Scalable graphic solutions making a comeback?

[Johnny Rotten]

Was just perusing some old Carmack comments made at slashdot and came across the following (from last december http://slashdot.org/comments.pl?sid=25432&cid=2762972 )

Dependent texture reads are the only really new thing in the last year or two (and only really got worked out right in the Radeon 8500), but next year is going to see floating point pixel formats, which was going to be one of Bali's truly important points. We should also see highly scalable boards built on consumer chips, which has been promised for years, but (with the exception of some 3dfx high end systems) not delivered properly.


I'm assuming he is refering to Creative and 3D Labs P10 architecture. But the implication (to me at any rate) is that there is more than one scalable design coming. Wonder who else he could be hinting at?

 

[Gollum]

Well, ATI still has its scalable MAXX tech, not sure if it's highly scalable or only 2-4 chip, but if I remember correctly they solved a lot of the problems their tech had with their first MAXX card since then and who knows, the R300 might be optimized to take good advantage of it.

Other than that, PVR tech proved to be scalable in the past (for the high end arcade boards at least) and nvidia chips (even current ones) are supposedly also capable of being used in multi-chip setups - maybe the next nvidia generation uses some of 3dfx' scalable tech and makes them even more suited for that.

I honestly don't see Parhelia being scalable, neither the Xabre and SavageXP, did I forget anything?

 

[duffer]

He's probably talking about the 3DLabs design, doesn't it have "n" functional units? You could imagine scaling that design so that different versions of that design having fewer or more vector units on the chip.

I doubt that we'll ever see useful multi-chip scaling again. Instead, we'll see more of the single-chip clock-rate scaling and pipeline-replication scaling that we have today.

Think of it from the chip designer's point of view: They want to make the most money for the least amount of work. Multichip designs are expensive to create -- not only in terms of chip architecture, but also pacakging, thermal and power board design, and driver design. It's much cheaper to to keep the same basic design and just play games with the clock speeds and number of internal units on the chip.

Also, look at the shape of the graphics card market. On the high-volume low-performance end, price is everything. On the low-volume high-performance end, card enthusiasts are willing to pay a large price premium for only slightly better performance. But even the card enthusiasts don't want to pay more than about $400 per board.

In the past it's been difficult for multi-chip solutions to be cheap enough to sell well. It's been much more successful to create slightly tweaked versions of the single-chip architecture, even though the tweaked versions are only somewhat faster than the high-volume versions.

As long as graphics card enthusiasts remain small in number, and uncritical in their buying habits, I expect we'll continue to see more of the same.