NV35 - already working, over twice as fast as NV30?
- Thread starter PrzemKo
- Start date
gokickrocks
Regular
aaXiom3D said:BRiT said:
Well just think 2sec... Where can you find 500MHz (1000MHz effective) DDR-I memory ?
hynix has em
aaXiom3D said:BRiT said:
Well just think 2sec... Where can you find 500MHz (1000MHz effective) DDR-I memory ?
The same place you can find an NV35, or for that matter the NV30 5800 Ultra -- the hardware fairy test labs...
Seriously though, Hynix has some 500Mhz DDR-I memory.
Who says it has to be DDR I? The source could have just as easily mistaken DDR II for DDR I, as he could've mistaken 500MHz effective for 500MHz real.
I admit 250MHz real sounds more in keeping with nV's core-mem clock-phasing, but would a 250/250 NV30 actually score 48fps at 16x12x32+4xAA+8xAF? Could a 250/250 (256-bit) 8x1 (z + color) NV35 actually score 111fps at the same settings that a 250/250 R300 scores 70 on the basis of an improved memory controller alone?
Or are we looking at GigaPixel's "free" FSAA finally coming to fruition?
(There, chew on that for a while. )
I admit 250MHz real sounds more in keeping with nV's core-mem clock-phasing, but would a 250/250 NV30 actually score 48fps at 16x12x32+4xAA+8xAF? Could a 250/250 (256-bit) 8x1 (z + color) NV35 actually score 111fps at the same settings that a 250/250 R300 scores 70 on the basis of an improved memory controller alone?
Or are we looking at GigaPixel's "free" FSAA finally coming to fruition?
(There, chew on that for a while.
)Anand didn't confirm any number for the tests - only the NV35 final specs.
Well, if this is 250/250 for the NV35...
For all we know, nVidia could have decided to use lossy compression everywhere
"Hey! How come the sky is pink?!"
More seriously though, my bet is still that nVidia's first attempt at a DDR-II memory controller/interface was a gigantic failure.
And I also believe ATI's Color Compression is less advanced than nVidia's, my bet is that ATI can only compress when all samples are the same and nVidia could compress when only some of the samples are identical.
nVidia probably also implemented better Z Compression - the current leader in Z Compression is the RV350, and that's kinda lame IMO
Uttar
Well, if this is 250/250 for the NV35...
For all we know, nVidia could have decided to use lossy compression everywhere
"Hey! How come the sky is pink?!"
More seriously though, my bet is still that nVidia's first attempt at a DDR-II memory controller/interface was a gigantic failure.
And I also believe ATI's Color Compression is less advanced than nVidia's, my bet is that ATI can only compress when all samples are the same and nVidia could compress when only some of the samples are identical.
nVidia probably also implemented better Z Compression - the current leader in Z Compression is the RV350, and that's kinda lame IMO
Uttar
Pete said:1- Who says it has to be DDR I? The source could have just as easily mistaken DDR II for DDR I, as he could've mistaken 500MHz effective for 500MHz real.
2- I admit 250MHz real sounds more in keeping with nV's core-mem clock-phasing, but would a 250/250 NV30 actually score 48fps at 16x12x32+4xAA+8xAF?
3- Could a 250/250 (256-bit) 8x1 (z + color) NV35 actually score 111fps at the same settings that a 250/250 R300 scores 70 on the basis of an improved memory controller alone?
4- Or are we looking at GigaPixel's "free" FSAA finally coming to fruition?
(There, chew on that for a while.
1- yep probably but the nv rep told me DDR-I
2- someone with a NV30 ultra to check this score at 250/250 ?
3- not just improved memory controler. More pipelines (8 vs 4), more bandwidth, more pipeline optimizations
4- you'll have to wait NV40 to see some Gigapixel ideas inside Nvidia silicon...
Informative post, aaXiom.
As to #3, I assumed they "doubled" the pipes, as I said z + color, as opposed to 8x1 (just) z. (I guess ppl refer to that as 8x0 around here, but I can't be expected to keep up with all the buzzwords flying around here--I've got zixels and shoxels out the yin-yang.) I'm still surprised NV35 could outperform R300 by 50%, clock for clock. Very impressive, if true.
