First REAL FX Benchmarks

[Ante P]

With 4X AA and 16x Aniso, there's no such thing as CPU limited though.

Yup there is, UT2003 botmatch is one of them.
I think Dungeon Siege and/or Morrowind would still be CPU-limited even with 4x FSAA and 8/16x Aniso.

Anyways, the scores that will really matter to me is 1280x960 and up with 4x FSAA/8x Aniso and up because that's basically what me 9700 Pro is able to do today.

 

[Nagorak]

With 4X AA and 16x Aniso, there's no such thing as CPU limited though.

Yup there is, UT2003 botmatch is one of them.
I think Dungeon Siege and/or Morrowind would still be CPU-limited even with 4x FSAA and 8/16x Aniso.

Yes, but the counterpoint that always seems to be ignored is who cares? I mean, you're getting 50 fps in a UT2003 botmatch or in DS, it's not really a big deal... The time you really need higher FPS is when you're playing an FPS-style game online which won't be CPU limited. For single player 50 fps is perfectly manageable, and for a top-down game like DS 50 fps is more than enough. Saying you're CPU limited in DS is almost like saying you're CPU limited in Minesweeper.

 

[Ante P]

With 4X AA and 16x Aniso, there's no such thing as CPU limited though.

Yup there is, UT2003 botmatch is one of them.
I think Dungeon Siege and/or Morrowind would still be CPU-limited even with 4x FSAA and 8/16x Aniso.

Yes, but the counterpoint that always seems to be ignored is who cares? I mean, you're getting 50 fps in a UT2003 botmatch or in DS, it's not really a big deal... The time you really need higher FPS is when you're playing an FPS-style game online which won't be CPU limited. For single player 50 fps is perfectly manageable, and for a top-down game like DS 50 fps is more than enough. Saying you're CPU limited in DS is almost like saying you're CPU limited in Minesweeper.

I was referring to benchmark numbers, not actual gameplay.

 

[Doomtrooper]

Not true..Dungeon Siege running with 6 players.. I host as the server on a high speed connection...on Veteran level my wifes machine is litteraly a slide show (she dies ALOT due to it too)..even with a 8500...while my wifes friend has a XP 2000 and 8500 and running fine.

More monsters, more players on Dungeon Siege equals more CPU limitations.

So again I say if you got a Athlon XP 2000 and you play UT 2003, a FX or a 9700 are going to be a wash in that scenario as they are platform limited...makes picking a video card much easier that way.

Different if you have endless pockets and are running Dual P4's or XP's on SCSII 160 drives etc..

 

[K.I.L.E.R]

Funny how I play DS on my P4 1.8GHz and when there are loads of monsters from originally being 1 monster I get jack all performance hit, as soon as I enable AF 128-tap my performance dips more and this is when there are no monsters on screen

Play at 1280x960.

 

[KimB]

I'm confused as to why the GF FX is slower in standard 1024*768*32, but faster with AA and Aniso? I thought the feeling was the GF FX would be faster stock, but slower with AA on. This seems to be the exact opposite of that situation?

CPU-limited, immature drivers.

 

[Doomtrooper]

Your system is slightly faster than my wifes..KT 133 (not even KT133a) with a 900 mhz Thunderbird on a 100 mhz fsb with a Radeon 8500 ...

I saw her frames hitting 4 fps...time for a upgrade shes not just into Solitaire no more

 

[Fred]

The FX has a significant fillrate advantage, so statistically you probably will see a sweet zone for the FX vs ATI at certain mid level resolutions and AA combinations in benchmarks that arent throwing that many huge textures around.

As you up the resolution, and in games that use a lot of texture passes, that advantage should naively ( we are ignoring many factors here. Like efficiency, cache architecture, etc) diminish.

Moreover, in situations where we are geometry limited, im not convinced that the FX will necessarily have an advantage here in the games of today. As a general rule, the more flexible an architecture is, the bigger a speed hit it takes. Im not sure if higher clocks will offset this drawback.

 

[K.I.L.E.R]

"Im not sure if higher clocks will offset this drawback."
Isn't that the point of the GFFX's massive clock speed?

 

[Dave H]

As you up the resolution, and in games that use a lot of texture passes, that advantage should naively ( we are ignoring many factors here. Like efficiency, cache architecture, etc) diminish.

Why would upping the resolution stress bandwidth more than fillrate? I can't think of any systematic reason why higher resolution would increase the required bandwidth per pixel.

OTOH, Mintmaster provided a convincing argument that the opposite is true--that upping the resolution tends to decrease required bandwidth per pixel, because there's no mip-map more detailed than the base texture, even if the mip-map selection algorithm would normally require one:

Another is the fact that when there's magnification rather than minification, you only need to sample the top level mipmap. Depending on the magnification, the texel to pixel ratio can be quite low, like in the 3DMark2001 fillrate tests or for lightmaps. This is especially true at high resolutions. In these cases texture bandwidth is nearly negligible, and in multitexture situations you really want another texture unit so that you can use extra bandwidth.

Similarly, multitexturing should tend to stress fillrate over bandwidth on a GPU with an nx1 pipeline organization, which all the DX9 cards have. If you're making multiple passes, than yes, that stresses bandwidth more because you also need to read what's already in the framebuffer. But with modern cards capable of 16x loopback, you need a better reason than multitexturing to be triggering multiple passes.

