Cost of IMR vs. TBDR - READ DESCRIPTION before voting.

[Joe DeFuria]

Just for sh*ts and giggles...

Say we had two graphics cards today. One was a traditional IR (immdiate mode renderer such as Radeon or GeForce) and the other was a TBDR (tile based deferred renderer, such as Kyro.) Now assume they both have the exact same specifications for:

1) DirectX Feature Support
2) Raw Fill Rate (pixel and texel)
3) Raw memroy bandwidth

Keep in mind we are talking about RAW SPECS, not "effective" numbers.

What would you guess the relative COST of the two cards to be.

 

[Crusher]

I'm enough of a procrastinator to know that putting things off until later always costs you more than doing them now.

 

[Xmas]

If the raw specs should be the same I see no reason why a TBDR would be cheaper.

 

[arjan de lumens]

Depends a lot on whether the IMR has optimizations like hierarchical Z, Z/framebuffer compression, and how big its framebuffer cache is, that sort of thing, which add to both cost and performance.

 

[Joe DeFuria]

Depends a lot on whether the IMR has optimizations like hierarchical Z, Z/framebuffer compression, and how big its framebuffer cache is, that sort of thing, which add to both cost and performance.

Right, and it also depends on whether the TBDR has other optimizations beyond what the Kyro currently has. (Larger tile buffers, it's own compression, etc.)

So the only way to really address this issue, is to assume the two cards appear on the market at the same time. (Perhaps making the relevant technologies viable at similar time frames for each architecture?) You'll have to guess for yourself which architecture would have more or less "additional costs" for performance enhancing features based on them being available at the same time.

 

[Tahir2]

"Sir, I have found you an argument; but I am not obliged to find you an understanding." -- Samuel Johnson

The only question I care about at this moment in time is who is/was Samuel Johnson cos he says/said some funny things [/quote]

 

[demalion]

Well,

If the IMR has design elements to increase efficiency, for the purposes of your poll it would likely: Either cost more than the TBDR (assuming just pipe count increase and HW T&L support being the changes for the TBDR, due to overdraw elimination), or cost the same if performance enhancements were placed on the TBDR.

If the IMR does not, then it could cost the same or even less (and it would perform significantly worse) than the TBDR.

In either case, the TBDR looks like a performance/cost winner, assuming DX 9 functionality is not a factor (the fill rate and bandwidth stipulations don't address this) or has no performance penalties for the TBDR. Or, atleast, that seems the most likely outcome, IMO, except that the performance enhancements seem like a big black box that may or may not be bigger than a bread box.

The vote seems like a coin toss to me. Hmm...I guess I'll give a vote of confidence to the prowess of Kristof and Krew...

 

[Joe DeFuria]

(assuming just pipe count increase and HW T&L support being the changes for the TBDR, due to overdraw elimination),

What are you talking about? Hardware T&L support? I said, assume both cards have the same exact DirectX support. Do I need to state the same exact DirectX support "in hardware?"

In either case, the TBDR looks like a performance/cost winner,

Sigh...THIS HAS NOTHING TO DO WITH PERFORMANCE/COST!!!

The vote seems like a coin toss to me.

And yet, you have problems with me in the other thread taking the middle road?!

 

[demalion]

(assuming just pipe count increase and HW T&L support being the changes for the TBDR, due to overdraw elimination),

What are you talking about? Hardware T&L support? I said, assume both cards have the same exact DirectX support. Do I need to state the same exact DirectX support "in hardware?"

Err...yeah. I did. I assumed DX 7 support.

For cost equivalency I included HW T&L and pipe count increase as stated to try and get the complexity of design to be similar based on the design complexity of designs I know of...

In either case, the TBDR looks like a performance/cost winner,

Sigh...THIS HAS NOTHING TO DO WITH PERFORMANCE/COST!!!

What are you talking about? My comments did in fact have to do with performance/cost, that's why they began "In either case, the TBDR looks like a performance/cost winner," in response to arjan's post about performance and cost ...I wasn't aware that discussing that was verboten after arjan had mentioned it It seemd to me pertinent to the process of "guessing for myself" which would have more or less "additional costs" for performance enhancing features.

And yes, arjan, you, and I were discussing the same type of performance and the same type of cost, since hierarchical Z, Z/framebuffer compression, for example, don't affect raw fill rate, raw bandwdith, and direct x feature support, but do reflect in the cost of chip implementation.

