NV40: Surprise, disappointment, or just what you expected?

[Nick Spolec]

I must say, I am both surprised and not surpirsed over the NV40.

I knew there was no way Nvidia would release another junker, and they would make sure it would be an impressive chip. But they still managed to surprise me with it's speed.

They especially surprised me with THE ABILITY TO TURN OF BRILINEAR. oF course, the NV40 is powerful enough that it doesn't need it.

All in all, I say, excellent release.

There is no way I will buy it though. It's gigantic cooler is frightening, and so are it's power requirements.

 

[KimB]

Well, the non-Ultra should be a bit more sane in terms of heat/power requirements. And I believe one reviewer did run the Ultra fine on a 350-watt PSU.

Anyway, I was surprised at how much more they packed into the design compared to the NV30 without even doubling the number of transistors. I guess that just goes to show how the NV30 design really was borked.

Oh, and I have to say: I told you so! There is no reason to believe that FP32 vs. FP24 is the reason for the performance difference between the R3xx and NV3x. With the amount of FP32 power of the NV4x, I see no reason that FP24 can be seen as a benefit.

 

[jjayb]

Oh, and I have to say: I told you so! There is no reason to believe that FP32 vs. FP24 is the reason for the performance difference between the R3xx and NV3x. With the amount of FP32 power of the NV4x, I see no reason that FP24 can be seen as a benefit.

Yup. You proved us all wrong. Being right on 1 out of 50 things about the Nv30 aint bad.

 

[Joe DeFuria]

Oh, and I have to say: I told you so! There is no reason to believe that FP32 vs. FP24 is the reason for the performance difference between the R3xx and NV3x. With the amount of FP32 power of the NV4x, I see no reason that FP24 can be seen as a benefit.

Lower precision has the indirect benefit of lower transistor counts needed to support it. (Presumably the same reason why NV40 doesn't support all operations FP32 mode.)

 

[Nick Spolec]

Oh, and I have to say: I told you so! There is no reason to believe that FP32 vs. FP24 is the reason for the performance difference between the R3xx and NV3x.

Sure it is, because the NV30 just couldn't do FP32 at an exceptable speed. It sucked at PS2.0 as it is... With FP32 on, it REALLY sucked.

The only reason Nvidia included in it the NV30 (unless it was an error in judgement) was to have a show piece, a one-up. "FP32 is better then FP24". What they don't tell you is that they will rarely ever be a situation that the chip does FP32.

But, I digress from my own topic....

 

[KimB]

Lower precision has the indirect benefit of lower transistor counts needed to support it. (Presumably the same reason why NV40 doesn't support all operations FP32 mode.)

Of course the NV40 supports all PS operations in FP32 mode. Regardless, with the incredibly strong performance of the NV40 with full FP32 shaders (current benchmarks indicate full FP32 speed, sometimes with speed drops, likely due to driver bugs, when dealing with FP16).

A few instructions will obviously not run at full speed in FP32, though, such as rsq and nrm.

 

[Wixard]

Overall I'm impressed. Glad to see nvidia still in the running.

Im just wondering if the nvidia mid/low end parts can be "soft moded" Im doubting it, but lets hpoe ati does something like this with the x800. (Anyone else think the X = 10? ) I noticed that. Oh well.

 

[Joe DeFuria]

Lower precision has the indirect benefit of lower transistor counts needed to support it. (Presumably the same reason why NV40 doesn't support all operations FP32 mode.)

Of course the NV40 supports all PS operations in FP32 mode.

I didn't say all pixel shader operations. I said "operations." (Such as floating point texturing / filtering.)

 

[Cali3350]

All i want to see is the $300 cards. As long as there reasonably better then my 9700Pro (which im not fully convinced) ill be happy.

 

[duncan36]

its amazing how much progress they made from the time they started down this track by releasing the 5700u 4x1 card to today with a 16x1 card.
They've been in red-alert mode for 9months+, their engineers deserve a vacation.

Its amazing how people claim you're nitpicking if you say you wont buy the card because of noise. I'd like to hear the card in 2d/3d modes to get a better idea.

 

[rwolf]

Oh, and I have to say: I told you so! There is no reason to believe that FP32 vs. FP24 is the reason for the performance difference between the R3xx and NV3x. With the amount of FP32 power of the NV4x, I see no reason that FP24 can be seen as a benefit.

You have proved nothing. The FP32 unit on the NV4X might be a different design. R3xxx has fast FP32 in everything except the pixel shader.

I don't see FP32 contributing to better IQ than FP24 and if it takes nearly double the transistors why bother?

The NV40 has equal or less IQ quality compared to R3XX on the odd occation in might be better.

Overall it looks like a pretty good design.

 

[Bjorn]

All i want to see is the $300 cards. As long as there reasonably better then my 9700Pro (which im not fully convinced) ill be happy.

I'm guessing that the 12 pipe versions of the new chips (R420 and NV4X) are going to be a whole lot faster then the 9700 Pro. Especially in games like Half Life 2.

 

[Joe DeFuria]

All i want to see is the $300 cards. As long as there reasonably better then my 9700Pro (which im not fully convinced) ill be happy.

Yup...I hope to see the following in a $300 card (from either ATI or nvidia)

12 pipes running at least 450 Mhz+, on a 500Mhz+ 256 bit bus. That would be a pretty killer card for $300, and a worthy upgrade from my 9800 non-pro (about the same as your 9700 Pro).

Such a card should translate into 2x the pixel shading perormance of our cards, and a bump or two up in resolution / AA for non shapding apps.

 

[Cali3350]

Uh, isnt the X800SE, the 8 pipe card with 128bit memory, ATI's $300 part?

 

[SlmDnk]

I am very pleasantly surprised. Now I'm waiting for ATI's next-gen card so that I can decide which one to buy

 

[Stryyder]

Uh, isnt the X800SE, the 8 pipe card with 128bit memory, ATI's $300 part?

Only if you believe the chart posted and [H]ard OCP which is second hand is genuine.

Remember even kyle said that he only believed some of the info on the chart.

 

[merlin371]

what is FP32?
sorry I'm a bit of a noob

 

[Cali3350]

Well, either way, BFG announced their product and the 6800NU is deff having 256bit bus and GDDR3 memory. Pair that with a 12 pipe chip and i supose even a 300mhz clockspeed would be a nice increase on my 9700Pro.

 

[EasyRaider]

$300 for a card with presumably less than half the performance of a $500 card would be an awful deal.

Nah, either X800SE will have a 256 bit bus, or it will cost less.

 

[Joe DeFuria]

Uh, isnt the X800SE, the 8 pipe card with 128bit memory, ATI's $300 part?

Only if you believe the chart posted and [H]ard OCP which is second hand is genuine.

Remember even kyle said that he only believed some of the info on the chart.

Correct.

I still insist that there's no way that the information on the SE is correct. Either the price is wrong, or the specs are wrong...or both.

ATI hasn't had $300 part with a 128 bit bus since before the 9700!