So, do we know anything about RV670 yet?

My bad. Got all excited when I saw 128-bit referenced here and 64-bit referenced in an R600 presentation. I wish they wouldn't do aggregate precision like that... Should stick to per component.

 

Some people claim that in order to support DX10.1, all NV has to do is enable support in drivers. They also claim whats known as 10.1 was intended to be used in 10.0, but one of vendors failed to implement whats needed so these things were cut from final spec, thus - somebody needed new hardware for "DX10 done right"
No idea if this is true or not.

 

Only thing similar to this that I've heard was dropping some memory virtualization to "optional" due nV not getting it right

 

RV680?

http://www.digitimes.com/bits_chips/a20071023PB200.html

2x RV680 = R680?
RV680 = MCM-Version of RV670, with on-die PCIe-bridge and A12-stepping?

 

nah, RV680 will not be MCM. I reckon it'll be more like the gemini cards we've seen so far. RV670 runs too hot to be placed in MCM.

 

Wow, Google translation got quite a bit better as of late!

 

Maybe a mistake, RV670XT GPU is 85W...

 

If it is Gemini, it is possible also by RV670.
As for RV680, optimization for multi GPU might be done.
For instance, HT is used for an internal connection.

 

I think so, too - though is HT fast enough? I was going to start a thread on the topic of intra MCM bus types a while back, but got lazy.

Jawed

 

Offtopic: Actually I checked back daily, but was still getting the message that the boards were not available. Up until a few minutes ago I still got that message. :S

Yes. Gather4 is the DX10.1 name for Fetch4. AMD has supported this since the R580, RV515 and RV530 in their hardware. nVidia needs to make a bigger change to their hardware to support Fetch4 in DX10.1. They'll have to make a change in their TMU's to support this...

 

I thought Fetch4 was ATi's answer to something Nvidia already had? Something to do with PCF acceleration?

 

For the record AMD also has PCF (added w/R600 I believe) which they were lacking in the R5xx days.

 

No, Fetch4 does a "scattered" fetch which then allows shader code to perform PCF. A normal fetch (point-sampled, say) reads one or more colour channels from a single texel. Fetch4 is a "scattered" fetch because data from 4 texels is read into one "pixel" in shader code, in a single fetch instruction.

http://www.beyond3d.com/content/reviews/2/4

PCF, per se, is a feature of the texture pipeline in D3D10 (i.e. it's no longer NVidia proprietary, protected by patent). I'm guessing that D3D10.1 generalises this fetch, allowing the ALU pipeline to get at these texels as though PCF had left them untouched.

Jawed

 

Thanks....so what manner of fetch did Nvidia use to drive their PCF implementation and why would it be necessary to change it for DX10 as suggested above?

 

NVidia's fetch for PCF didn't provide the option to deliver those texels (as channels of a single .rgba texel) into the ALU pipeline. Fetch simply delivered the (single-channel) shadow texels as a quad for PCF filtering within the TMU, doing something funky (erm, someone else will have to explain it) with the bilinear filtering units of the TMUs to produce the PCF result.

So Fetch4 and presumably Gather4 just provide "texel passthrough", from the texturing pipeline to the ALU pipeline.

Jawed

 

Some frutzilla info (i don't think so its a surprise to anyone).
Geforce 8800GT is faster than RV670XT

 

Hm... Ok, it's 55 nm. Sorry.

Does it matter? Which CPU do I need to see this?

 

I'm not sure if this was already posted:



edit: Fixed link, thanks Kaotik.

 

Your link has extra "http//" in it
But that's indeed one of the best improvements in crossfire I believe, the ability to use more monitors with CF than before (you were limited to 1 monitor before, afaik?)