Another demo

Now, if only I had a DX9-capable GFX board... *sighs*

PowerVR, release a cheap Series5 with PS/VS 3.0 PPPPPLLLLLZZZZ! :)


*G*
 
Yeah, those are normal framerates. In 1024x768 I'm getting 363fps. Got a report earlier from a guy with a FX 5600 who got 43fps in that same resolution, though I figure he probably had vsync on, though that would still indicate less than 85fps. Now with Hyp-X reporting 63fps (though not stating resolution) I'm getting a little concerned about the GFFX performance. I mean, yes it's slow on floating point fragment shaders, but THIS slow ... the 9700 is like 6x faster. Can't see any reason why this particular shader would underperform this much. :?

Did anyone run the Fire demo on the GFFX btw? If so, how many fps?
 
i get around 644fps on average on 1280x960 with 16xAF and 6xAA (set in desktop and in the demo)...but then my card is clocked to 510/380 on the core

ran the demo on the 5200 (clocked at 400) i have lying around...
on 1280x960 with no AF and no AA, i get around 73fps

edit: resolution was at 1280x960, not 1280x1024

note: cpu @ 3.6 w/ 300fsb
 
Humus said:
shaders, but THIS slow ... the 9700 is like 6x faster.

I'm finding it to be closer to 7-9x faster (than the 5800U, with only slightly better numbers for the 5900U) in typical DX9 situations (especially those with an abundance of dependent texture reads) myself :? I was really hoping for much stronger DX9-class support from the hardware guys (all except ATI, who has been doing quite a bit better than I was expecting) than we've been seeing.
 
Ilfirin said:
Humus said:
shaders, but THIS slow ... the 9700 is like 6x faster.

I'm finding it to be closer to 7-9x faster (than the 5800U, with only slightly better numbers for the 5900U) in typical DX9 situations (especially those with an abundance of dependent texture reads) myself :? I was really hoping for much stronger DX9-class support from the hardware guys (all except ATI, who has been doing quite a bit better than I was expecting) than we've been seeing.

Maybe something else is wrong?
 
Humus said:
Now with Hyp-X reporting 63fps I'm getting a little concerned about the GFFX performance. I mean, yes it's slow on floating point fragment shaders, but THIS slow ... the 9700 is like 6x faster. Can't see any reason why this particular shader would underperform this much. :?

Note this is an 5600.
But way slow it is.

You should see ATI's pipedream demo running on it. :!:
Altough running is definately not the right word... :p

Did anyone run the Fire demo on the GFFX btw? If so, how many fps?

Runs with exactly same speed.
 
Hyp-X said:
Runs with exactly same speed.

So 63fps there too? That would make the R9700 only 3.25x faster in that demo. Getting 205 fps myself. If it's exactly the same performance then it may hint that there's something more than just slow shaders. Perhaps some kind of driver problem.
 
Humus said:
Perhaps some kind of driver problem.

Maybe, but it scales properly with resolution.
(The first thing I checked if I had a VSync type of problem, but appearently not.)

The two demos appear to have about the same number of instructions and are likely shader limited, so I don't see why having the exact same speed is surprising.

I'll check what I can come up tweaking the shaders.

Update: It scales with the number of instructions as well.
 
Well, the Fire demo has only one instruction considered complex, and that's a SIN. SIN has native hardware support in the GFFX and runs at the same speed as normal ALU instructions. On R9700 though it is expanded into like 8 instructions. So for the GFFX that should be a 8 instruction shader to execute, while the Radeon has like 15.
In the Electro demo there's 9 instructions of which one is complex, a POW. To my knowledge there's no native support for this in either hardware but is expanded to 3 instructions. So 11 vs. 11 instructions there. So the GFFX should run the Electro demo slow, while the Radeon should run it faster (which it does too). There are of course other factors too though ...
 
Humus, just for extra information, it actually takes the GFX 2 cycles to execute a pow instruction (per pipeline). I believe the R3xx architecture executes it in 1, but I'm not 100% on that one.
 
Back
Top