[Jawed]

Originally Posted by geo

I'm not saying there aren't reasons to use clock domains. It's just that in all the great many times that Fast14 has come up --and it's come up on these boards regarding every new ATI design since it was licensed-- this is the first time that I can recall where it's suddenly become a necessity of its use that clock domains be a required feature for us to be looking for as evidence that Fast14 was included. Maybe I just have a really bad memory on that point.

My point was merely that it's most practical to concentrate the ultra-extreme clocks in places where the maximum benefit is derived. We just don't know how to work out that benefit in detail, though

There's no reason why similar reasoning doesn't apply to R600 as it does to G80. It's just a pity we're so in the dark, lol. Ironic we "know" more about it because of what Orton's said than just about any other single source.

Jawed

[Razor1]

Originally Posted by geo

I'm not saying there aren't reasons to use clock domains. It's just that in all the great many times that Fast14 has come up --and it's come up on these boards regarding every new ATI design since it was licensed-- this is the first time that I can recall where it's suddenly become a necessity of its use that clock domains be a required feature for us to be looking for as evidence that Fast14 was included. Maybe I just have a really bad memory on that point.

It doesn't need to be a requirement,

I'm leaning towards scalers on the r600 because a rumor like scalers just doesn't pop up out of no where specially since that rumor was well before the g80 news, any news or rumors.

If for arguement sake that rumor that mintmaster heard was just to throw off the scent of the g80 and laid down on the r600, and this 1 ghz rumor is for real, wouldn't 64 or 96 vector ALU's be more then sufficient to give the r600 enough shader power to match its 512 bit bus? Doesn't matter if the vector units aren't fully utilized either. At 750 the r600 would have enough theoretical performance to match the g80 with 64 vec 4 ALU's, at 1 ghz it should be able to get a substatial lead.

Although this kind goes out of the way of "more" programming flexibility then scaler ALU's, thats why the scalar ALU's are making alot more sense once the marketing stuff is in there.

[Jawed]

Jawed

[Arun]

Originally Posted by Jawed

If you can reconcile that with performing fp16 or fp32 texture filtering, then why not, eh?...

Somehow, I completely fail to see what you mean. You aren't assuming fp16/fp32 filtering is done via loopback, are you? :s


Uttar

[Jawed]

Originally Posted by Uttar

Somehow, I completely fail to see what you mean. You aren't assuming fp16/fp32 filtering is done via loopback, are you? :s

Nope, I'm not assuming anything.

I'm waiting patiently for the revised G80 architecture article - when you guys have dug some more.

Jawed

[neliz]

Uhm.. X2100 is Alive?

Saw M71-M when looking at the pcisig list http://www.pcisig.com/developers/com...ors_list/pcie/ to confirm l'inq's story about g80 not being pcisig compliant.

(sorry for old news, please ignore me now..)

but doing some random googling for it showed up this funny quote

Now, for a real comparison, the GeForce 7900 GTX is more along the lines of a Radeon X2100 GT Pro card (not the X1600 Pro), but just in looking at the examples above, you can see that the Radeon card that is probably 90% as good as the GeForce card costs less than 40% of the price.

http://www.wcreplays.com/forums/printthread.php?t=69017

[INKster]

Originally Posted by neliz

Uhm.. X2100 is Alive?

Saw M71-M when looking at the pcisig list http://www.pcisig.com/developers/com...ors_list/pcie/ to confirm l'inq's story about g80 not being pcisig compliant.

(sorry for old news, please ignore me now..)

but doing some random googling for it showed up this funny quote

http://www.wcreplays.com/forums/printthread.php?t=69017

X1950 XTX is not on the PCISIG listing.
Does that mean it's also not PCI-Express compliant ?

[neliz]

Originally Posted by INKster

X1950 XTX is not on the PCISIG listing.
Does that mean it's also not PCI-Express compliant ?

Clean your glasses and read again please, R580 is listed. (I don't think they register every board that some wiley taiwanese comes up with.)

[Razor1]

quadro 4400 or 4500 aren't on that list either or the 5500 or the 4000 or the 4500x2.

[neliz]

Originally Posted by Razor1

quadro 4400 or 4500 aren't on that list either or the 5500 or the 4000 or the 4500x2.

Well, like above, every chip is only listed once, so only g71 will show up and not every card bastardization that follows and that you come to love, be it a GT, GTX or GX2 or their workstation inbred brethren.

AMD's pci 1.1 list isn't completely tidy but does a good job at separating chip and card

[WaltC]

Originally Posted by Bouncing Zabaglione Bros.

