GeForce 3 At Tom's Hardware!

[pen^2]

wow, thx for a good pic, now thats some quality from a non-technological consumer's view, i think its gonna be just another solid product with high initial cost, the card's sheer performance being the appeal rather than the marketing gimmick.

 

[BenSkywalker]

Robo-

"How does the OGMS look? haven't seen it in action, I must admit if it actually CAN do this (70ish @ 1024x32 bit w/2x2) then I will be VERY VERY interested."

Not sure what FPS it will be hitting with 2x2 FSAA but with final drivers, look at the benches that we have seen and how poor the GF3 is performing in 16bit, it is clear that they still have quite a bit of work to do. As far as how does it look, go over to the B3D boards and there is a comparable screenshot from the V5 6K running 8X with LOD bias at -2, I would say they are very comparable. The texture clarity in Dave's shots is extremely impressive, I spent quite some time looking those over(with a "magnifying glass" at that).

"as far as T&L is concerned, I forget I'm dealing with someone who's CPU is about half the speed of mine. Of course, you forget you're dealing with someone whose CPU is twice the speed of yours, as well."

I'm running in the 700MHZ range right now and should be upgrading again soon. Ignoring that though, Rev has posted numbers with his rig, linkage. ~50% boost on his GHZ T-Bird running 1024x768 32bit color. Using MDK2 is a bad example in many senses(though it is nice to have a proper benchmark), it is like comparing Quake2(MDK2) to Quake3(Giants), MDK2 used it but didn't use the extra polygon complexity to utilize it, just more complex lighting(Giants uses both). That was also running the 6.31s and the latest drivers have improved Giants performance under Win9X a rather significant amount for myself(along with Jukka). The overall visuals in the games simply aren't comparable, Giants makes MDK2, Quake3 or UT look very dated.

"I haven't played Sacrifice or Giants (been too busy with SB2001 and NHL2001 - WHOOT!!! Still waiting for the Mario Lemieux roster update ), so I'll have to plead ignorance in that case."

Sacrifice it isn't so much about performance. They use a LOD system on the models where they change the amount of polys on the fly. With hard T&L this isn't very noticeable, they flicker and deform all the d@mn time under software T&L. One of the developers showed some benches(which they don't have in the game yet) and he was maintaining higher poly rates and faster FPS running hard T&L compared to soft. Sacrifice is not Giants in terms of visuals(nothing I have seen is close), but I would definately rate it as the second best looking game I have ever seen(and in an artistic only perspective arguably the best looking period).

"I still don't get why they can't implement a lod bias slider. It's incredible the difference that can make."

Check out those screenies Dave posted a link to and browse around the B3D boards for some V5 6K shots with the LOD bias adjustment, I don't think nVidia needs one(though I have to say I'd like to see it myself).

Dave-

Where have you been?

 

[Silhouette90]

>"How does the OGMS look? haven't seen it in action, I must admit if it actually CAN do >this (70ish @ 1024x32 bit w/2x2) then I will be VERY VERY interested."
>
>Not sure what FPS it will be hitting with 2x2 FSAA but with final drivers, look at the >benches that we have seen and how poor the GF3 is performing in 16bit, it is clear that >they still have quite a bit of work to do

Ben: More anisotropic filtering, more performance hits. The digit-life guys also have a look in the performance of the ansitropic filtering and after 16 taps, there is a huge performance hit and frame rates begins to crawl at very high resolutions. Fortunately, Their new FSAA technique is pretty cool and does what it should do without a performance hit.

For 16 bit rendering, they correct the memory access problem in gf2 series which is the main bottleneck for 32 bit rendering.(i.e. peak fill-rate 1600 Mtexel, actual fill-rate around 550 Mtexel for 32 bit, 1100 Mtexel for 16 bit for gf2. In gf3, now both 16 bit and 32 bit approachs around 1300 Mtexel). Thats why the performance scores of 16 bit does not effected much when compared to Geforce2 Ultra + do not expect any enormous improvement in 16 bit therefore.

 

[BenSkywalker]

"The digit-life guys also have a look in the performance of the ansitropic filtering and after 16 taps, there is a huge performance hit and frame rates begins to crawl at very high resolutions."

Driver revision- 10.5, not a shipping GF3 driver by any means. In fact, they had to use a reg hack to enable anisotropic, how much time do you think they have spent optimizing for it? I'm sure a lot of people greatly appreciated the numbers they provided, it would serve people well though to remember the early numbers we saw on the GeForce using the not for GF series 2.08 Detonators. They scored quite poorly when compared to the vanilla TNT2 which now we know to be far off the mark. Am I saying not to trust those benches as represenative of final GF3 performance? Absolutely.

"Thats why the performance scores of 16 bit does not effected much when compared to Geforce2 Ultra + do not expect any enormous improvement in 16 bit therefore."

