Are you ready ? (Nv 30 256 bit bus ?)
- Thread starter Bjorn
- Start date
Well, let's assume that it can dispatch 2 128-bit non-interdependent shader ops per cycle, and perhaps 4 (in special cases) at 64-bit precision. Then it's possible that it would take 25 cycles for a 100 op shader.
Also there is some possibility that scalar ops and texture reads (which also could as instructions) could be dispatched in parallel as well.
On the other hand, knocking your fillrate down by a factor of 25 to 100 is going to be slow no matter what unless you have good deferred shading/occlusion culling.
You have to ask yourself: If the NV30 has 125M transistors, but only a 128-bit bus, what are all the other transistors budgeted for? Framebuffer compression? ATI has it. More complex AA? ATI has it. Better Hyper-Z/LMA? ATI has it. They can't beat ATI with an equivalent set of compression/efficiency features, but with half the real bandwidth. Therefore, they must be doing something more exotic, or it is total PR BS.
Also there is some possibility that scalar ops and texture reads (which also could as instructions) could be dispatched in parallel as well.
On the other hand, knocking your fillrate down by a factor of 25 to 100 is going to be slow no matter what unless you have good deferred shading/occlusion culling.
You have to ask yourself: If the NV30 has 125M transistors, but only a 128-bit bus, what are all the other transistors budgeted for? Framebuffer compression? ATI has it. More complex AA? ATI has it. Better Hyper-Z/LMA? ATI has it. They can't beat ATI with an equivalent set of compression/efficiency features, but with half the real bandwidth. Therefore, they must be doing something more exotic, or it is total PR BS.
While I usually aplause NVIDIA on most of the things they're doing, Kirk is bullshitting again... 
I don't understand how could anyone conclude from what he said that NV30 will have a 256-bit memory bus (i'm not confirming nor denying whether it's 256-bit or 128-bit, it's just weird that one would conclude it is indeed 256-bit from what he said).
I would conclude from what he said that NV30 uses the fastest memory available today (as rumours have been saying - DDR-II, and 500mhz I assume), while also having very effective bandwidth saving techniques, thus, the effective bandwidth statement.
Kirk didn't talk about physical bandwidth, which could further mean that NV30 will indeed have a 128-bit memory bus.
Think of it like that:
R300 has a 256-bit memory bus. Assuming that NV30 is indeed 128-bit and also assuming it's as effective as rumours have been saying, it would be fairly intresting to see how things develop... ATI and NVIDIA, both took a different path to follow. Perhaps NV30 will be faster than R300 (by quite a large margin) on certain apps, but slower and on par on others? (i assume aa & aniso perfomance, since benchmarking without them doesn't make sense... at least until DOOM III comes out...). That would put NVIDIA in a pretty good position, where they could sell the high-end NV30 part at approximately the same price as R9700pro costs right now and also offer higher perfomance. That would also leave them plenty of room for expanding with NV35, where they could use a 256-bit memory bus. After all, if they could beat R300 with 128-bit, what sense would be in using 256-bit? (except making the chips more expensive to make, which is not good).
I for one wouldn't care one bit if NV30 would indeed utilize a 256-bit memory bus, as long as it's perfomance is greater or on par with R300 (the former being more valuable to me ).
P.S
There are some rumours floating around with some really interesting details, quite suprising to say the least, but I'll save that till after NV30 is announced...
I don't understand how could anyone conclude from what he said that NV30 will have a 256-bit memory bus (i'm not confirming nor denying whether it's 256-bit or 128-bit, it's just weird that one would conclude it is indeed 256-bit from what he said).
I would conclude from what he said that NV30 uses the fastest memory available today (as rumours have been saying - DDR-II, and 500mhz I assume), while also having very effective bandwidth saving techniques, thus, the effective bandwidth statement.
Kirk didn't talk about physical bandwidth, which could further mean that NV30 will indeed have a 128-bit memory bus.
Think of it like that:
R300 has a 256-bit memory bus. Assuming that NV30 is indeed 128-bit and also assuming it's as effective as rumours have been saying, it would be fairly intresting to see how things develop... ATI and NVIDIA, both took a different path to follow. Perhaps NV30 will be faster than R300 (by quite a large margin) on certain apps, but slower and on par on others? (i assume aa & aniso perfomance, since benchmarking without them doesn't make sense... at least until DOOM III comes out...). That would put NVIDIA in a pretty good position, where they could sell the high-end NV30 part at approximately the same price as R9700pro costs right now and also offer higher perfomance. That would also leave them plenty of room for expanding with NV35, where they could use a 256-bit memory bus. After all, if they could beat R300 with 128-bit, what sense would be in using 256-bit? (except making the chips more expensive to make, which is not good).
