B3D Should do this w/ the GFX...

[incandescent]

I think something that could be very telling as to which design is more efficient (NV30 v R300) would be to downclock the GF FX to about 300/300 and oc a Radeon9500 Pro to 300/300 (most can go A LOT higher) and then do a comparison. heck you could even OC the GF FX, boot a game up and watch it clock itself down to 300 core and see what happens...


sorry if it's been suggested b4.

 

[demalion]

I think you're plagiarising Joe. For shame!

 

[incandescent]

I think you're plagiarising Joe. For shame!

lol like i said, sorry if its been suggested before, I've been extremely sick with valley fever the past couple weeks and havent done much reading so if it has been suggested before you can lock/delete this thread.

 

[Mintmaster]

I was looking at that issue, and it looks like NV30 is some 20% behind ATI in that regard, which is not like the NVidia we knew since the GF3.

From the info here (1600x1200 to keep GPU clock scaling linear):
http://firingsquad.gamers.com/hardware/ati_radeon_9500_pro_overclock/default.asp
The 256-bit bus give the 9700 an advantage over the 9500 PRO of:
- 41% in Q3 with FSAA/AF
- 29% in UT2003 w/o FSAA/AF
- 15% in UT2003 with FSAA/AF
From Anand:
- 37% in SS2 w/o FSAA/AF
- 21% in SS2 with FSAA/AF (1024x768)

From Anand's review, GFFX Ultra has an advantage over the 9700 (non PRO, quality AF) of:
- 13% in Q3 with FSAA/AF
- 42% in UT2003 w/o FSAA/AF
- 19% in UT2003 with FSAA/AF
- 27% in SS2 w/o FSAA/AF
- 15% in SS2 with FSAA/AF (1024x768)

Do some math, and you get a 9500 PRO at 500/500 (85% overclock) having an advantage over a GFFX Ultra of:
- 16% in Q3 with FSAA/AF
- 1% in UT2003 w/o FSAA/AF
- 35% in UT2003 with FSAA/AF
- 18% in SS2 w/o FSAA/AF
- 21% in SS2 with FSAA/AF (1024x768)

The same advantage would apply to a GFFX clocked at 270/270 (obviously).

Now, I know I used multiple sources, and there are problems with linearity (although the GFFX and GFFX Ultra scale almost exactly with clock speed for these tests), etc.

Still, this gives you a big reason why everyone is disappointed with the performance of the GeforceFX, on top of the ordered grid AA, noisy cooler, size, heat, price...

 

[Evildeus]

Done by a french site ( a good one)

 

[Doomtrooper]

Interesting.

 

[Mintmaster]

That french site has very low FX numbers.

Is it downclocked? To which speed, 300/325 (like GF4 Ti4600) or 275/270 (like the 9500 PRO) ?

EDIT: I just went to the site. They're all 275/270. Thanks.

 

[Mintmaster]

That graph really paints a telling story. Do you think NVidia just pulled an Intel here, making up for lower per-clock efficiency with a higher clock?

That would explain the need for the crazy cooling, as well as why no-one predicted such high clock speeds for NV30 when rumours were just floating around.

 

[Evildeus]

That french site has very low FX numbers.

Is it downclocked? To which speed, 300/325 (like GF4 Ti4600) or 275/270 (like the 9500 PRO) ?

EDIT: I just went to the site. They're all 275/270. Thanks.

Sorry i forgot to put the speed (yeah it's 275/270).

Yes definitively interesting... unfortunately for Nvidia

 

[Luminescent]

What these benchmarks indicate to me are poor Nvidia implementations of Aniso and AA in the driver set, indicating it is quite immature. This aside, clock for clock, it seems the FX is comparable to the 9500/9700 architecture, albeit memory and compression schemes. Seems the 128 bit bus and the drivers are restricting the FX greatly.

Mintmaster, how do you conclude lower per-clock efficiency for the FX if all the cards are clocked at the same rate. If anything, the FX is a little more efficient clock for clock than the 9500, flawed aniso/aa aside. It seems that if it weren't for its 128-bit bus, clock for clock, FX would best a 9700 pro (not in image quality, unless it could implemented custom AA with the fragment shader in FP mode at a respectable rate).

 

[demalion]

That graph really paints a telling story. Do you think NVidia just pulled an Intel here, making up for lower per-clock efficiency with a higher clock?

Woah, it is nowhere that bad.

True, the AA and aniso are seemingly less efficient, but they are still better than the GF 4.

 

[Luminescent]

These benchmarks tell an interesting story. As I previously stated, the FX seems by no means to be the inefficient clock/op architecture many thought it would. Perhaps it is crippled by its 4 32-bit bus partitions bus as opposed to the 2 64-bit ones of the 9500 pro (if they are indeed 64-bit).

 

[Evildeus]

What these benchmarks indicate to me are poor Nvidia implementations of Aniso and AA in the driver set, indicating it is quite immature. This aside, clock for clock, it seems the FX is comparable to the 9500/9700 architecture, albeit memory and compression schemes. Seems the 128 bit bus and the drivers are restricting the FX greatly.

Mintmaster, how do you conclude lower per-clock efficiency if they all the cards are clocked the same way. If anything, the FX is a little more efficient clock for clock than the 9500, aside from the flawed aniso/aa implementations. It seems that if it weren't for its 128-bit bus, clock for clock, it would best a 9700 pro.

How do you conculde that the issue with AA/AF comes from the driver? I hope for Nvidia but it's not that clear for me at least.

As for effeciency, i agree with you. How do you infer this Mintmaster?

 

[Luminescent]

Maybe I should have stated that AA and Aniso implementations may be suffering as a result of a driver issue (Uttar located some discrepancies in 2X MSAA; may indicate 4X and the other modes are not completely up and running, issues not present with the Geforce 4). If not, the z-sampling and blending implementations of the AA units are just not up to standard. Hopefully it is the former.

 

[KimB]

I think that they should have at least also shown some benches with those settings at a lower resolution. The FX seems to do rather poorly at 16x12 w/ 4x FSAA. I think it's due to memory limitations (judging from Anand's FSAA shots, or lack thereof, it looks like the FX doesn't downsample the framebuffer until scanout, meaning the backbuffer takes up just as much space as the front...significantly increasing memory size requirements).

 

[Luminescent]

That possiblity slipped my mind Chlanoth 8), should have thought of it.

 

[Evildeus]

I think that they should have at least also shown some benches with those settings at a lower resolution. The FX seems to do rather poorly at 16x12 w/ 4x FSAA. I think it's due to memory limitations (judging from Anand's FSAA shots, or lack thereof, it looks like the FX doesn't downsample the framebuffer until scanout, meaning the backbuffer takes up just as much space as the front...significantly increasing memory size requirements).

Well the 9500 pro should suffer the same limitations...

 

[2B-Maverick]

I think that they should have at least also shown some benches with those settings at a lower resolution. The FX seems to do rather poorly at 16x12 w/ 4x FSAA. I think it's due to memory limitations (judging from Anand's FSAA shots, or lack thereof, it looks like the FX doesn't downsample the framebuffer until scanout, meaning the backbuffer takes up just as much space as the front...significantly increasing memory size requirements).

well, dear Chalnoth.. look at what they benched against: a 9500Pro... with a 128Bit mem-interface just like the GFFX.

NV seems a tiny bit more efficient without AA/AF. But once you set those to HQ levels.. OUCH.