I await GP tech with great interest.
As to #3, I assumed they "doubled" the pipes, as I said z + color, as opposed to 8x1 (just) z. (I guess ppl refer to that as 8x0 around here, but I can't be expected to keep up with all the buzzwords flying around here--I've got zixels and shoxels out the yin-yang
.) I'm still surprised NV35 could outperform R300 by 50%, clock for clock. Very impressive, if true.I await GP tech with great interest.
Thanks for the info aaXiom, very interesting!
BTW, MuFu said at nV News ( and I posted it at www.notforidiots.com/GPURW.php , too ) that the NV35 supports *both* DDR-I and DDR-II. So it sounds possible they'd use 500Mhz DDR-II for the ultra, but 400Mhz DDR-I for the regular, for example.
I'd be surprised if the NV35 truly had double the pipelines. I'd guess they're using some clever tricks, and that in practice each of their 8 pipelines is less efficient than each of the 4 pipelines on the NV30.
But they'd also fix some major bugs in the NV30 pipelines, so that in practice each slightly slower or faster than a NV30 pipeline.
And then you've got two times the pipelines, giving roughly double performance.
Wouldn't be surprised if the performance difference between FP32/FP16/INT12 got smaller, too.
As for NV40... Very good to know we'll get some GigaPixel stuff ( frankly, that's what I had hoped for ) , but are we getting some Rampage stuff in it too?
You know what I mean... "Revolutionary Gateless Design" j/k
I mean, the Rampage was supposed to have things like partly dynamic gates, microcode programmability, and so on. Any chance we'd get some of that on the NV40?
Uttar
BTW, MuFu said at nV News ( and I posted it at www.notforidiots.com/GPURW.php , too ) that the NV35 supports *both* DDR-I and DDR-II. So it sounds possible they'd use 500Mhz DDR-II for the ultra, but 400Mhz DDR-I for the regular, for example.
I'd be surprised if the NV35 truly had double the pipelines. I'd guess they're using some clever tricks, and that in practice each of their 8 pipelines is less efficient than each of the 4 pipelines on the NV30.
But they'd also fix some major bugs in the NV30 pipelines, so that in practice each slightly slower or faster than a NV30 pipeline.
And then you've got two times the pipelines, giving roughly double performance.
Wouldn't be surprised if the performance difference between FP32/FP16/INT12 got smaller, too.
As for NV40... Very good to know we'll get some GigaPixel stuff ( frankly, that's what I had hoped for ) , but are we getting some Rampage stuff in it too?
You know what I mean... "Revolutionary Gateless Design"
j/kI mean, the Rampage was supposed to have things like partly dynamic gates, microcode programmability, and so on. Any chance we'd get some of that on the NV40?
Uttar
Thanks aaXiom3D for clearing up the NV35 speeds, that is what I thought but couldn't really prove. The test shows the advantage of a 256bit bus of the NV35 over a 128bit bus of the NV30 is what I get from the results. The numbers I think does compare to the R300 core and card design but to conclude it is faster without true IQ comparisons and knowing the quality of the output I think would be premature.
Now since the latency of DDRII ram is higher then DDRI ram would also affect the NV30 to NV35 results, meaning at the same speed DDRI ram would have an advantage.
Anyone want some overclocked results of QIII?
Now since the latency of DDRII ram is higher then DDRI ram would also affect the NV30 to NV35 results, meaning at the same speed DDRI ram would have an advantage.
Anyone want some overclocked results of QIII?
Is that what the test shows? Or does it show that by lowering the clock speed to a point where it is no longer bandwidth limited on either card, that the NV35 can push twice as many pixels.
It seems to me the numbers tell more about the pixel pipes than the bus width.
Edit: Though what Axiom says (that the memory was actually 250/500 instead of 500/1000) would suggest both to me.
It seems to me the numbers tell more about the pixel pipes than the bus width.
Edit: Though what Axiom says (that the memory was actually 250/500 instead of 500/1000) would suggest both to me.
If you wanted to prove the pixel pipe count, you wouldn't use AF + 4XFSAA. With AF the increased number of texture samples would eliminate the difference between 4x2 and 8x1 and 4X FSAA is bandwidth limited.