Opinions needed on this Interview

[trinibwoy]

And, if this is the case then the comparison should be SFR against tiling.

Do you know enough about each to make that comparison for us Dave? How valid is Chris' assertion that redundant texture fetches at tile boundaries will waste significant bandwidth?

 

[Dave Baumann]

That would entirely depend on the architecture. For instance, as best I can tell NVIDIA's rendering scheme appear to be allocated according to availability (i.e. the next available quad gets the next work dispatched to it), in which case there probably would be a lower efficiency overall for Tiling across chips for GF6 (as, at a quad level, the next work to be dispatched is likely to be within similar localities as the work on the other quads, hence texture fetch hits from their L2 texture cache of that new work might still be high). However, ATI's dispatch methodology is already tiled - each of the quads in the chip are assigned tiles when the render target is initialised, so that texture locality issue is already built into their rendering scheme and "super tiling" over that doesn't really introduce any extra inefficiencies from a texture locality perspective.

 

[Jawed]

It's also worth remembering that ATI can vary the size of the tiles.

I believe the argument about textures being required on both cards is null.

Supertiling also has a problem with over-fetching textures (as did scanline interleaving) because of the number of textures that cross the tiling boundaries. This means that each GPU would have to fetch the same texture for neighboring tiles. In SFR, we have one edge where textures can cross boundries. In supertiling, you have many, many more edges and the problem is multiplied linearly with the number of edges. These double-fetched textures can eat up valuable bandwidth.

The reason I say this is that textures have a lifetime in graphics card local memory that's longer than 1 frame. Since the on-die texture caches are tiny, pretty much all texturing across a frame is done from local memory rather than caches (i.e. the cache supports repeated accesses, not the first access). So if all textures end up on both cards (when wouldn't they?...), I don't think it's valid to say this is a bottleneck for supertiling.

But I do think the geometry duplication concern is entirely valid. Eric admitted to this. But ATI, for whatever reason, already has a comfortable lead on geometry. With unified shading hardware that'll prolly increase.

Finally, I don't see why ATI is barred from using AFR for benchmarking. In fact, sadly, we might find that it's the effort required to implement AFR that's held-up MVP.

Unless ATI wanted to hold MVP back until an ATI dual-PEGx16 mobo was ready to make a double-splash.

Jawed

 

[Dave Baumann]

It's also worth remembering that ATI can vary the size of the tiles.

Not entirely sure about this, but I think the sizes only make sense within the bounds of the granularity of HierZ (although the larger the better is probably the case here).

 

[Jawed]

Super-tile size seems to be one of those parameters that you'd tweak very carefully. Sort of similar to how ATI tweaks the memory differently for 16-pipe and 12-pipe cards.

Jawed

 

[JoshMST]

Wow, I slept like absolute crap last night. I kept waking up in a cold sweat with visions of Dave/Rys/Reverend and others all mercilously hounding my article! Last time I release an article that late at night! I might need a vacation from this stuff!

Anyway, there is absolutely no reason why ATI can't support AFR. Now, supposedly NVIDIA does have silicon built into the nForce 4 Ultra/SLI that handles the SFR/AFR calls, and like Chris said in the article, there isn't supposed to be any driver overhead for those methods (which I actually find a bit strange considering the lower resolution performance of a SLI setup). Anyway, I think it would be silly for ATI to rely on a supertiling solution all the time when AFR would work great in many other situations.

I think that Chris' comments about wasting of bandwidth to texture accesses is a bit overstated here, and I really think that both NVIDIA's and ATI's solutions will suffer much the same penalty when using tiling/SFR. As for geometry scaling... in AFR it is not a issue (for either side obviously), but I wonder how much of an issue it is for NVIDIA's SFR? Sure, each card splits half the frame, but even in SFR does each card have the full geometry content loaded up for that frame, or just for the 1/2 of the frame they are working on?

Definitely many questions are not answered here, and unfortunately I did not ask these questions. Funny how it works that you think of better questions once the interview is over!

 

[Geo]

By the way, what do y'all make of Demirug's speculation here?:

http://www.beyond3d.com/forum/viewtopic.php?t=22018&postdays=0&postorder=asc&start=7

 

[digitalwanderer]

Funny how it works that you think of better questions once the interview is over!

Try putting them together and submitting 'em to nVidia for a follow-up interview, those are always popular.

 

[neliz]

Unless ATI wanted to hold MVP back until an ATI dual-PEGx16 mobo was ready to make a double-splash.
Jawed

In my opinion it would only be LOGICAL to demonstrate a complete producte (MVP board+graphics.)
I'm pretty sure Ati'has an advanced southbridge that, besides the 1 or 2 Peg lanes for the integrated graphics will also have some sort of QOS (PCI latency?) built in to make sure that the component with the biggest bandwidth need actually receives it as well...

 

[Geo]

Btw, how do we feel about that AGP 3.0 question/answer? Frankly, I didn't know that the spec supported two cards --but somehow I'm uncomfortable with NV pointing at bandwidth limitations on the AGP bus as the deciding factor in that era. I recall there being no advantage seen initially to 8x vs 4x in benchmarks. Doesn't seem more likely that a major part of the timing to the move to SLI when it did was at least as much a function of NV's own growing strength in the chipset market making it feasible for them to lead the move from both angles --cards and mobos?

 

[Rys]

Btw, how do we feel about that AGP 3.0 question/answer? Frankly, I didn't know that the spec supported two cards --but somehow I'm uncomfortable with NV pointing at bandwidth limitations on the AGP bus as the deciding factor in that era. I recall there being no advantage seen initially to 8x vs 4x in benchmarks. Doesn't seem more likely that a major part of the timing to the move to SLI when it did was at least as much a function of NV's own growing strength in the chipset market making it feasible for them to lead the move from both angles --cards and mobos?

