Anand has the details about r520,rv530,rv515

[Hellbinder]

Batches.. We don need no stinkin Batches...

 

[Jawed]

Am I missing something? We seem to be both agreeing and arguing different things.

You're missing the fully decoupled texture pipes, which I believe are common to all of R5xx, as well as Xenos.

Batches need to be scheduled to shader or texture pipes. The count of pipes is immaterial. Their being decoupled from each other is the key concept. Some kind of advanced scheduler is required.

Jawed

 

[CMAN]

Can someone explain the X-X-X-X?
I'm guessing the first X is the number of pipelines and the third X is the number of TMU's per pipe. But I'm not really sure.

 

[kemosabe]

Sorry, can't participate in the ROP/texturing/batch debate, but Baumann clearly insinuated that we should look to 4-1-3-2 to comprehend why RV530 is described as having 12 "shader processors". So if he's not being disingenious about it, obviously R580 (16-1-3-1) must have 48 of these shader processors, regardless of what the other numbers mean. By December, R520 variants will quickly be relegated to the performance segment and a $600 Radeon X1900 will rise from the ashes.

 

[_xxx_]

By December, R520 variants will quickly be relegated to the performance segment and a $600 Radeon X1900 will rise from the ashes.

I predict that too. It's actually the only way for ATI to re-gain any high-end market share. If they can pull it off, that is...

 

[Geo]

So if he's not being disingenious about it, obviously R580 (16-1-3-1) must have 48 of these shader processors, regardless of what the other numbers mean.

The alternate theory I've seen hereabouts, in an attempt to finesse transistor bloat (I think), is that the 3 might be "ops" rather than "units". Whether this is a distinction with an actual difference remains to be seen.

 

[Razor1]

http://img.pconline.com.cn/images/bbs4/2005/09/29/1127924803613524.jpg

more leaked benchmarks

 

[ERK]

Those "relative performance" numbers are very interesting:

RV515 - 1
RV530 - 3
R520 - 5

Jawed

Sorry for jumping in late here...
It seems the 5 is just a 4, but adjusted for clock speed differences.
The 4, to perhaps state the obvious, coming from the notion that R520 is basically 4 RV515s architecturally (16-1-1-1 = 4x 4-1-1-1) in other words, four quads vs one quad.

 

[IbaneZ]

http://img.pconline.com.cn/images/bbs4/2005/09/29/1127924803613524.jpg

more leaked benchmarks

There's Sassens numbers again.

 

[Arty]

There's Sassens numbers again.

Yup, rehash of a rehash of the rehashed numbers.

 

[DegustatoR]

TMU - ROP/TMU - PS/TMU - Z/TMU

or

ROP - TMU/ROP - PS/ROP - Z/ROP

 

[ERK]

The alternate theory I've seen hereabouts, in an attempt to finesse transistor bloat (I think), is that the 3 might be "ops" rather than "units". Whether this is a distinction with an actual difference remains to be seen.

I think it must be 'units' or 'processors.' I don't buy significant heavy/light shader differences, either. Just keep in mind that the shader power of the X850 is much less than the 7800GTX, and would be even more soundly beaten if not for its higher clock speed. The GTX basically has three times the units!

It is looking like the X1800 will also have a shader (per clock) deficiency compared to the 7800, also likely having only 16 shader processors to compete with the 7800's 48 shader ALUs (but again at a higher comparative clock speed).

I think the leaked benchmarks support this view, but we'll see shortly when the reviews come out.

In that light, I am thinking the R580 will indeed have thrice the shader power (processors) of the R520, and only then will surpass the 7800 in raw shader throughput.

Speculation throughout, of course,
ERK

 

[Rys]

It is looking like the X1800 will also have a shader (per clock) deficiency compared to the 7800, also likely having only 16 shader processors to compete with the 7800's 48 shader ALUs (but again at a higher comparative clock speed).
ERK

If you're counting 'shader ALUs' and stating that number for G70, you should count for R520 too in the same way, giving 32 vec+scalar ALU pairs, 2 per fragment processor.

 

[PeterAce]

If you're counting 'shader ALUs' and stating that number for G70, you should count for R520 too in the same way, giving 32 vec+scalar ALU pairs, 2 per fragment processor.

But are R520 ALUs ?:

Vec3 + Scalar (4D like R300, R420)
or
Vec4 + Scalar (5D like Xenos)

And if there are 2 per fragment processor (vertical) is there a crossbar between them? Do they have the same instruction distrubution?

And where does this leave R580's ALUs?

Are the mini-ALUs (pre-ALUs) Vec+Scalar pairs still there as well?

 

[Joe DeFuria]

Don't mind me, but...

AAARRRRRRGGGGHHHHH!!!

Ok...back to the thread...

 

[Moloch]

But are R520 ALUs ?:

Vec3 + Scalar (4D like R300, R420)
or
Vec4 + Scalar (5D like Xenos)

And if there are 2 per fragment processor (vertical) is there a crossbar between them? Do they have the same instruction distrubution?

And where does this leave R580's ALUs?

Are the mini-ALUs (pre-ALUs) Vec+Scalar pairs still there as well?

You're making my head spin

 

[Razor1]

There's Sassens numbers again.

hehe sorry should have checked before I posted good call.

 

[KimB]

I think it must be 'units' or 'processors.' I don't buy significant heavy/light shader differences, either. Just keep in mind that the shader power of the X850 is much less than the 7800GTX, and would be even more soundly beaten if not for its higher clock speed. The GTX basically has three times the units!

Not really. The X850's pipes can be described as:
Tex + ALU + mini ALU

...whereas the GTX can be described as:
Tex/ALU + mini ALU + ALU

So if you're doing just MAD's, yes, the GTX will act like it has twice the number of ALU's per pipeline. But not for all shaders will it be that much better.

 

[Jawed]

TMU - ROP/TMU - PS/TMU - Z/TMU

or

ROP - TMU/ROP - PS/ROP - Z/ROP

I dare say the concensus round here is that all R5xx will have double-rate Z. Sure, it's a guess - but the implication that only RV530 would have double-rate Z doesn't really make sense.

Jawed

 

[Razor1]

Not really. The X850's pipes can be described as:
Tex + ALU + mini ALU

...whereas the GTX can be described as:
Tex/ALU + mini ALU + ALU

So if you're doing just MAD's, yes, the GTX will act like it has twice the number of ALU's per pipeline. But not for all shaders will it be that much better.

can't both main ALU's do either mad or mul operations if they are free though? So it would be more like 2.5 x the shading power in most instances?