Geforce FX question

[BoardBonobo]

With the delays in the launch of the GF FX being roughly equal to the time it usually takes them to do a refresh part, what are the chances that they have developed a 256bit memory bus version for the high end.

How hard would it be to redesign the crossbar for this and would it be feasible for nVidia to have doen in the time they have had?

 

[KimB]

256-bit bus? Almost nil. Anyway, it would be a good idea to judge whether or not nVidia needs to move to a 256-bit bus only after seeing the GeForce FX in action.

 

[sas_simon]

What I would like to know is if they scrapped the nv35 and will go straight to the nv40 in autumn time of this year.

 

[Hellbinder]

256-bit bus? Almost nil. Anyway, it would be a good idea to judge whether or not nVidia needs to move to a 256-bit bus only after seeing the GeForce FX in action

I cant even understand this statement.

You do realize that if R300 had 500mhz ram it would absolutly CRUSH the NV30 in FSAA at High resolutions.. of course you do. There are only so many ways to do FSAA with a large number of Samples. Unless you employ some form of Pure Edge AA similar to FAA you are going to need bandwidth.

As it is with the Ram that Nvidia is using they are only about 3GB behind which is simply not goiing to impact them that much in their current situation. They will have a big advatage in AF simply do to their Fillrate advatage. However,, It is PLAINLY obvious that the R350 is going to be pushing a minimum of 25GB bandwidth and possibly 30+GB. Combined with higher fill rate there is simply NO WAY that the Nv30 is going to compete with that kind of a Bandwidth defecit. I dont see how you can even debate that.

Ram Speed is simply not going to improve enough by the end of this year to be able to make them competative with a 128bit bus. Its just that simple. ill bet 20$ right now as it is that the 9700pro beats the Nv30 is most apps with pure 4x FSAA, and i would watch VERY closely what Nvidia is trying to pass off as 4x FSAA when it comes time to compare.

 

[JD]

I think it will depend on Ati's next card, then nv might release 256bit bus to be competitive. Though this would be a case when fill rate is the predominant factor in rendering pipeline and with doom3/shadow volumes this might as well be the case. The gffx offloaded lots of stuff from fixed function pipeline so the bottlenecks could be elsewhere depending on the game played. So 9700 with its higher bandwidth than gffx might pull ahead while gffx with its pixel shader engine might be faster and vice versa type thing.

 

[Ostsol]

I think that it is indeed quite assured that NVidia's next generation of video cards will have a 256 bit bus. The GeforceFX already -requires- the huge copper heat-spreader across the back of the card to prevent the RAM from getting too hot. There may be a way to correct the issue in future iterations of DDR-II, but it is likely far easier to increase memory bandwidth by going to a 256 bit bus than trying to push more speed out of RAM -- especially when heat becomes an issue.

 

[Sage]

256-bit this, 128-bit that, 16GB this, 19GB that...... these are all the absolute max burst speed! what matters is how much they can actually get out of that 16 / 19GB. DDRII has higher latency which would mean that at the same clock speed, it would be slower than DDRI.... if that was all that mattered. The fact is there is a whole lot of factors determining how much bandwidth you can actually use. if you want something to read, go to http://arstechnica.com/paedia/b/bandwidth-latency/bandwidth-latency-1.html

 

[KimB]

]You do realize that if R300 had 500mhz ram it would absolutly CRUSH the NV30 in FSAA at High resolutions.. of course you do. There are only so many ways to do FSAA with a large number of Samples. Unless you employ some form of Pure Edge AA similar to FAA you are going to need bandwidth.

But it would be very hard to build a Radeon 9700 with a 256-bit bus that operates at 500MHz, and very expensive (lots of pins + high clock means lots of noise). I still think that if we see a R300 derivative with 500MHz DDR2, it will be using a 128-bit bus.

Btw, when I said "next to zero" I was speaking of the GeForce FX. The probability of the NV35 having a 256-bit bus is quite a bit higher, but I still don't think it's likely.

As it is with the Ram that Nvidia is using they are only about 3GB behind which is simply not goiing to impact them that much in their current situation. They will have a big advatage in AF simply do to their Fillrate advatage. However,, It is PLAINLY obvious that the R350 is going to be pushing a minimum of 25GB bandwidth and possibly 30+GB. Combined with higher fill rate there is simply NO WAY that the Nv30 is going to compete with that kind of a Bandwidth defecit. I dont see how you can even debate that.

