X1800xl

[EasyRaider]

When I look at new games and the shaders they use, the ALU/TEX instruction ratio is usually greater than 4:1. Texturing power is less important on new games.

Does that account for anisotropic filtering?

Why is the R520 16-1-1-1 and not 8-1-3-2, if universally, it's more important to have higher ALU to pixel/texel rate for "new games?"

I've been wondering about that too. It just doesn't make any sense. It seems that RV530 or R520 or both are badly balanced. You could argue that the R520 is designed for fast AF, but then what about RV515 and R580?

I'm thinking 8-2-3-2 would have been most optimal for R520 and 4-2-3-2 for RV530.

 

[Jawed]

I really can't work out where the X1600XT is targetted. Is it a replacement for X700Pro or X800XL?

Jawed

 

[Joe DeFuria]

Basically, when the mid-range part only has 25% of the raw pixel/texel throughput of the high-end part...it just screams "gaping hole in lineup" to me. I generally expect the low end part to have 25% of the high-end...not the mid-range.

It does seem odd to me that ATI would apply the "more relative shader ALU power" to one part in the line-up, but not the other. (upcoming R580 notwithstanding. )

 

[EasyRaider]

I really can't work out where the X1600XT is targetted. Is it a replacement for X700Pro or X800XL?

In terms of price, definitely X800 XL. In terms of performance, well, it's somewhere in between.

 

[OpenGL guy]

Does that account for anisotropic filtering?

No. Texture latency is very hard to judge as the HW is very good at hiding it.

 

[EasyRaider]

Basically, when the mid-range part only has 25% of the raw pixel/texel throughput of the high-end part...it just screams "gaping hole in lineup" to me. I generally expect the low end part to have 25% of the high-end...not the mid-range.

It does seem odd to me that ATI would apply the "more relative shader ALU power" to one part in the line-up, but not the other. (upcoming R580 notwithstanding. )

In all fairness, the low end is 25% of the high end, while the midrange at least is very good for ALU ops. Otherwise, I completely agree.

(edit)
Also, it's fair to assume the midrange should offer much better value for money than the top of the line. Comparing X1600 XT with X1800 XT, the performance/price ratio seems about the same to me.

 

[RoOoBo]

Latency? The problem with high AF modes is bilinear throughput! Sure DOOM3 has shaders with 4 - 5 ops per texture access, and DX games may have even higher rates, but if each texture access requires 8 bilinear samples you don't have enough ops to run while your texture unit is busy. The GPU becomes completely texture unit limited (tomorrow I may post a graphic with one such test case).

 

[Crusher]

When there are 7800 GTs going for $360, the X1800 XL will have to have a very quick price drop to see many sales, I think.

 

[caboosemoose]

One thing that strikes me about RV530 is that it has approximately the same tranny count as R420. Imagine a 90nm R420 clocked to 600MHz as a midrange part...

I know that it's a little more complicated than that and you'd have all sorts of issue to address (like the problem of whther you got to 256-bit bus on a midrange chip to keep the bandwidth you'd need to feed such a chip etc etc etc), but I concur with the general feeling that RV530 is probably ahead of its time and will be a dead chip by the time the software environment matches its abilities perfectly.

 

[Xmas]

No. Texture latency is very hard to judge as the HW is very good at hiding it.

But you can't hide the sampling costs. If the ALU:TEX instruction ratio is 4:1, but at the same time the average cycle ratio is 1:4 (and with high quality AF, it most likely is even beyond that), what's the point of having 3 times more ALUs than TMUs?
Arithmetic performance does become more important, but today, I don't think a mid-range part with a 3:1 ALU:TEX ratio is a good idea. And from what I've seen so far, RV530 doesn't impress me at all, price/performance-wise.

 

[rwolf]

It seems like the new chips do great in DX and not opengl. Perhaps opengl guy could explain why?

 

[Jawed]

How many textures per shader are going to be AF'd?

Is it just going to be one or two surface textures?

My understanding has been that where textures are used for non-surface purposes (e.g. lighting lookups or depth of field effects) the shader is either performing un-filtered texturing or very simply filtered texturing in those cases.