There's no advantage seen in games because they don't do GPU-to-host writebacks, across that direction of the bus. Texture and data uploads on AGP are fairly quick, it's just getting data back off the GPU that's slow.

Depending on the traffic they'd want back from the hardware, AGP might well be too slow.

 

[Geo]

There's no advantage seen in games because they don't do GPU-to-host writebacks, across that direction of the bus. Texture and data uploads on AGP are fairly quick, it's just getting data back off the GPU that's slow.

Depending on the traffic they'd want back from the hardware, AGP might well be too slow.

Ah, good distinction. Thanks. Tho if most of the cross-talk is across that inter-card connector? You'll note they later point at that as a major advantage to their system, not stressing the bus with the crosstalk.

 

[JoshMST]

Oh man, when I started posting last night I was pretty sick and doped up on cough syrup. I think I misrepresented some things, and it really makes me look like an ass to NVIDIA and Chris Daniel.

Here is what I should have said instead of "I look like a mouthpiece to NVIDIA... they didn't answer all my questions... etc."

I was approached by NVIDIA to do an interview, and this was honestly more as a favor to me and helping to get my website up in traffic. I have known BB for years, and whenever we see each other at trade shows and whatnot, we typically step away from the writer/PR guy and have a really good time. So, for them coming to me was a huge favor, and a show of some confidence.

I then prepared a list of questions that I wanted submitted to Chris. BB looked over these and found that quite a few questions were undesirable. These reasons were not that they were pointedly anti-NVIDIA, or asked embarrasing questions, but rather that they tried to pry answers into "unannounced products" or were directly comparing NVIDIA's stuff to "ATI's unannounced products" that they probably know a lot about (since they all use TSMC as a foundry partner, ATI and NVIDIA know alot about their competitor's roadmaps and upcoming products).

We then worked on several revisions of these questions. Again, the guys at NVIDIA were not terribly interested in getting into a pissing contest with ATI, and I can't blame them one bit. Sure, we all would have liked to have seen a "So what does NVIDIA's SLI have over ATI's MVP". However, since that is a competitors unannounced product, they can't comment on it. So we worked around that and we got in the "supertiling" questions.

I really should have been a lot more political about this, and just simply explained this as "my first several revisions of questions were handed back to me because they were out of line most of the time, and the interview would have been full of "we don't talk about unannounced products"". Having such a interview was a big thing for my little old site, and my previous comments were posted in haste and delusion. Yes, interviews give these companies a chance to talk about their products in a very positive light So, I really do not feel like a mouthpiece here, and I think that after the work I did to get them to answer that many questions, especially the ones dealing with supertiling, was a pretty good feat. Instead of working with me, NVIDIA could have just handed the questions back to me as is with "we can't talk about those things due to their status as unannounced products". As it was, the guys at NV worked with me to edit my questions so that we could get down to issues that they could talk about, and give the readers some very solid information about the product and how it compares to "other possible technologies" and where SLI is heading (all without having to say "we don't talk about unannounced products").

Gah, drugs, ill health, no sleep, a 2 year old kid that refuses to go to bed, and a plethora of other issues that have rained down upon my life outside of PenStar have all contributed to some ill-thought posts by me. I really misrepresented myself in those first comments, and I apologize to everyone here. Piss poor tactic by me. Remind me not to take heavy duty cough syrup late at night then force myself to stay up late and post here.

 

[jb]

Josh,

its not a bad thing as you brought a some good info. You just have to understand that it helps NV as well as they are going to spin anything off in the best light they can. But its intresting to read these things and compare them to some of the recent ATI interveiws as the style of the answers are almost night and day..

Thanks again!

 

[trinibwoy]

Instead of working with me, NVIDIA could have just handed the questions back to me as is with "we can't talk about those things due to their status as unannounced products". As it was, the guys at NV worked with me to edit my questions so that we could get down to issues that they could talk about, and give the readers some very solid information about the product and how it compares to "other possible technologies" and where SLI is heading (all without having to say "we don't talk about unannounced products").

And they say Nvidia isn't a friendly place Don't be so rough on yourself man - I think everyone agrees that the interview was valuable. At least you know the next one will be even better minus the cough syrup and all that

 

[trinibwoy]

the style of the answers are almost night and day..

What's the difference? I don't know where to find the recent ATI interviews else I'd check myself!

 

[DemoCoder]

That would entirely depend on the architecture. For instance, as best I can tell NVIDIA's rendering scheme appear to be allocated according to availability (i.e. the next available quad gets the next work dispatched to it), in which case there probably would be a lower efficiency overall for Tiling across chips for GF6

The NVidia employee was talking about the inefficiency of supertiling vs SFR, not the inefficiency of ATI supertiling vs NVidia tiling. (does ths latter even exist?) The issue is that coarse division of the scene over cards is likely to produce less texture fetch duplication than fine-grained division.

 

[JoshMST]

Oh man, I just read through some of my responses from last night again... I never even approached NVIDIA about previous AFR solutions and the ATI MAXX!!! I feel like I just had a out of body experience! I am literally horrified. I am glad I was not at a bar or something last night trying to get a date. What the hell would have come out of my mouth?

 

[DemoCoder]

I don't think there is anything wrong with what you did. Stop beating yourself over the head. And don't let overly criticial and biased people dictate the way you conduct yourself.

 

[Dave Baumann]

The NVidia employee was talking about the inefficiency of supertiling vs SFR, not the inefficiency of ATI supertiling vs NVidia tiling. (does ths latter even exist?)

Errr, yes.

The issue is that coarse division of the scene over cards is likely to produce less texture fetch duplication than fine-grained division.

And the effects of that depend on the architecture at hand, as I said in the replay.