We've certainly see them do things that were a dead-end (eg PP hints, shader replacement), and we've also seen them elect not to do things that we've seen from ATI, and then had Nvidia follow a few years later (improved AA, gamma corrected AA, angle independant AF, 256 bit bus, etc). There's no doubt that Nvidia and ATI have made different design decisions and decided to emphasise different technologies over the years. Sometimes one of them gets to what the market wants first, sometimes it's the other.

That's not necessarily a comment on the quality of their designs, but more the difficulty of predicting what developers and consumers will want a couple of years after they've started working on their next-gen technology.

I think that what 3d gpu consumers--and therefore hardware devs--want to pursue is Image Quality coupled with Performance. As noted, some R&D avenues get us there, and some are dead-ends. This is always mitigated and balanced by the fact that game developers always seek the widest possible markets for their products, which means that backwards compatibility is an absolute must for the game developer at any given point in time. When Epic and other developers talk about their upcoming "DX10 titles," I read those comments as indicative that these products will "fully support" some/many aspects of DX10 while at the same time those titles will also "fully support" some/many aspects of the earlier and far more common DX9-level hardware, too. I do not believe that for game developers the situation between DX10 and DX9 will be as "either-or" as believed in some quarters. To spend any amount of money developing a "DX10-hardware-required" title makes no economic sense at all at the present time. That's the reality, imo.

But of course the marketing situation is going to be much different, isn't it?... They're going to "either-or" us to death, much like they did when PCIe (vs. AGP) was first launched. Reality, though, will proceed apace at a much slower rate. IMO, of course...

[Razor1]

Originally Posted by neliz

Well, like above, every chip is only listed once, so only g71 will show up and not every card bastardization that follows and that you come to love, be it a GT, GTX or GX2 or their workstation inbred brethren.

AMD's pci 1.1 list isn't completely tidy but does a good job at separating chip and card

hmm no look at the geforce 6800 and 3400 quadro these were the same chips. or the nv42m or nv42glm, the gf7 class chips are different chips from professional to consumer thats why the soft mod doesn't work anymore with rivatuner.

But I don't think they are going by "same" chips only once anyways, ATi has all thier professional line and consumer lines listed.

[silent_guy]

Originally Posted by Demirug

But the real reason IMHO is that power consumption does not scale well with clock. The DIE would be smaller but much hotter.

Why the die would be hotter for the same final performance?

Unless you're going to extremes and start running into barriers (think CPU and even at 1.35GHz we're not there yet), performance/Watt increases with increasing clock speed.

I think the reason is because it doesn't fit the architecture (e.g. ROP tied to MC) and because it's harder to design (and the benefit is too small to make it worth doing it.)

[icecold1983]

Originally Posted by Demirug

Well, the G80 dual texture filter units could be a hard enemy in high AF settings.

well r580 already has an almost non existent performance drop when enabling af. also when does nda on r600 expire?

[INKster]

Originally Posted by neliz

Clean your glasses and read again please, R580 is listed. (I don't think they register every board that some wiley taiwanese comes up with.)

That's another graphics card.
R580 is not listed, only a "X1900 XT". Since the memory chips and speeds are different (probably certain aspects of the PCB too), so is the power consumption, therefore, i am still right. Or is this last one an aspect not tested by the SIG ? They test complete cards or just a GPU now ?
Like i said, there is no X1950 XT/XTX on that list. Prove me wrong.

What about RD600 or NF650/680i ?
They are also not on PCISIG.

[Kaotik]

Originally Posted by INKster

That's another graphics card.
R580 is not listed, only X1900 XT. Since the memory chips and speeds are different (probably certain aspects of the PCB too), so is the power consumption, therefore, i am still right. Or is this last one an aspect not tested by the SIG ?
Like i said, there is no X1950 XT/XTX on that list. Prove me wrong.

What about RD600 or NF650/680i ?
They are also not on PCISIG.

There was no R580 X1900, so one can assume it's still meaning the X1950's, which name might have changed on later date from planned X1900's?

The chipset lists seem a bit lacking all together, I for one couldn't spot i975 from there at all.

[Ailuros]

Originally Posted by geo

I find it impossible to take the 1GHz talk seriously when it is tied to the idea that they are actually aiming for 2GHz. If I had to ballpark it right now then my best guess is it's not going to be lower than 700MHz, but not so high as four digits. And I still haven't heard anything to make me take clock domains seriously.

