Just because their arch is designed for high ILP it doesn't mean they cannot transform their dot products into multi-cycle MUL, MAD..MAD sequences, potentially freeing up unused lanes. On the other hand they might have enough free lanes anyway in the average case that perhaps it's not worth the hassle to implement such an 'optimization'.
If you are overclocking why do you care about design standard limits? So you drop another .1 volt over the wires and get them a little warmer ... care? It's not like there are circuit breakers in them.
Thanks, yeah, I've been referring to "naked vertices" as these attribute-less points produced by TS, but not being clear on the fact that they're in patch space that can't be used as input to Setup, which is why DS converts back to the application's space and then clothes them in attributes.
If the bandwidth of the 256bit model isn't fully utilized, then yes.
Yes, that's what I had in mind.
That increases the latency of the DP, which will generally increase wasted lanes as that generally increases the latency of the entire shader.
Yes, you could argue that there's typically 1-1.5 lanes free per clock in pixel shading. I honestly haven't spent much time on vertex shaders to have a good idea what happens there.
Really? HD4870 is normally considered ~30% faster than HD4850 with 4xMSAA. While HD4770 occasionally exceeds HD4850, it's normally a bit slower.
Equal - yes, it can. But you can buy 4870 for as low as $130 now which gives Juniper XT a hefty $70 premium for DX11 support. That's a bit too much of a premuim from my point of view.
It's fixed per draw call but the actual count of attributes per vertex can be as high as 16 (vec4) in D3D10, I believe. 10.1 and 11 support 32 if I remember right.
I'm pretty sure he's just making up numbers. You can't just double the TDP of the single gpu card (for reference: 4870 160w (seems too low - 4890 is set to 190 and using about the same), 4870x2 286w). They can use better chips, the board power will probably be lower etc.
Even more funny is this: http://www.fudzilla.com/content/view/15508/1/
Linking the fake graphs based on fudo's own numbers - but at least he can claim to be right that way
Actually the claim shifted between then and now. The orginal claim was RV770, now its moved to RV790...Even more funny is this: http://www.fudzilla.com/content/view/15508/1/
Linking the fake graphs based on fudo's own numbers - but at least he can claim to be right that way
There is nothing to do for it, if you OC the card and it pulls more current through the wires than what 6-pin power cables are rated that is exactly what happens. The system doesn't just magically stop working if you cross the line in the sand.
I compared 512MB models to avoid influence of different VRAM capacity (1680*1050, AA4x/AF16x, computerbase performancerating). HD4870 1GB is significantly faster than HD4770 512MB (up-to 40%) but that's due its VRAM capacity, not its bandwidth... HD4770 1GB results would be very interesting, I think it could outperform HD4850 512MB.