Jawed

 

[Junkstyle]

Even if XL shipped with the same chip as the XT there is no way you will get XT performance because the memory is clocked at 1000MHz where the XT is clocked at 1500Mhz...good luck with that OC.

 

[jvd]

the xl is good . Its only as fast as the 7800gt but it does hdr / fsaa so why wouldn't i pick it up over the 7800gt

 

[Bob]

How many textures per shader are going to be AF'd?

They typically all end up with the same or very similar AF ratios.

 

[SugarCoat]

Perhaps this should be in the XL thread or one of the numerous X1000 threads, but here goes.

First of all, there is nothing to unlock, so your only hope is overclocking.

Second, ATi is using an older revision of the chip, one with a soft ground problem that did not provide sufficient yields for the XT, for early XL cards. This will most likely mean that the XLs are very poor overclockiers. Once this old stock is sold out (quantity unknown), the situation will likely improve.

Have to agree ATI made a mistake with stock clocks on it. With a card with its power driven mainly by its clocks its a shame ATI chose these and one has to wonder if it was a long time ago. Surely they wouldnt of done this if 200MHz more on the memory and maybe 50 on the core would of bumped it over a GT and just under a GTX (which it almost is in some cases).

The bum core thing should only of taken up a couple batches, and until i see some overclocking results or some XL's missing their heatsinks we wont know how long they last.

Someone correct me if im wrong but if the assumption of bad cores on the XL's is correct, then the core should be missing a metal layer in its ID correct? XT i believe was 5? I havent been able to find a review/preview site that actually removed the heatsink to verify.

The biggest reason though perhaps to wait for the XT may not be its clocks but a 512mb card for 550 is a really starting to look nice as we get closer to Vista and games get more advanced. So anyone looking for a long term purchase may want to wait for one of those.

We also have no idea what the first few months of driver revisions will bring for performance improvments, and there is definitly room there for improvment; i dont care what anyone says about them launching with mature drivers. Theres quite a few games where the X850s performance is just scraping under an XL and XT and that simply shouldnt be happening, and i doubt will be for long.

Yet another question, am i wrong to assume that the RV530s launch performance does not bode well at all for this "monstor" the R580 is assumed to be as well? Looks fantastic on paper but real world differences may sing a different tune.

 

[Lezmaka]

the xl is good . Its only as fast as the 7800gt but it does hdr / fsaa so why wouldn't i pick it up over the 7800gt

I guess that would depend on if you think hdr+aa is worth an extra $70-$100.

 

[ANova]

Even if XL shipped with the same chip as the XT there is no way you will get XT performance because the memory is clocked at 1000MHz where the XT is clocked at 1500Mhz...good luck with that OC.

You didn't read did you.

 

[Moloch]

I guess that would depend on if you think hdr+aa is worth an extra $70-$100.

Extra?
We'll see how much extra it is a few weeks to a month for prices to stabilise.
hdr and aa are def worth it for slower type games though for fps games it's not needed nearly as much as say an rpg game.

 

[RoOoBo]

How many textures per shader are going to be AF'd?

Is it just going to be one or two surface textures?

My understanding has been that where textures are used for non-surface purposes (e.g. lighting lookups or depth of field effects) the shader is either performing un-filtered texturing or very simply filtered texturing in those cases.

Jawed

In DOOM3 it looks like more textures than I expected are actually set with high AF (or the OpenGL framework is writing the wrong MAX_ANISO value which I doubt). If the IHVs then think that it's a mistake and transform most of those to no AF it's a different matter that should be discussed between the engine designers and the driver designers.

What do you think is the bottleneck in this DOOM3 frame? Texture unit utilization is measured in the address ALU (but for the simulator it's the same that filter ALU utilization). Rendered at 1024x768 and 8xAF. Real GPUs execute more shader instructions per cycle than the simulator. And you can't get better performance from the TU unless you implement 2 bilinears/trilinear per cycle or better.

Performance degradation from 12:12 to 12:4 is around 30%. The 12:8 configuration didn't look that bad (5-10%) and 12:4 could become the sweet spot when larger shaders (2x at least more arithmetic ops) are used and if the number of texture fetches doesn't change much. Or if AF isn't used ...