A Few 9700 Screenshots

[DemoCoder]

You still need to issue two instructions, you need twice as many ports on your register file, and even though instruction to instruction data-dependency is low you still need a scoreboard to resolve them.

Cheers
Gubbi

Why? Just assume that the compiler/driver has resolved them. This can be done completely using static analysis due to the nature of the shader's restricted programming language. NVidia, for example, will know the timings of their units, and the compiler, in the case of an unresolvable dependency, can insert delay instructions as needed.

 

[Pete]

Well, I didn’t really see any banding but then I thinks its going to be difficult to see any issues with straight textured titles.

Dave, people have said Soldier of Fortune II shows quite a bit of banding. I'd give that a whirl, if you can.

As for the rest of you: doesn't this constant bickering get boring at some point? Just enjoy what ATi and nV are giving you. There's no need to waste B3D's bandwidth rehashing the same argument over and over and over and over again.

 

[Ichneumon]

SOF2 shows quite a bit of banding if you crank up the gamma on the 9700. However, it does that on any other card too that i've seen it on.

I'll try it without compressed textures and see if that changes it... i just did a quick run of it last night as i'm still exploring the card.

 

[Gubbi]

You still need to issue two instructions, you need twice as many ports on your register file, and even though instruction to instruction data-dependency is low you still need a scoreboard to resolve them.

Cheers
Gubbi

Why? Just assume that the compiler/driver has resolved them. This can be done completely using static analysis due to the nature of the shader's restricted programming language. NVidia, for example, will know the timings of their units, and the compiler, in the case of an unresolvable dependency, can insert delay instructions as needed.

What you describe is a VLIW, not a superscalar shader, which is what I thought you meant originally. - And as I stated VLIW makes more and more sense as the low level instruction scheme for a shader, the more I think about it.

However, there is a (low) upper limit to how complex you can make a shader unit before you get diminishing returns. If you can pack 3-4 operations into one shader VLIW instruction, chances are you'll have alot of idle slots in there, inflating shader program size and wasting silicon real estate (idle execution units).

I still believe you'd be better off with twice (or 50% more) single issue shader units than more complex ones. Of course with gobs of pixel pipes, you eventually run into problems with waste when rendering lots of small polys, so you'd have to decouple the pixel shaders so that the pixel shaders can operate on different polys in the same cycle.

Cheers
Gubbi

 

[DemoCoder]

What you describe is a VLIW, not a superscalar shader, which is what I thought you meant originally. - And as I stated VLIW makes more and more sense as the low level instruction scheme for a shader, the more I think about it.

No, I did not explicitly mean VLIW (although I agree with you that it is probably the most efficient implementation) . I mean that the compiler schedules and routes operations to functional units in the same way that compilers do register allocation today.

Let's say you have K functional units from a set F. Consider the following program:

X = T * C
Y = N * L
Z = R * V
MC = Ka + X
DC = Kd * Y
SC = Ks * Z
FC = MC + DC + SC

X, Y, and Z are data-independent, so I explicitly schedule (via some extra bits) these operations to functional units F_1, F_2, F_3. DC and SC are also data independent. Same there.

(F_1) dp4 X, T, C
(F_2) dp4 Y, N, L
(F_3) dp4 Z, R, V
(F_1) add MC, Ka, X
(F_2) mul DC, Kd, Y
(F_3) mul SC, Ks, Z
(F_1) add FC, MC, DC
(F_2) nop
(F_3) nop
(F_1) add FC, FC, SC

Now, if the GPU is VLIW, than you'd use tree-covering in the compiler back-end generation phase to pack this schedule into the proper VLIW words. On the other hand, you could just as easily have a GPU that handles these instructions via pipelining and prefetch as normal, except that it doesn't have to use a scoreboard to figure out which units are busy and which are ready. The GPU can just blindly execute these instructions without worrying about a conflict/hazard.

Yes, you are always going to have situations where some units go idle, but you have the same situation of idle silicon by adding additional pixel pipelines. I would argue that increasing parallelism at more granular levels is better than increasing it at more course levels.

