r420 efficiency

[hovz]

in looking at the theoretical benhmarks of nv40 and r420, both the xt and pro are less then 70% of their theoretical maximum. thats pretty bad. i have read that ati sacraficed some non af and aa performance, but this much?

 

[rwolf]

I would imagine that the GPU doesn't run flat out all the time because of CPU limitations. You still have the game code and driver code running on the CPU.

 

[PeterAce]

There's still some tuning work to be done on the X800, especially for AA. As someone mentioned in a review, the memory controller is very programmable, however it's not always obvious what the best settings are (i.e. what's best for one mode may not be best for another) and the combinations are enourmous. Expect large improvements with AA as we find more optimal settings. Stay tuned.

 

[hovz]

what about the non aa stuff. the xt fillrate is like 5.7 gigapixels, its maximum is like 8.3. the pros is like 3.2, its maximum is 5.7

 

[hkultala]

in looking at the theoretical benhmarks of nv40 and r420, both the xt and pro are less then 70% of their theoretical maximum. thats pretty bad. i have read that ati sacraficed some non af and aa performance, but this much?

They lack memory bandwith.
Clock the chip to 100 MHz and keep the momory at same speed and you get much closer to the theoretical.

But reality is not filling untextured or single-textured huge-sized triangles, so this kind of "fillrate achieved by synthetic tests" does not matter much.

 

[hovz]

in looking at the theoretical benhmarks of nv40 and r420, both the xt and pro are less then 70% of their theoretical maximum. thats pretty bad. i have read that ati sacraficed some non af and aa performance, but this much?

They lack memory bandwith.
Clock the chip to 100 MHz and keep the momory at same speed and you get much closer to the theoretical.

But reality is not filling untextured or single-textured huge-sized triangles, so this kind of "fillrate achieved by synthetic tests" does not matter much.

the xt ha smore bandwidth than the nv40 yet tis results are lower. and i understand this isnt a real game but if the r420 is only performing at 70% or less of its potential in filling pixels than its not performing as well as it could in games.

 

[FUDie]

in looking at the theoretical benhmarks of nv40 and r420, both the xt and pro are less then 70% of their theoretical maximum. thats pretty bad. i have read that ati sacraficed some non af and aa performance, but this much?

They lack memory bandwith.
Clock the chip to 100 MHz and keep the momory at same speed and you get much closer to the theoretical.

But reality is not filling untextured or single-textured huge-sized triangles, so this kind of "fillrate achieved by synthetic tests" does not matter much.

the xt ha smore bandwidth than the nv40 yet tis results are lower. and i understand this isnt a real game but if the r420 is only performing at 70% or less of its potential in filling pixels than its not performing as well as it could in games.

If you're looking at the color read/write only results, then those aren't very interesting. How often does a game do color reads/writes with no Z? I'd say rarely, if at all. The X800 XT's Z fillrate looks quite good, and since Z tends to take more bandwidth than color, the chip looks quite balanced.

-FUDie

 

[Mintmaster]

FUDie, the Z-buffer only takes more bandwidth than colour on old architectures, like Rage128 or GeForce2 and earlier. All chips have Z-compression now, so colour takes up plenty more bandwidth.

The Z-fill numbers aren't very encouraging either, because there's plenty of bandwidth. They should be reaching the peak. Just look at NVidia's actual 12GPix fillrate for Z only. I'm curious if they can maintain that rate with stencil reading and writing, but they probably can.

Finally, you can compare to the 9500 PRO, which has 9% less bandwidth per pipe per clock. The X800 XT should have 3.8 times the fillrate (> 7 GPix/s), but falls well short.

What's really discouraging is the multitexturing fillrate. There's really no excuse for that to be low, and I have no idea why the texel fillrate decreases when you add more textures than 2 or 3. Latency is all I can think of for the culprit.

There are significant efficiency problems with the X800 XT, and I'm disappointed. I assume they're coming from the latency of GDDR3, but they had plenty of time to optimize that since R300 was released. It mostly involves just inserting some FIFO's into the hardware, but maybe they didn't want to increase the die size any more. Maybe they even reduced FIFO sizes to squeeze 16 pipes onto the chip. Hopefully it is indeed just a matter of tweaking a few values.

On the bright side, it seems as though game performance is pretty close to where it should be, using the "9500 PRO times 3.8" rule in fill limited situations. I haven't checked too much, though, since the 9500 PRO hasn't been tested in a review with recent games for a while now.