Well IMHO 700 sounds too modest and 1000 too hot

[Ailuros]

Originally Posted by icecold1983

well r580 already has an almost non existent performance drop when enabling af. also when does nda on r600 expire?

Depending on the scenario there is no such thing as non existent even on G80. If you're testing scenarios like FEAR where AF almost doesn't make any difference then it's obvious that you can say that for all these GPUs AF is "for free".

Demi was obviously pointing at the very high bilerp rate of the G80, which in worst case scenarios shows less than half the drop with AF than with former GPUs.

[Ailuros]

Originally Posted by Razor1

It doesn't need to be a requirement,

I'm leaning towards scalers on the r600 because a rumor like scalers just doesn't pop up out of no where specially since that rumor was well before the g80 news, any news or rumors.

If for arguement sake that rumor that mintmaster heard was just to throw off the scent of the g80 and laid down on the r600, and this 1 ghz rumor is for real, wouldn't 64 or 96 vector ALU's be more then sufficient to give the r600 enough shader power to match its 512 bit bus? Doesn't matter if the vector units aren't fully utilized either. At 750 the r600 would have enough theoretical performance to match the g80 with 64 vec 4 ALU's, at 1 ghz it should be able to get a substatial lead.

Although this kind goes out of the way of "more" programming flexibility then scaler ALU's, thats why the scalar ALU's are making alot more sense once the marketing stuff is in there.

Fine and dandy so far. I on the other hand was aiming at units like TMUs/ROPs. Unless of course I'm completely wrong and pure shader throughput is the be all end all that defines all performance characteristics of D3D10 GPUs.

Depends of course as to what ALU:TMU ratio everyone really expects in the end. A 1:1 scenario is out of the question; more like something either 4:1 or 2:1 depending on how creative your math is. Then it comes down to what each unit is capable of...

[Dave Baumann]

Originally Posted by INKster

R580 is not listed, only a "X1900 XT".

If you look at the title of the list "R580" or "R580 Crossfire" is listed as the actual product.

[RobertR1]

Didn't know where to post this but....

When the R600 does come out we got the technical analysis of it on B3D, would it possible to dedicate a couple of pages and compare the R600 to Xenos, from both a hardware and software perspective, on their differences and similarities. That would be highly appreciated by many (I hope ).

[INKster]

Originally Posted by Dave Baumann

If you look at the title of the list "R580" or "R580 Crossfire" is listed as the actual product.

I don't want to keep pounding this anymore but..., "R580" as in X1900 GT, "R580" as in X1900 XT/XTX, or "R580" as in X1900 Crossfire (slower than the standard XT) ?
Could be any of those and still not be a X1950 XTX with GDDR4 (the current AMD top of the line graphics card, for all intents and purposes).

[Acert93]

Originally Posted by RobertR1

Didn't know where to post this but....

When the R600 does come out we got the technical analysis of it on B3D, would it possible to dedicate a couple of pages and compare the R600 to Xenos, from both a hardware and software perspective, on their differences and similarities. That would be highly appreciated by many (I hope ).

Seeing as consoles are often one of many "lowest common denominators" it would be interesting, but to sum it up "R600 rules Xenos" But yes, seeing how ATI's second generation GPU compares and contrasts, and why different decisions were made, and if there may be any baselines established due to their relationship, and so forth, would be interesting. Fine grained performance comments would also be interesting, but without benchmarking of Xenos it is kind of hard to compare some things. Would be an interesting article though.

[Unknown Soldier]

Originally Posted by INKster

I don't want to keep pounding this anymore but..., "R580" as in X1900 GT, "R580" as in X1900 XT/XTX, or "R580" as in X1900 Crossfire (slower than the standard XT) ?
Could be any of those and still not be a X1950 XTX with GDDR4 (the current AMD top of the line graphics card, for all intents and purposes).

From my understanding, 'R580' as in the entire range.

Beyond3D's R580 chip description - X1900
Beyond3D's R580 chip description - X1950


or if you like

http://www.beyond3d.com/misc/chipcomp/

Please note that the X1950 XTX for all intents and purposes is still the R580 .. just(as quoted) "with tweaks for improved GDDR4 support".

US

[Dave Baumann]

Originally Posted by INKster

I don't want to keep pounding this anymore but..., "R580" as in X1900 GT, "R580" as in X1900 XT/XTX, or "R580" as in X1900 Crossfire (slower than the standard XT) ?
Could be any of those and still not be a X1950 XTX with GDDR4 (the current AMD top of the line graphics card, for all intents and purposes).

It can encompass the lot.