So, do we know anything about RV670 yet?

[Bouncing Zabaglione Bros.]

I can tell you that the projected performance of the RV670 is.... comparable to R600... At least that's what I've been told. And various sources have now also confirmed that R600 has reached it's EOL.... Yes, you read that right... End of lifetime for the R600. So there has got to be some replacement in the high end.

So the high end replacement for R600 will be still only competing with Nvidia's second string product? Or are AMD relying on two-R670-on-a-board-crossfire in order to compete at the top end?

 

[AnarchX]

With 825MHz it should be also perform in some cases better. I also hopping on a bit improved MC and little better ROPs (working color-compression? ), so it should be worthy for the "2950XT" name.

Above this should be the 2950X2X placed as top-of-the-line, acording the rumor-mill...

R680 is still questionable - Fudo said it does not exists/was canceled
But what is with the FireGL/FireStream-cards? Here should make 512Bit & 500+GFLOPs a lot more sense.

 

[Bouncing Zabaglione Bros.]

I would have thought that reproducing the same product that already fails to compete against it's competitors isn't exactly going to set the tills ringing. Okay, so it's on a smaller process, but if performance isn't any better, then R670 will still get stomped by Nvidia's current offerings, let alone their refresh.

Wonder how long it's going to be before R670 gets EOL'ed in favour of R700? Sounds like the sooner the better if it's only on par with the R600 and it's various compromises.

 

[INKster]

I would have thought that reproducing the same product that already fails to compete against it's competitors isn't exactly going to set the tills ringing. Okay, so it's on a smaller process, but if performance isn't any better, then R670 will still get stomped by Nvidia's current offerings, let alone their refresh.

Wonder how long it's going to be before R670 gets EOL'ed in favour of R700? Sounds like the sooner the better if it's only on par with the R600 and it's various compromises.

R670 ? Don't you mean RV670 ?
The "V" stands for "Value", not high-end.

 

[vertex_shader]

I would have thought that reproducing the same product that already fails to compete against it's competitors isn't exactly going to set the tills ringing. Okay, so it's on a smaller process, but if performance isn't any better, then R670 will still get stomped by Nvidia's current offerings, let alone their refresh.

You forget one thing, rv670 is a value performance GPU its not the successor of r600.

 

[Mariner]

So have we decided how feasible a 2 x RV670 Gemini-style card is as the new high-end option for AMD?

About as feasible as a 2 x G92 high-end card? Must admit I've not been following the speculation too closely.

 

[Pressure]

Considering that AMD already placed CrossFire Radeon HD 2900XT against the Geforce 8800 Ultra, I am not that surprised.

I'm a bit concerned if this is any indication that two RV670 boards (with four GPUs in total) is going to compete against one Geforce 8800 Ultra.

 

[Jawed]

Or are AMD relying on two-R670-on-a-board-crossfire in order to compete at the top end?

I thought that's taken as read from the point at which R600-refresh appears.

Jawed

 

[Jawed]

With 825MHz it should be also perform in some cases better. I also hopping on a bit improved MC

In what respects do you think it's deficient?

and little better ROPs (working color-compression? )

Colour compression isn't working currently in R600? 4xAA samples per loop is what ATI needs I reckon but I doubt any enhanced capability in the remaining R6xx GPUs simply because the focus will be on the D3D10.1-specific additions.

But what is with the FireGL/FireStream-cards? Here should make 512Bit & 500+GFLOPs a lot more sense.

I've honestly got no idea, but I wonder if apps that run on these GPUs care how many GPUs they're running on.

Jawed

 

[AnarchX]

In what respects do you think it's deficient?

The buckling in memory-scaling looks a bit suspect:
http://www.3dnews.ru/video/512_bit_memory_bus_efficiency/

4xAA samples per loop is what ATI needs

Is this not the case in current R6xx GPUs?

 

[Jawed]

The buckling in memory-scaling looks a bit suspect:
http://www.3dnews.ru/video/512_bit_memory_bus_efficiency/

Considering how tweakable the MC is under driver control and the complexity of latency v GDDR3/4 v bandwidth: I don't think it's possible to conclude there's a deficiency based on those results. Earlier ATI GPUs' MCs behave oddly when you run memory "out of spec", too.

Is this not the case in current R6xx GPUs?

Nope, 2xAA for the lose.

Jawed

 

[AnarchX]

Nope, 2xAA for the lose.