The vote seems like a coin toss to me.

And yet, you have problems with me in the other thread taking the middle road?!

Since that other thread is a discussion of performance versus cost, as I've mentioned there...

This vote is not (as per your stipulation it is raw and not effective fillrate and bandwidth, for example), though there was a discussion that was (see above) in relation to the common factor between these two discussions (cost, defined in a way that makes sense to me).

 

[K.I.L.E.R]

Company that starts making TBDR but before they were making cards that only do IMR, the TBDR will cost more in terms of R&D ( which we aussies pay a 15% tax on for all IT goods, goodness knows what reason ) as the company will have to take a bit of time to research it before jumping on the bandwagon. Remember, time = money.

 

[Dave H]

K.I.L.E.R. brings up a good point. To flesh this out a bit:

...it depends what you mean by "cost". Does "cost" mean:

1) # of transistors (after all, same bandwidth -> same PCB costs?)
2) actual marginal cost per good GPU die from 3rd-party fab
3) marginal cost per die along with total R&D/driver development costs
4) marginal cost per die plus R&D/driver costs amortized over total sales
5) above plus IP licensing fees
6) wholesale price of GPU for board integrator
7) wholesale board price
8) retail board price


As we all know, "cost" (as calculated in certain of the above ways) will presumably vary a great deal depending on which IHV is developing these hypothetical TBDR/IMRs. Are we comparing, say, the costs to Nvidia of doing NV40 as an evolutionary IMR vs. utilizing their Gigapixel IP and moving to a TBDR? Are we comparing, say, a new PVR TBDR to a new IMR from ATI?

Offhand, I'm inclined to choose "cost = # transistors" because it is removed enough from real costs that I don't feel bad then going on to assume hypothetical equally-talented engineering teams. From there I would assume the IMR has the full complement of current bandwidth-saving/overdraw reducing technologies implemented, because honestly, none of us is too interested in a chip that doesn't. And therefore I would assume the TBDR "costs less", because I think the transistor count necessary to process scene geometry and implement a tile-sized on-die depth-sorting cache is less than the transistor count for the IMR efficiency features. Note that this assumption is based only on comparing how IMR transistor count has increased as such features have been added, to guessing how many transistors were required for the TBDR functionality of e.g. Kyro, which is to say it's based on not very much.

In the end, this conclusion is so IMO meaningless to the real-world desirability of TBDR vs. IMR that I won't bother voting until the question is clarified a bit.

PS - funny how item #8 in the list turns into the "cool shades dude" emoticon...

 

[Mariner]

Misae,

Samuel Johnson was an Englishman who collated the first english dictionary during the 18th Century. Try the following web site for more information:

http://www.samueljohnson.com

A very interesting chap with a good wit.

In response to Joe's poll - I don't see why there should be any difference in cost (assuming he's talking about money) between the two architectures.

Unless, of course, the TBDR hammered the performance of the IMR - in this case then they would obviously charge more for the better performing chip!

 

[Joe DeFuria]

...it depends what you mean by "cost". Does "cost" mean:

Fair question...difficult to answer. The "actual" answer is: "the price it would have to sell at to break even." So that would include R&D as well as part costs.

So, we have to assume that both parts sell in equal volumes.

In more practical terms it might be close to "7) wholesale board price ".

Definitely not "retail board price", because that is more dictated by supply / demand and actual performance, not cost.

 

[Dave H]

So, we have to assume that both parts sell in equal volumes.

With this in mind, I decided to ignore the fact that I don't really know the answer and instead vote for IMR costing more.

However, in the real world a GPU from one of the Big Two is going to sell a hell of a lot more units than one from a smaller IHV, even if it does have the legendary (perhaps mythological) TBDR price/performance advantage. And that will affect amortized R&D/support costs quite a bit.

 

[Simon F]

Misae,

Samuel Johnson was an Englishman who collated the first english dictionary during the 18th Century. Try the following web site for more information:

http://www.samueljohnson.com

Gosh. That site seems to say the quote I use might not be from Mr Johnson.

 

[Mariner]

I always thought that the quote you use was from Mark Twain, but I can't think where I heard this - pretty sure I've read it somewhere but I can't remember where!