 

[Hellbinder]

here is a Quote from a similar thread to this one at Rage3d.

Hi again!

I've run the benchmark again (this time without stuttering ... precache was deactivated... sorry for that!)

my scores...

@ 398,25/756Mhz (almost r350 speeds ?...

150 (oced r300) vs. 143 (default nv30)

config:
2.4GHz P4 (slow ram timings @ 2.5-6-3-3) 133MHz FSB

...and please mind that an Atlon XP 2600+ with nForce2 should perform better than my P4 at default speed!

update: I get almost IDENTICAL results @ 2.88GHz (less then 1 frame difference)

Pretty much indicates what the R350 is going to do to the GFFX ULTRA. And thats only at just under 400mhz.

 

[Evildeus]

]Pretty much indicates what the R350 is going to do to the GFFX ULTRA. And thats only at just under 400mhz.

Yes if they can do the 400/400

 

[Bambers]

I'm confused as to why the GF FX is slower in standard 1024*768*32, but faster with AA and Aniso? I thought the feeling was the GF FX would be faster stock, but slower with AA on. This seems to be the exact opposite of that situation?

I think its the AF hitting the r300 here. AF is more of a fillrate than memory bandwidth hit and since the gf fx is going to have lots of spare texturing power with its limited memory bandwidth I think this is why its pulling ahead with the 4xAF. It would be interesting to see 2xFSAA only and no af.

 

[PiNkY]

Why would upping the resolution stress bandwidth more than fillrate? I can't think of any systematic reason why higher resolution would increase the required bandwidth per pixel.

Well technically bandwidth per pixel stays the same, you just have a lot more pixels to draw. Yet, your average texture cache efficency **should** increase with higher resolutions (more consequtive accesses to the same texture), so (c.p.) overally the demands should be more in the fillrate department (sorry for my lack of english language skills).

But the performance advantages of GFFX @ 4AF,2AA do not really surprise me, as both cards are propably mainly fillrate limited here, as with all their compression techniques the increase in Text./FB bandwidth will not be as serious and GFFX has lots more fillrate. Also UT2k3 uses lots of triangles, so ATI's adaptive method of AF does not safe them as much work as in older games (as the rel. amount of visible screen area will not consist of isotrophic surfaces).

Also the raw bandwidth advantage of 9700p should be partly offsetted by the higher granularity of GFFX memory access schemes (half clock@256bit < full clock@128bit when using 4way X-Bar with 32bit word size).

 

[K.I.L.E.R]

I will let you guys in on a dream I have had:
Copy/Paste

Guys, the GeforceFX is 120% faster than the GF4 Ti4600 using 8x FSAA.
I was disapointed when playing JK2.
Anything over 4x FSAA you will start to feel the heat. FSAA of modes <=4x to run awesome, anything over that and the performance drop starts increasing pretty heavily until reaching it's apex performance drop when using 8x FSAA.

Those UT2K3 scores with those IQ options on ARE the GeforceFX's sweetspots.

All this was in a "dream" btw.

Tip: Remember this post for future reference.

 

[Qroach]

hmmm, and is what is said about the new ATI card R350 is true about it being only 10 % faster than the FX, then what's the point in even releasing it? The FX seems pretty close in performance to the new radeon at least in the test above (need to see more), but f the R350 is only 10 % faster than that, well, blah.

 

[epicstruggle]

until doom3 is released. what benchmarks would you use to see:
--how bandwidth limiting the card is
--implementations of aa and af
--IQ
--whatever else you think would be relevant

so what games/applications would be good to show the above stated goals.

later,

 

[Dave Baumann]

hmmm, and is what is said about the new ATI card R350 is true about it being only 10 % faster than the FX, then what's the point in even releasing it? The FX seems pretty close in performance to the new radeon at least in the test above (need to see more), but f the R350 is only 10 % faster than that, well, blah.

So, by that reasoning, whats the point in releasing GFFX?

 

[nggalai]

OK, update from Thilo:

- he's waiting with publishing more benches until he's sure no NDA have been distributed among the likes of Anandtech and THG.

- drivers were detonator 42.63

- Terratec has disclosed their European pricing: GFFX 5800 Ulta @ €659 and GFFX 5800 @ €579 suggested retail price. It's not clear whether the non-Ultra will sport FXFlow or a more traditional cooling solution (at least Terratec doesn't seem to be sure about it).

- Overall (performance) impression of the GFFX5800 Ultra is very good, but with some disclaimers: apparently, in UT2003 4xMSAA is a lot slower than 2xMSAA. Thilo also posted this for comparison: "4x fsaa-only is slower than 8° AF-only, in UT2003."

ta,
-Sascha.rb

 

[Dave Baumann]

Terratec has disclosed their European pricing: GFFX 5800 Ulta @ €659 and GFFX 5800 @ €579 suggested retail price. It's not clear whether the non-Ultra will sport FXFlow or a more traditional cooling solution (at least Terratec doesn't seem to be sure about it).

Going by Quafro FX the only one with a single slot fan is the Q FX 1000, running at 300MHz. Q FX 2000 is at 400MHz with a two slot fan (but without the heatpipe/FlowFX solution).