 

[FUDie]

FUDie, the Z-buffer only takes more bandwidth than colour on old architectures, like Rage128 or GeForce2 and earlier. All chips have Z-compression now, so colour takes up plenty more bandwidth.

This is false. Not every pixel will require a color read/write, but nearly every pixel will require a Z read and many a Z write as well.

The Z-fill numbers aren't very encouraging either, because there's plenty of bandwidth. They should be reaching the peak. Just look at NVidia's actual 12GPix fillrate for Z only. I'm curious if they can maintain that rate with stencil reading and writing, but they probably can.

That's odd, I saw numbers for the X800 XT hitting close to it's maximum in Z only tests.

Finally, you can compare to the 9500 PRO, which has 9% less bandwidth per pipe per clock. The X800 XT should have 3.8 times the fillrate, but falls well short.

Again, not interesting to me.

There are significant efficiency problems with the X800 XT, and I'm disappointed. I assume they're coming from the latency of GDDR3, but they had plenty of time to optimize that since R300 was released. It mostly involves just inserting some FIFO's into the hardware, but maybe they didn't want to increase the die size any more. Hopefully it is indeed just a matter of tweaking a few values.

You're disappointed that the fastest hardware isn't efficient in some theoretical tests? Weird.

On the bright side, it seems as though game performance is pretty close to where it should be, using the "9500 PRO times 3.8" rule in fill limited situations. I haven't checked too much, though, since the 9500 PRO hasn't been tested in a review with recent games for a while now.

Who cares about the 9500 Pro when the X800 XT is double the performance of a 9800 XT in many cases?

-FUDie

 

[hovz]

the point is it can and SHOULD be higher

 

[FUDie]

the point is it can and SHOULD be higher

No, the point is to be the fastest, and it is. Being very efficient in some theoretical tests is all fine and dandy, but if it doesn't win you benchmarks, who cares? Measuring "pure color fillrate" when applications don't use "pure color fillrate" is useless.

-FUDie

 

[hovz]

even in textures its perfoming with low efficiency

 

[karlotta]

even in textures its perfoming with low efficiency

And?
why should i care? is there a better way to ascess value/performance. Go buy a 9800pro its less than 200$US( if you look ) or spend 400$US on a X800pro....Is efficency realy only about fillrate?

 

[Heathen]

the point is it can and SHOULD be higher

that's only in synthetic tests and being honest sybthetic test mean very little, it's the game performance I care about and that's (as far as I can tell) exactly where it should be.

 

[hovz]

the point is it can and SHOULD be higher

that's only in synthetic tests and being honest sybthetic test mean very little, it's the game performance I care about and that's (as far as I can tell) exactly where it should be.

nope that should be higher too

 

[jvd]

the point is it can and SHOULD be higher

that's only in synthetic tests and being honest sybthetic test mean very little, it's the game performance I care about and that's (as far as I can tell) exactly where it should be.

nope that should be higher too

sigh you read what opengl guy said . Why continue .

The focus now on video cards is aa and af .

I have yet to play a game on my 9700pro with out aa and af. I don't believe with a new 500$ card i would turn it off for any reason.

So even if non aa performance is lacking there is still plenty of performance to be gained in aa tests in which you would use on a 500$ card .


From reading your posts I just feel like your trolling.

Give ati's drivers the same thing you give nvidia's drivers. They are both the first sets of drivers for brand new video cards .

IN 3 -6 months both will get gains in performance by tuning the drivers for the new cards .

 

[hovz]

what do i give nvidias drivers? nothing but crticism, thats what i give them. i give crticism where its due. im sorry if ive offended you by not praising ati and nvidia for their efforts

 

[jvd]

what do i give nvidias drivers? nothing but crticism, thats what i give them. i give crticism where its due. im sorry if ive offended you by not praising ati and nvidia for their efforts

Funny . All i see is you complaining about ati and on top something already answered to u

 

[martrox]

This is another of these misguided threads that people need to think before they post. As has been stated here more than a few times, it's the uses that the card was designed for that counts. If a card performs well in synthetic tests, but poorly in real world usage, is that a good thing?

 

[trinibwoy]

I have yet to play a game on my 9700pro with out aa and af. I don't believe with a new 500$ card i would turn it off for any reason.

Guess you never played BF Vietnam then AA/AF doesnt work on me 9800PRO. I must say from my personal experience Nvidia (GF2Pro, TI4200) does have better drivers - I got out before the FX fiasco