Yes, there is a way. While I will admit that it isn't looking very likely right now, there is still the possibility of the GeForce FX's bandwidth savings methods being that much better. But, it will be very hard for nVidia to fight against ATI's much more mature drivers in the R350. (ATI's drivers for the Radeon 9700 seem pretty poor right now. If they can significantly improve them by the time the FX is released, that's the only thing that I see as giving nVidia lots of trouble)

Ram Speed is simply not going to improve enough by the end of this year to be able to make them competative with a 128bit bus. Its just that simple. ill bet 20$ right now as it is that the 9700pro beats the Nv30 is most apps with pure 4x FSAA, and i would watch VERY closely what Nvidia is trying to pass off as 4x FSAA when it comes time to compare.

It's not about RAM speed. It's about not using the bandwidth in the first place.

 

[KimB]

I think that it is indeed quite assured that NVidia's next generation of video cards will have a 256 bit bus. The GeforceFX already -requires- the huge copper heat-spreader across the back of the card to prevent the RAM from getting too hot. There may be a way to correct the issue in future iterations of DDR-II, but it is likely far easier to increase memory bandwidth by going to a 256 bit bus than trying to push more speed out of RAM -- especially when heat becomes an issue.

I'm not so sure. I do still think that this 256-bit bus thing will fade in time, just like the multi-chip boards of the past didn't last very long. It's not economical. I think that the signal noise in the 256-bit bus parts will prevent them from really getting that close to the frequency of their 128-bit bus counterparts, which should help to keep the 256-bit bus parts from really dominating.

Another item of importance is what we've seen nVidia state in the past, that the focus is shifting off of memory bandwidth and towards computational efficiency. That is, if NV30-style shaders indeed require more processing power compared to memory bandwidth, then there won't really be a need to move to a 256-bit bus.

 

[Ichneumon]

Another item of importance is what we've seen nVidia state in the past, that the focus is shifting off of memory bandwidth and towards computational efficiency. That is, if NV30-style shaders indeed require more processing power compared to memory bandwidth, then there won't really be a need to move to a 256-bit bus.

Of course they stated that! They don't HAVE a 256bit bus. They would be commiting PR suicide to say that a 256bit bus is good when they don't have one!!

We haven't reached the point of enough bandwidth to do "good enough" FSAA and advanced filtering at "good enough" speeds across the board. We're almost there... but we're not there yet. I don't expect the 256bit bus to go away anywhere soon on bleeding-edge cards.

You talk about the complexity of having a 256bit bus on a card... What about the complexity of dealing with overheating ram chips on a 128bit bus? They're both technological problems that will be solved in time with advancements to each tech. Don't assume that one is necessarily easier than the other... No one thought the 9700 could possibly be clocked at the speeds it is on .15 proces... and yet here it is. And it was done 6 months before Nvidia could figure out how to make DDR2 work on a 128bit bus with their problem chip.

 

[KimB]

Of course they stated that! They don't HAVE a 256bit bus. They would be commiting PR suicide to say that a 256bit bus is good when they don't have one!!

Yes, that is a definite factor in coming forward with a statement like that. But it still doesn't mean they're wrong. For example, the one texture per pixel pipeline of both the R300 and NV30 almost makes it seem like both companies agree with this.

You talk about the complexity of having a 256bit bus on a card... What about the complexity of dealing with overheating ram chips on a 128bit bus?

And a 256-bit bus will have to deal with that in the exact same way.

And it was done 6 months before Nvidia could figure out how to make DDR2 work on a 128bit bus with their problem chip.

The problem looks to have been .13 micron process issues, not anything to do with RAM.

 

[Ichneumon]

And a 256-bit bus will have to deal with that in the exact same way.

The problem looks to have been .13 micron process issues, not anything to do with RAM.

Yes, a 256bit bus will have to deal with ram heat problems, but it doesn't *NEED* ram AS fast as a 128bit bus to acheive its greater bandwidth.