If you increase it at a very granular level (like the P10), then you have opportunities for allocating resources where they are best needed. If you increase it coarsely by including more monolithic pixel pipelines, when one of them goes idle, you are wasting *far more silicon* You would argue of course for allowing these pipelines to be more independent. I would argue that you should take this further and allow resources within a pipeline to be more independent. I think the Trident XP4 takes this approach.

I would advocate a more general approach. Include lots of reusable units and spend logic on complicated routing architecture so that the compiler can hook up the units together as they are needed and route data between them (ala MAJC)

Adding more pipelines might be the brute force way of doing it, but you could achieve the same thing by going multichip and just using extra GPUs to boost performance. It just doesn't seem as elegant to me.

 

[KimB]

I would argue that you should take this further and allow resources within a pipeline to be more independent. I think the Trident XP4 takes this approach.

Not entirely sure what you mean by allowing the resources within the pipeline be more independent, but it stands to reason, to me, that the XP4's pipelines are more dependent upon each other than the pipelines of other chips today.

 

[DDM_Reaper20]

Did any of the reviews mentioned anything? The only thing I remember is "we didn't connect the externalt power, but the system booted up nevertheless". Not much of a problem, is it?

Got this from HotHardware, first page of their review:

Since the R300 VPU draws so much power, it may actually stress the capability of some systems to deliver enough power through the AGP connector. This additional power source to the board ensure cross platform compatibility. We did try powering up with the connector unplugged and the system indeed would not post.

And I really don't feel like responding to the rest of your post, SvP. I don't like to respond to personal attacks.

That's why there's the power connector, to ensure that systems will power up (of course, anybody trying to connect it to a dirt-cheap POS 200W PS will be in for a big surprise). Not connecting it sounds like asking for trouble to me. There'd be legitimate grounds for complaint if there was no power connector.

As to refreshes and all, what was the exact reaction when nVidia announced the GeForce 3 Ti500 and kin . . . ? Which is superior in which departement to the GeForce 3, aside from speed . . . ?

Let's face it, it's quite a PR boost for ATI to have come out with Radeon 9700 Pro before Nvidia has anything to show. It's not the end of the world, of course, but it's pretty impressive IMO.

Edit: grammar

Best,
DDM_Reaper20

 

[alexsok]

Let's face it, it's PR boost for ATI to have come out with Radeon 9700 Pro before Nvidia has anything to show. It's not the end of the world, of course, but it's pretty impressive IMO.

Obviously it is, ATI wanted to be first and all I can do is congratulate them on a job well done!

 

[Dio]

That's why there's the power connector, to ensure that systems will power up. Not connecting it sounds like asking for trouble to me. There'd be legitimate grounds for complaint if there was no power connector.

The R9700 BIOS will shut the card down and prevent booting if it detects the power connector is not attached. The design of the board power circuitry is such that there is the possibility of board (not chip) damage if the chip puts a large current draw when the external power connector is not on.

Some early boards didn't have this feature in the BIOS - although I think all the boards that have gone outside ATI should have, it is not impossible a couple have slipped through. I have booted mine without the power connector on as well, and it survived, but there was nearly a serious muscle control failure when I realised how close I could be to frying a board that good

 

[KimB]

That's why there's the power connector, to ensure that systems will power up (of course, anybody trying to connect it to a dirt-cheap POS 200W PS will be in for a big surprise). Not connecting it sounds like asking for trouble to me. There'd be legitimate grounds for complaint if there was no power connector.

I still think that there is legimitate grounds for complaint. Mainly, since it requires so much power, it's going to be producing lots of heat. In this situation, the two are one and the same (power consumption and heat dissipation). Granted, this shouldn't make much difference for the enthusiast, as that person should know the implications of the card already. But, this card should not be marketted to somebody who is new with computers and, "Just wants the best," say, to upgrade their manufactured system (Dell, Gateway, Compaq, etc...).

 

[Dave Baumann]

But, this card should not be marketted to somebody who is new with computers and, "Just wants the best," say, to upgrade their manufactured system (Dell, Gateway, Compaq, etc...).

Good grief!!