They are under 60FPS 1600x1200 16bit Quake3 with a GF3, I would gladly wager that those numbers will improve significantly with final drivers. Look to MPixel/Mtexel or any other measurement you want, the GF3 is performing significantly slower in 16bit then it should be.

 

[Silhouette90]

Hey Ben,


<<<<&quot;The digit-life guys also have a look in the performance of the ansitropic filtering and after 16 taps, there is a huge performance hit and frame rates begins to crawl at very high resolutions.&quot;>>>>

<<Driver revision- 10.5, not a shipping GF3 driver by any means. In fact, they had to use a reg hack to enable anisotropic, how much time do you think they have spent optimizing for it? I'm sure a lot of people greatly appreciated the numbers they provided, it would serve people well though to remember the early numbers we saw on the GeForce using the not for GF series 2.08 Detonators. They scored quite poorly when compared to the vanilla TNT2 which now we know to be far off the mark. Am I saying not to trust those benches as represenative of final GF3 performance? Absolutely.>>

Hmm yep! You may be true, but still anisotropic filtering uses lots of texture data to compute the color value of the pixel. I think 8-tap means 8 texture data, so the amount of data required for each pixel doubles by doubling of the number of the taps. I think my performance hits discussion is still true, but they can improve the performance-of course-for every tap configuration, which makes 16+ taps configurations also viable

Note: I heard (of course I can not validate this note) that Nvidia does not supply their boards to reviewers because of immaturity of their drivers. I think thats what you talk about. And I am pretty sure that they would increase the performance considerably because they have the best driver group in the area with no doubt.

<<<<&quot;Thats why the performance scores of 16 bit does not effected much when compared to Geforce2 Ultra + do not expect any enormous improvement in 16 bit therefore.&quot;>>>>

<<They are under 60FPS 1600x1200 16bit Quake3 with a GF3, I would gladly wager that those numbers will improve significantly with final drivers. Look to MPixel/Mtexel or any other measurement you want, the GF3 is performing significantly slower in 16bit then it should be. >>

Even if the drivers matures, I do not still expect a huge difference in 16 bit. As far as I understand GF2 series also does its job well in 16 bit area, and this card would not improve it considerably much. However, I am really happy to see that 32 bit shines now. And if 32 bit performance is almost same as 16 bit performance, why to use 16 bit then?

 

[BenSkywalker]

I think I was reading what you were saying improperly-

&quot;Even if the drivers matures, I do not still expect a huge difference in 16 bit. As far as I understand GF2 series also does its job well in 16 bit area, and this card would not improve it considerably much.&quot;

I don't think it will be much, if any faster then the GF2U in 16bit, right now though it is getting its' @ss handed to it I think it will run much faster in 16bit then the scores on DigitLife indicate, but I would be very surprised to see it perform significantly faster then the GF2 boards we already have. BTW- I never use 16bit when I can use 32bit, I'm looking forward to 64bit color for that matter

&quot;You may be true, but still anisotropic filtering uses lots of texture data to compute the color value of the pixel. I think 8-tap means 8 texture data, so the amount of data required for each pixel doubles by doubling of the number of the taps. I think my performance hits discussion is still true, but they can improve the performance-of course-for every tap configuration, which makes 16+ taps configurations also viable&quot;

I think there will clearly be a performance hit, but I expect it to shrink a bit with driver improvements. Right now its' about a ~35% performance hit, I wouldn't be shocked to see it in the 25% range which is a price I would definately be willing to pay(it looks reeeaaalll good).

 

[Silhouette90]

Ben: Pretty cool dicussion, thanks pal!!!

See you around!!

PS: I am looking forward for an MX version (it is the only one my money can buy ).
Ps2: I am really curious about radeonII. Hope it rocks as well..

 

[Hawk]

Hey Pidge, I don't understand why you and Anandtech can't get 1600x1200 to work in UT. I'll had that option since day one...

 

[BFG10K]

Yes and no,If the V5 6k had stayed in development until today it would more than likely be a total different animal.

I doubt it. It would still be the same size, have the same power requirements and still require four fans. All in all it would still be an unwieldly card to use.

 

[DaveB3D]

But once you used it, you didn't want to use anything else.. that I know from experiance..

 

[RoboTECH]

&quot;But once you used it, you didn't want to use anything else.. that I know from experiance.. &quot;

oh STFU curse you!!!

BTW, does that weirdo still PM you at all saying you're me and I'm you?

 

[pidge]

Hawk,
Its a limitation of Unreal Tournament. It reads a fixed amount of supported resolutions from your video cards drivers starting from the lowest resolutions (I forgot what the number was). Anyways since Detonator 3, NVIDIA has raised the number of supported resolutions and so the maximum resolutions that are available on UT has dropped and I can only get up to 1024x768. With the older Detonator 3 drivers, I was able to get up to 1280x1024 but I guess NVIDIA has added more resolutions again. This is just a flaw with UT.