I for one wouldn't care one bit if NV30 would indeed utilize a 256-bit memory bus, as long as it's perfomance is greater or on par with R300 (the former being more valuable to me
).P.S
There are some rumours floating around with some really interesting details, quite suprising to say the least, but I'll save that till after NV30 is announced...
LeStoffer said:
Tell me about it. I was actually hoping it would have a 128-bit bus and 4 pipelines with a radical departure in how it's rendering. Their's plenty of preformance to be had by making your design more effecient--not just increasing the 'big number/nomenclature' that people on the internet can run around and use to engage in yet another cock-measuring contest.
The fact Nvidia and ATI both may employ the very same techniques to save bandwith and clock cycles and thus Nvidia can't beat ATI on this path, it's not so obvious, imho.DemoCoder said:
Despite the fact Nvidia had/has a less efficient hierarchical zbuffer implementation than ATI, in term of efficiency I believe Nvidia has beat ATI all the times on the IMR efficiency front.
With R300 ATI made a very big jump and surpassed Nvidia, but I expect Nvidia to get the efficiency crown with the NV30. Maybe they will employ some exotic techinique as you pointed out, or maybe they will just improve the current methods, dunno.
As was said in other discussions on this forum probably there is a lot to do to increase IMRs efficiency.
Just to make an understanding, I believe that on the AA front there is a lot of room for improvements.
However...we have to wait just a couple of weeks to know.
ciao,
Marco
well look at it this way. if nvidia releases the nv30 with 500mhz ram and a 128bit bus in january-febuary time period. Ati could then in febuary-march time period put out a new video card using a 256bit bus and 500mhz ram (or faster) all while having a 5-6 month period between highend parts. Nvidia's only option would be to ride it out with the nv30 as the high end or release a faster speed version of the nv30 pissing off all the people who spend 400 or whatever a month or two before on the slower nv30.
Yes I am sure that the nv35 will. But when can they release it? I bought the 9700pro in sept. I spent 400 bucks on it. Now you can bet i would never buy another ati product if a month later they released a video card faster than mine at the same price. Now if ati responds to the nv30 with a r300 that is faster than the nv30 buy a small amount or even a r350 with a 8x2 set up and all the extras that the nv30 will have over the r300 and more than what will nvidia do. If they are trumped right after the launch of the nv30 how can the answer that. Would we really see a nv35 or a speed bin of the nv30 out just a month or two after the normal nv30? How much would that piss of nvidia fans or will they just bend over and take it like many did with the geforce sdr and ddr crap ?
It all comes down to how would ATI respond and how powerful of a response that would be...
I'd have to agree that they respond the way you anticipate (or whatever it is you're doing there ), it would hurt NVIDIA greatly, as bringing in NV35 shortly after NV30 wouldn't fit well into NVIDIA's plans... (hell, NV30 isn't supposed to be put to mass production until January or so...).
I'd have to agree that they respond the way you anticipate (or whatever it is you're doing there
), it would hurt NVIDIA greatly, as bringing in NV35 shortly after NV30 wouldn't fit well into NVIDIA's plans... (hell, NV30 isn't supposed to be put to mass production until January or so...).Well maybe they will have an NV30 Ultra prepared as well to top anything from ATi that they will not release before any R350's are announced. Although that would be risky, as they might not be able to top everything ATi could cook up.
A more realistic countermeasure would be a dual-chip NV30 board. Theoretically even the GF2 has multichip support - someone mentioned that Nvidia was ready to defeat the V5-6000 with such a beast back then. It would be a ridiculous beast for sure, with insane power requirements and price, but it only had to be faster than any new Radeons. We all know that the real business is in the mass market, and highend parts only matter for PR - so even such a card would be enough to save their face...
A more realistic countermeasure would be a dual-chip NV30 board. Theoretically even the GF2 has multichip support - someone mentioned that Nvidia was ready to defeat the V5-6000 with such a beast back then. It would be a ridiculous beast for sure, with insane power requirements and price, but it only had to be faster than any new Radeons. We all know that the real business is in the mass market, and highend parts only matter for PR - so even such a card would be enough to save their face...
GF2's and on are supposed to have AGP bridges built in, however they've never really functioned properly AFAIK. I think the V5 6000 reply would have been GF2 Ultra, however seeing as 6000 never materialised they held it back and released it once they realised NV20 wouldn't be on time.
From a conversation I've had with someone at ATI I get the impression that there area already multi-chip variants of R300 out there -- however not anything sensible for the consumer, but larger workstation boxes with many chips for DCC uses.
From a conversation I've had with someone at ATI I get the impression that there area already multi-chip variants of R300 out there -- however not anything sensible for the consumer, but larger workstation boxes with many chips for DCC uses.
