Dreamcast past
Fafalada:
Those polygons are always thrown at the scene Lazy,
But they don't always produce pixels, which was where I was comparing the relevancy of peak polygon figures between PSP and PS2.
 

Fafalada

Fafracer forever
But they don't always produce pixels, which was where I was comparing the relevancy of peak polygon figures between PSP and PS2.
Well my point was that the PSP number stated is for peak transform, and triangle setup, which is equivalent to PS2s raw numbers.
I don't think culling really affects those numbers (on PS2 you generally don't need it, and on PSP I assume it's free and the fillrate is probably more optimal with it enabled(seeing that eDram is not as fast)).
How PSP polygon numbers translate to realworld will depend on specs for other operations (lights, skinning etc.). If the lighting pipeline happened to be horribly slow for instance, the peak number could be Less applicable then for PS2.
 

Panajev2001a

GAF's Pleasant Genius
Tekky said:
Assuming those specs are anything like the specs for various chips they presented at ISSCC, then you can safely assume that:

1) the polygon figure applies to orthogonal projection only. Divide it by at least 2 to get a "real" (but still "peak theoretical") number.

Doing simple Perspective Projection the EE can push about 102.85 MVertices/s: you can get that down on VU1 to 5 cycles (using the EFU's extra FMAC and FDIV) and 7 cycles on VU0 (all per Vertex).
 

Panajev2001a

GAF's Pleasant Genius
Fafalada said:
Well my point was that the PSP number stated is for peak transform, and triangle setup, which is equivalent to PS2s raw numbers.
I don't think culling really affects those numbers (on PS2 you generally don't need it, and on PSP I assume it's free and the fillrate is probably more optimal with it enabled(seeing that eDram is not as fast)).
How PSP polygon numbers translate to realworld will depend on specs for other operations (lights, skinning etc.). If the lighting pipeline happened to be horribly slow for instance, the peak number could be Less applicable then for PS2.

True, but you do have a 2.6 GFLOPS VFPU to fall-back on... not too much, but the Dreamcast could do 396 Gazillion polygons per second with a 1.4 GFLOPS FPU.
 

Panajev2001a

GAF's Pleasant Genius
Phoenix said:
I also heard that its got 4MB of integrated memory instead of 8MB. So I have to ask again - does Sony shop at a memory store where the prices never go down?

The original number was 12 MB.

8 MB of embedded main RAM, 2 MB for the Media Engine and 2 MB for the GPU.

They took out the 8 MB of embedded main RAM, they added a memory controller and 32 MB of external RAM to function as main RAM/System Memory.

35 MPolygons/s is a peak figure: it is the same as the Triangle Set-up Engine peak performance as well.


Something I'd like to point out is that now the Media Engine which is powered by another R4000i does include an FPU (that sits on one of the R4000i's CO-OP pipes).

At 333 MHz this means a peak of almost 0.67 GFLOPS.
http://www.extremetech.com/slideshow_viewer/0,2393,l=&s=201&a=133950&po=3,00.asp
 

nathkenn

Borg Artiste
i never understood these number, after running multiple passes and adding playable game data games dont push past even hundreds of thousands of poly's, even next gen
 

Panajev2001a

GAF's Pleasant Genius
nathkenn said:
i never understood these number, after running multiple passes and adding playable game data games dont push past even hundreds of thousands of poly's, even next gen

???

Hundreds of thousands of polygons ?
That figure you know can mean something like 59+ MPolygons/s at 60 fps and considering that some people on Xbox managed >20-25 MPolygons/s I do not think you are very close.
 

Fafalada

Fafracer forever
Panajev said:
True, but you do have a 2.6 GFLOPS VFPU to fall-back on... not too much
I didn't want to get the discussion too convoluted so I stuck to GPU discussion only.
Anyway the VFPU is rated slighty higher then a single VU, so it's actually "quite much" in that aspect.
 

Panajev2001a

GAF's Pleasant Genius
Fafalada said:
I didn't want to get the discussion too convoluted so I stuck to GPU discussion only.
Anyway the VFPU is rated slighty higher then a single VU, so it's actually "quite much" in that aspect.

Hehe, I was only kidding: that is a lot :).

I think VFPU is VU0 minus VIF0, in the lower pipe-line we have only the FDIV (the branching can be done with normal instructions present in the MIPS core) and there are no micro-memories.

At least they would be able to re-use the macro-mode instruction set.

2.6 GFLOPS seems to me not the peak number: at 8 FP ops per cycle (if it was something like VU0) this would mean that you would get an original clock of 325 MHz.

So, VFPU might either be quite different from VU0 or they took into account some limitations that the VFPU might have in the final 2.6 GFLOPS number.
 

Panajev2001a

GAF's Pleasant Genius
nathkenn said:
have you ever modeled before? a million poly's is a whole hell of a a lot

It sure is, Unreal Engine 3 outdoor scene plan for more than 100 MPolygons in the level.