 

[Bigus Dickus]

But, this card should not be marketted to somebody who is new with computers and, "Just wants the best," say, to upgrade their manufactured system (Dell, Gateway, Compaq, etc...).

You're not serious are you? Like previous generations of video cards haven't been extremely hot or power hungry.

GeForce256, GF2 Ultra, V5 5500...

IMO, adding the power connector actually ensures stability (given an adequate PSU) for the "average Joe" that might not have the best motherboard around.

 

[Galilee]

You have to use a powercord on a CD-burner also, but anyone should and can buy one. It's no big deal.

 

[KimB]

[quote="Bigus You're not serious are you? Like previous generations of video cards haven't been extremely hot or power hungry.

GeForce256, GF2 Ultra, V5 5500...[/quote]

Which actually supports my argument. If those were hot and power hungry, then how much hotter/more power hungry is a board that needs external power?

 

[Doomtrooper]

As GPU's advance in technology and are being developed with 110 +million transistors heat is something that is unavoidable.

Look at the .13 Micron Athlon XP and Pentium 4..both processors are pumping out 60-80 watts of heat....Graphic processors are going to see the same heat issue.

The 9700 draws approximatley 50 watts of power...with todays high powered machines this isn't a big deal at all.

 

[Dio]

Which actually supports my argument. If those were hot and power hungry, then how much hotter/more power hungry is a board that needs external power?

There's no difference; unless you assume systems are being sold with less headroom on the power supply now than they used to be. Changing a video card then or now for a more power-hungry one always would take a hit to the power supply.

For example, the motor on the average harddrive during spin-up consumes an awful lot of power. And the videocard does not draw max current except when running 3D applications, at which point harddrives are not spinning up. So there is 'headroom'.

I've got two systems here that are power-marginal and sometimes fail to boot. And the results from changing video cards (even between say Savage4 at one extreme and R9700 at the other) makes _no_ difference in boot possibility. Detaching a case fan helps more.

As a percentage of the overall MB power, the graphics card consumes less now than it did when AGP was new.

 

[Entropy]

Look at the .13 Micron Athlon XP and Pentium 4..both processors are pumping out 60-80 watts of heat....Graphic processors are going to see the same heat issue.

The 9700 draws approximatley 50 watts of power...with todays high powered machines this isn't a big deal at all.

A major difference is that you can fairly easily use a heatsink with a large dissipation area on the CPU. On the GPU however, you are very limited in terms of height, or you will block PCI slots. The solution is to increase airflow instead. Alas, that will increase noise correspondingly.

I successfully converted my R8500 to passive cooling via a Zalman copper heatsink. It works well, though it cost me a PCI slot. I doubt that will be a sufficient cooling solution for the R9700 however.

Power draw/cooling noise (and lack of support for 3D-glasses) is my primary concern about the R9700. YMMV.

Entropy

 

[KimB]

I've got two systems here that are power-marginal and sometimes fail to boot. And the results from changing video cards (even between say Savage4 at one extreme and R9700 at the other) makes _no_ difference in boot possibility. Detaching a case fan helps more.

As a percentage of the overall MB power, the graphics card consumes less now than it did when AGP was new.

Just as you said, the graphics card doesn't use max power at bootup (it should be using very little). Instead, a power-hungry (and therefore hot) graphics card would instead cause system instability while playing 3D games. Anyway, I'd like to see what happens when a number of people have 9700's...

 

[Doomtrooper]

Entropy there is low profile solutions out there that doesn't required the loss of a PCI slot...the Vantec Iceberg and this one here works very well..

http://www.pcpowercooling.com/products/cooling/card/video_cool/video_cool.htm

 

[Entropy]

Entropy there is low profile solutions out there that doesn't required the loss of a PCI slot

Thanks for the suggestion. Actually, my plan for the future is to use a high quality low-profile 1U CPU cooler, and lower the voltage to the fan. Alternatively, Zalman has a heat-pipe design that isn't offered on the market yet that allows the use of heat dissipation blocks on either side of the gfx-card. Combined with a quiet fan on a fan-bracket, that might do the trick. Going with water-cooling is a third option, but I'm wary of pump and fan noise adding up.

Entropy