Certainly .13 process was the primary contributor to the FX delays... but is DDR2 speeds and heat problems going to be resolved faster than someone else can come up with a way to make a 256bit bus even more economical? I expect a lot will depend on where companies want to be spending money... on expesive ram, or on chip/board design.

You don't think that dustbuster cooler and ramsinks costs money on the FX? So Nvidia is foisting off costs for making FX boards to its board manufacturers who have to include that rediculous cooling solution, which comes out of Their margins. ATI manufacturers have costs too, but much of that has already been absorbed by ATI in the design making the 256bit bus possible.

You don't think that 3rd party manufacturers see things like that? There's good reasons why 3rd parties have been flocking to ATI lately.

 

[Pete]

The problem looks to have been .13 micron process issues, not anything to do with RAM.

Ich meant that ATi pulled two rabbits out of its hat: a performance leader with both a supposedly-too-expensive-and-topped-out .15 process and a supposedly-too-expensive-and-complicated 256-bit memory bus.

Whatever, I'm tired of speculation and theoretical specsmanship. I'm ready for the reviews. The ideal scenario for a cheapskate gamer like me would be a high-performing GF|FX trumped a month later by a faster R|10K. Let the price wars ... begin! For real this time.

 

[Ichneumon]

The problem looks to have been .13 micron process issues, not anything to do with RAM.

Ich meant that ATi pulled two rabbits out of its hat: a performance leader with both a supposedly-too-expensive-and-topped-out .15 process and a supposedly-too-expensive-and-complicated 256-bit memory bus.

Thank you. You interperet me better than I write me.

 

[Dave Baumann]

You talk about the complexity of having a 256bit bus on a card... What about the complexity of dealing with overheating ram chips on a 128bit bus?

And a 256-bit bus will have to deal with that in the exact same way.

Errr, uh

I think another point that was being made is that this isn't the case, since by going 256-bit you can still get more bandwidth by utilising standard chips that don't require any cooling.

 

[Evildeus]

The problem looks to have been .13 micron process issues, not anything to do with RAM.

Ich meant that ATi pulled two rabbits out of its hat: a performance leader with both a supposedly-too-expensive-and-topped-out .15 process and a supposedly-too-expensive-and-complicated 256-bit memory bus.

I'm sorry, but it doesn't say anything on the cost effective part, it just say that you can do it, not at what cost

 

[Sharkfood]

But it would be very hard to build a Radeon 9700 with a 256-bit bus that operates at 500MHz

How so? Several vendors are already selling R300's at 400-405mhz with the current core, unmodified. The totally silent Heat-pipe solutions are all running at 400+ mhz already.

I can't imagine how it would be "very hard" to get either a modified R300 core or .13u RX00 core to 500mhz with the same 256 bit bus.

It all comes down to the quality/price of memory at this stage. This seems to be the major stiffling point moreso than anything else.

 

[Dave Baumann]

But it would be very hard to build a Radeon 9700 with a 256-bit bus that operates at 500MHz, and very expensive (lots of pins + high clock means lots of noise).

GeForce FX and Radeon 9700 both feature the same number of RAM chips, so from the RAM side they both have the same number of pins and traces. 9700 only uilises a 8 layer PCB, unlike GFFX's 10 (at the moment) so they still have more to play with on the PCB side before it gets to be signicantly more complex or expensive than GeForceFX.

 

[Fuz]

I'm not so sure. I do still think that this 256-bit bus thing will fade in time.

I have to say that I totally disagree with you on this one. I think its here to stay. Nv will jump on the 256-but bus, er... I mean bit bus soon enough. My guess, we will see a 512-bit bus in the next 3 years.

 

[Tagrineth]

I'm not so sure. I do still think that this 256-bit bus thing will fade in time, just like the multi-chip boards of the past didn't last very long. It's not economical. I think that the signal noise in the 256-bit bus parts will prevent them from really getting that close to the frequency of their 128-bit bus counterparts, which should help to keep the 256-bit bus parts from really dominating.

Hmm... I wonder how many people downplayed the 128-bit bus when it started replacing 64-bit?

Also only two companies had multi-chip boards, really... 3dfx and ATi.

ATi's boards had problems with WinNT kernels, and around R200 they fixed the problem... but didn't need multichip to compete at the time And they definitely don't right now.

And 3dfx.. well, they're kinda, uh, too dead to compete at the moment.