60 Milliong polygons per second means 1 Million polygons per frame and the game S.T.A.L.K.E.R. already achieves that in peak stuations.
 

Fafalada

Fafracer forever
Panajev said:
Hehe, I was only kidding: that is a lot
As long as we don't compare it to SH4 of course. Then it's tiny :D

At least they would be able to re-use the macro-mode instruction set.
I'm afraid I can't really comment on that :p

that the VFPU might have in the final 2.6 GFLOPS number.
Sounds just good old number rounding to me.
You know, if you took 166.5 and multiplied it by 4, the pixel fillrate would be... well... ;)
 

cybamerc

Will start substantiating his hate
All that power is nice in theory but when the games barely look better than what was released on the DC what good is it?
 

nathkenn

Borg Artiste
100m poly level in the unreal tech demo was impressive, but it just sounds like sloppy modeling. wow, what i could do with 100m polys. the only way i can even think to have that much detail is by running a million smooths on a model for no reason heh. the only models i've seen that have that many polys is stuff generated by scanners and those models are useless for games
 

Panajev2001a

GAF's Pleasant Genius
Fafalada said:
As long as we don't compare it to SH4 of course. Then it's tiny :D

Even scaling the SH-4 at 33 MHz would produce about 2.3 GFLOPS which is still coming short of what the VFPU can do.

I'm afraid I can't really comment on that :p

Hehe, soon we will have Austin Game Conference in which Sony promised their best and most in-depth "Programming the PSP presentation".

We will know :).

Still, seeing that basically every PlayStation 2 developer uses VU0 in macro-mode (between those who use the VU0 at all), it would be only smart of Sony to build on that familiar ISA.

Fast PlayStation 2 to PSP conversion... hehe we uncovered a clue, I think :).

The MAcro-Mode ISA is not really more limiting than the Micro-Mode one, but less efficient as you cannot really use the dual-isse nature of the VU0's VLIW engine (1 VLIW word = 2x64 bits instructions per cycle).

Sounds just like good old number rounding to me.
You know, if you took 166.5 and multiplied it by 4, the pixel fillrate would be... well... ;)

I know, very evil... that is why they rounded the clock-speed to 166 MHz in the presentations :).
 

Fafalada

Fafracer forever
Panajev said:
Even scaling the SH-4 at 33 MHz would produce about 2.3 GFLOPS which is still coming short of what the VFPU can do.
But SH4 has the magic powers, like you said it can transform gazilion polygons :p

Hehe, soon we will have Austin Game Conference in which Sony promised their best and most in-depth "Programming the PSP presentation".
We will know
Well it's their decision when/if talk things public, poor saps like us only have the right to remain silent. Wonder if they'll go into any lowlevel detail or just stick to their APIs for that though.

but less efficient as you cannot really use the dual-isse nature of the VU0's VLIW engine (1 VLIW word = 2x64 bits instructions per cycle).
Actually the problem is that macro instructions occupy the whole R59k pipe - no dual issuing with macro ops, not even CPU instructions.

cybamerc said:
All that power is nice in theory but when the games barely look better than what was released on the DC what good is it?
Give it time - there was early DC stuff that looked like filtered PSOne games too, some of it even released looking like that.
 

Panajev2001a

GAF's Pleasant Genius
http://www.extremetech.com/slideshow_viewer/0,2393,l=&s=200&a=133950&po=10,00.asp

Less than 500 mW for the PSP CPU while playing MPEG4 AVC video with audio (ATRAC3+ is used for the audio).

cybamerc said:
All that power is nice in theory but when the games barely look better than what was released on the DC what good is it?

They do not look better than this ?

duck_1.jpg
 

Panajev2001a

GAF's Pleasant Genius
Fafalada said:
But SH4 has the magic powers, like you said it can transform gazilion polygons :p

:D

Well it's their decision when/if talk things public, poor saps like us only have the right to remain silent. Wonder if they'll go into any lowlevel detail or just stick to their APIs for that though.


Actually the problem is that macro instructions occupy the whole R59k pipe - no dual issuing with macro ops, not even CPU instructions.

Yikes, forgot about that... well, PSP's R4000i will not have a problem, it is only a single-issue processor not a 2-way superscalar core like the 5900i :).
 

cybamerc

Will start substantiating his hate
Panajev2001a:

> Do we have released games on PSP yet ?

So what's your point? Software in development is the equivalent of a tech demo now?
 

Panajev2001a

GAF's Pleasant Genius
cybamerc said:
Panajev2001a:

> Do we have released games on PSP yet ?

So what's your point? Software in development is the equivalent of a tech demo now?

Considering the development environment (software emulators till past E3 this year), early documentation (VRAM usage, the VFPU, etc... all details that took a while to trickle down to developers), RAM change (many games were targeted at 8 MB of main RAM), etc... I'd say yes.

The pictures of games in development that have recently come out (Armored Core) did look pretty nice: no difference between the PlayStation 2 and the PSP shots really.
 
Top Bottom