Source?

AA sample rate 47488 Msamples/sec

http://www.beyond3d.com/resources/board/331
-> This considering single-cycle 4xMSAA or not?

 

[Jawed]

That's wrong.

http://forum.beyond3d.com/showpost.php?p=1021537&postcount=155

Jawed

 

[AnarchX]

Than R6xx antialiasing-performance is really no suprise in combination with the shader-resolve.

But why than they went to 8x MSAA and does not let the three-cycle step - 6x MSAA, as a choice?

 

[Jawed]

Than R6xx antialiasing-performance is really no suprise in combination with the shader-resolve.

Shader resolve genuinely isn't a problem, unless you want 200fps+ in some moderately old game. Or like to draw graphs showing max frame rates.

But why than they went to 8x MSAA and does not let the three-cycle step - 6x MSAA, as a choice?

That I dunno. It's been mystifying me for months but never seemed important enough to dig... Could be just marketing...

D3D10.1, onwards, may make multiples of 4 mandatory (erm, not sure about that to be honest)?

Still, I fell off my chair multiple times when discovering R600's slow AA - I stupidly thought Xenos was the start of a new zixel/AA-sample regime.

As far as I can tell the best defence for R600's current ROP configuration is that it's relatively fast when doing FP16+MSAA. So you could opine the lack of games that use this capability, or call it forward looking. Still there's always 3DMk06's SM3+HDR game tests when AA is turned on:

http://www.xbitlabs.com/articles/video/display/diamond-viper-hd2900xt-1024mb_21.html#sect0

Jawed

 

[AnarchX]

Shader resolve genuinely isn't a problem, unless you want 200fps+ in some moderately old game. Or like to draw graphs showing max frame rates.

Or just a high shader-utilization.
In oZones "Fur" Rendering Benchmark(extrem multi-pass, MADD-using) I saw a drop-down of ~50% on R600 and ~20% on R580, when 4xMSAA was activated.

 

[Jawed]

Or just a high shader-utilization.

AA resolve is essentially a constant amount of time for given frame dimensions and AA level, regardless of the time spent generating the frame. So as frame complexity increases, the AA resolve costs less and less of the total frame rendering time. e.g. 1ms of AA resolve time when the frame rate is 30fps/33ms is 3% versus 1ms out of 60fps/17ms is 6%.

In oZones "Fur" Rendering Benchmark(extrem multi-pass, MADD-using) I saw a drop-down of ~50% on R600 and ~20% on R580, when 4xMSAA was activated.

Can't tell much from that without knowing the absolute numbers (and comparing them with theoreticals). You need to know if and how bottlenecks shifted... e.g. if R600 rendered 4x faster than R580 without AA...

Jawed

 

[ShaidarHaran]

Or just a high shader-utilization.
In oZones "Fur" Rendering Benchmark(extrem multi-pass, MADD-using) I saw a drop-down of ~50% on R600 and ~20% on R580, when 4xMSAA was activated.

RV560 w/4xMSAA @ 1280x1024 loses only 4% of performance compared to no AA @ default clocks (575/690). Seems to scale well w/GPU clock too, @ 635/690 and 4xMSAA 6% drop compared to no AA. Scaling the memory in addition to the GPU maintains the same performance drop as stock clocks produce of 4% with clocks of 635/790.

Given RV560's "half R580" nature (in terms of functional units and bus-width) what does this imply the bottleneck is for R580 AA performance in this benchmark?

 

[Robbitop]

4xAA samples per loop is what ATI needs

I doubt this would be a too serious bottleneck. It'd only cost a single clock for the RBEs to loop in order to place 4 subsamples per pixel into the framebuffer. What's a single clock compared to maybe a hundred ones for the calculation of a single pixel?

 

[Jawed]

I'm thinking of Z-only type stuff, shadow buffers being the obvious example (but also Z-only pass just ahead of colour rendering). The AA is more of a side-effect from being able to renders lots of Z fast. R600 has the bloody bandwidth...

The only reason I can think of is ATI's focus on FP16 performance - by prioritising that over 4xAA. It's a pretty big deal theoretically, but it hasn't paid dividends so far.

HDR pixels are always going to be rendered slowly, whereas there are plenty of reasons why Z-only pixels need to be rendered as fast as possible.

I wish there was some decent analysis of R600 AA performance, to get a clearer picture...

Jawed