GeForce6800 w/ 8 pipes? (no,12)- Full review complete!

I missed this one, thanks DemoCoder!

But if that is the case why does the Ultra do so much better than the non-Ultra? Does it have 16 alpha blenders? (I'm not trying to be sarcastic DC, I'm trying to understand but I'm a tad too stupid I fear. )

 

Basically, what DC seems to be saying is that nV realized the bandwidth limitation for 6800plain allows for advertising 16 pixels pipes while conveniently only having to actually have 8 (since bandwidth won't support 16, why not just have 8 and say you've got 16, anyway, etc.?) nV's gotten a lot of mileage out of that approach in recent years... (With nV3x it was let's talk about ps2.0 but show 'em 1.x, Let's talk about about fp32 but show 'em fp16, and with nV40 so far it's been let's talk about shader 3.0 but show 'em 2.0, etc. ad inifinitum.)

Heh... I really have no clue as to why these results are as reported here. Either the 6800 samples looked at so far actually only have 8 pixel pipelines, or else all of the samples mentioned here are broken thus far, or else this is some kind of "SE" 6800 chip which nV's AIB partners are hoping they can get away with selling without having to reveal their Special Edition nature. The last possibility I would rate with high probability as we saw a lot of that with some nV3x products, IIRC. It seems clear that the 6800plain product nV officially announced is supposed to have >8 pixel pipes--but, again, we've seen how that goes as well. I admit it would be nice to know...

 

Just when I begin to think I have a handle on something, I get smacked back to reality. Hopefully DC or someone equally knowledgable can answer me these (TIA):

1. Are alpha blenders a normal pipeline/quad accessory, or are they part of the memory controller?
2a. If they're part of a quad, are there, say, two alpha blenders per quad, meaning the 6800 will only have six (with one quad disabled)?
2b. Or are alpha blenders part of the memory controller, and thus the 6800, like the 6800U, has eight (because it keeps the full 256-bit memory bus)?
3. How many alpha blenders do the NV35, R350, and R420 each have?
4. When Baron and others said NV40 would essentially act like an 8x2 part, was that just in reference to the ratio of pipes to bandwidth, or also to the number of alpha blenders?
5. Are alpha blenders in any way tied to the number of pipelines you can claim a GPU has?

 

edit: just got that from nVidia Europe this evening

 

seems the .rom file can not be downloaded.

try this plz : http://www.gzeasy.com/ours/cho/p212new.rar

 

No, that's not what I said at all doofus. The NV40 has 16 pipes and in all circumstances it behaves as if it has 16 pipes, unless you are doing alpha blending, in which case, the thoroughput is similar to 8-pipes. You will find ATI's chip behaves similarly in that they do not have enough bandwidth, and alpha blend fillrate will be lower than non-alpha blend. (8 bytes per pixel * 16 pipes * 500Mhz = 64 gigabytes per second)

Since you can't hope to achieve this, halving the number of blender units makes perfect sense. If you can't hope to write 16 alpha blended pixels per cycle, why include hardware for it?

NV40 has 16 pipes, each pipe has 2 shader units and 2 ROPs.

 

I'm still not getting why the Ultra scores so much higher than the non-Ultra than if that is the case, guess I'm just a mega-dufus.

 

Ok DC so if it is indeed blender units that are the limiting factor, then a card with 6 blenders at 335MHz should give about 60% of the performance of 8 blenders at 400MHz.

Assuming not otherwise limited.

 

Well, 6200 Mtexels/s for the 6800U is pretty damn high for 8-pipe throughput! I think this is wishful thinking. Obviously, ufon did the multitexturing fillrate test.

Well, 16 blend units would be useful for single or double channel formats.

Not a very common scenario, of course, and those formats are much more useful if they're high precision. I think NVidia's move was very smart, and likely enabled them to do fp blending at a die cost comparable to ATI's 16 blending units.

It seems now that NVidia will be shipping 8-pipe 6800NU's, which makes a lot more sense for a $299 part. Maybe the GT will now be 12-pipe.

I'm pretty sure ATI will have one too once the 9800's are depleted a bit. I did some simple probabilites a while back, and if the probability of a quad failing is less than 20%, there will be as more 16-pipe parts than 12-pipe parts. I doubt this is the case, and suspect there will be many more 12-pipe parts, indicating a higher failure rate which in turn would predict many 8-pipe parts. ATI won't be throwing these out.

 

Link works, hope Ufon can try it on his 6800

 

I believe a reason for that could be because, as a few people have said here (including a quote from NV europe posted by Ante P), something isn't right with the card, whether it's borked or has an old/bad bios.

 

hmm the BIOS I got from leadtek and the BIOS posted in this thread both says "Incorrect file format" when I try to flash them :/

 

Well, then, I don't know what you meant to say, since obviously the numbers reported here for the 6800 indicate 8 pixel pipes, but the numbers reported here for the 6800U indicate 16--as many have pointed out to you. The differences reported here are far too great to be accounted for by bandwidth and clock differences, if both gpus tested were indeed 16 pipe chips. You seem to be overlooking that aspect of the matter entirely.

 

http://www.pegasoft.cz/soft_down_free/nvflash446.pegasoft.cz.zip

boot in ms dos, type "nvflash -f p212new.rom" . (you need to unrar the file first ......)

 

? the new bios result slower than the old one ?

 

Thanks much for the info--so it's fair to conclude the 6800 is 12 pipes normally, then, and not 8 or 16?

 

"EEPROM and Chip ID mismatch" or something like that
chip is 0401 and BIOS is for 0402 (or the other way around)

and you can't override the safety precation so it just doesn't allow me to flash...