Johnathan256
Newcomer
Honestly, what does the amount of raw bandwidth matter if NV30 can work it's magic with less? David Kirk said:
Does anyone think that he would have said that if he knew the competition's hardware was on par already if not better? What does it matter what kind of memory interface NV30 has? Sure we would all prefer 256-bit, but if it doesn't need it to perform to its potential then so what?
Go ahead and flame away.
Does anyone think that he would have said that if he knew the competition's hardware was on par already if not better? What does it matter what kind of memory interface NV30 has? Sure we would all prefer 256-bit, but if it doesn't need it to perform to its potential then so what?
Go ahead and flame away.
There is no way that NVidia can beat ATI by shipping cards that have faster commodity DRAM. Why? Because any fancy DDR/DDRII memory, no matter how fast, ATI can buy the same memory and put it in their cards.
So it doesn't matter if NVidia's game is to ship a card with 500Mhz DDR-II memory. ATI, or third party OEMed, could reply instantly with a R300 variant using the faster memory. If it requires a silicon tweak, it would be trivial for ATI to change the memory controller and ship an R310 or whatever.
It is way easier for ATI to adopt new commodity memories than it is for NVidia to go from 128-bit to 256-bit bus. It virtually requires a redesign of the memory controller and system.
I noticed that picture of the NV30 "chip" with a dense pin array as well. Who knows? I am almost hoping that Nvidia has some exotic deferred shading architecture. Another brute-force 256-bit IMR would be, well, predictable.
Of course, a 256-bit deferred shading IMR would be even better.
So it doesn't matter if NVidia's game is to ship a card with 500Mhz DDR-II memory. ATI, or third party OEMed, could reply instantly with a R300 variant using the faster memory. If it requires a silicon tweak, it would be trivial for ATI to change the memory controller and ship an R310 or whatever.
It is way easier for ATI to adopt new commodity memories than it is for NVidia to go from 128-bit to 256-bit bus. It virtually requires a redesign of the memory controller and system.
I noticed that picture of the NV30 "chip" with a dense pin array as well. Who knows? I am almost hoping that Nvidia has some exotic deferred shading architecture. Another brute-force 256-bit IMR would be, well, predictable.
Of course, a 256-bit deferred shading IMR would be even better.
I'd read into this a 128 bit bus too, much as I would like it to be 256 bit with a 32GB/second or higher bandwidth, with the LMA III giving a 50% boost to give an effective 48GB/sec bandwidth. After all the hype and speculation 16GB/sec or thereabouts seems a bit lame. If Nvidia says twice as fast as a GF4 I'd expect at least 20GB/sec, if not more. Given they are now playing the 'effective' game, maybe 14GB sec + 50% gives 20GB/sec == twice as fast as GF4.
I guess in 9 days we should all know and the speculation will die down as we wait for meaningful benchmarks.
But whatever, I see R300 and NV30 may be a step function in graphics cards. From here on in things may be so powerful that everything now looks great. It may be R350 or NV35 are the cards that really deliver everything, even so they should only be 6-12 months away. I can wait until they get it right. I have 1 GF3, 1 GF2 GTS, 3 GF2 MX400s and a GF1 DDR. I will wait until things are great before I move again to upgrade.
Fingers crossed 18th Novemeber is like an early Christmas.
I guess in 9 days we should all know and the speculation will die down as we wait for meaningful benchmarks.
But whatever, I see R300 and NV30 may be a step function in graphics cards. From here on in things may be so powerful that everything now looks great. It may be R350 or NV35 are the cards that really deliver everything, even so they should only be 6-12 months away. I can wait until they get it right. I have 1 GF3, 1 GF2 GTS, 3 GF2 MX400s and a GF1 DDR. I will wait until things are great before I move again to upgrade.
Fingers crossed 18th Novemeber is like an early Christmas.
g__day said:
Funny you should mention step functions. On nVidia's website, they have slides from the Art Futura Conference. The slides are in Spanish. But one slide showing the convergence of real time rendering and offline rendering shows the combination of nv30 and cG to be a step function in real time rendering.
The link to the slides on the nVidia website:
http://developer.nvidia.com/view.asp?IO=ArtFutura_Conf (first set)
megadrive0088
Regular
at this point, I'm willing to accept that Nv30 most likely has a 128-bit bus, and that it has extremely good bandwidth saving features. I just hope the Nv35 moves upto 256-bit.
not to be crude, but frankly, Nvidia needs a wider bus..... to shove all that bull sh_t they've been feeding everyone, like rendering FF and GeForce"4"MX, back where it belongs
not to be crude, but frankly, Nvidia needs a wider bus..... to shove all that bull sh_t they've been feeding everyone, like rendering FF and GeForce"4"MX, back where it belongs
