Where’s the math?

[plutoniumman]

What’s the math that says a GPU is too fast for a certain port?

ie if someone says graphics card A is the fastest a pci-e 8x port can support... And that if you put the much faster graphics card B in, it wouldn’t perform any better than card A (in pci-e 8x), because the port is the limiting factor.

What’s the math behind it? All my life I’ve kinda been going on guess work, wether or not a port would have sufficient thru-put to support a GPU to its fullest potential.

Why does throughput matter so much, anyway? Once the GPU has the graphics loaded, isn’t it in the GPU’s memory? After this point, isn’t most of the data going through the port just the CPU updating the GPU on where to render the geometry?

 

[FuryRosewood]

whoever is telling you that you will be bottlenecked by a x8 bus is sadly mistaken, even synthetic benchmarks running sli at x8 vs x16 showed like a very very very marginal (did i mention very?) improvement in performance.

 

[Spesh]

whoever is telling you that you will be bottlenecked by a x8 bus is sadly mistaken, even synthetic benchmarks running sli at x8 vs x16 showed like a very very very marginal (did i mention very?) improvement in performance.

This is absolutely right. Current GPU's don't use the full bandwidth of x8, let alone x16

 

[kobaj]

You guys are missing the point. He is saying why don't we run HD6990s in AGP ports? Because the port would be a bottle neck.

Imagine the CPU as city A and the GPU as city B. City B needs food and water from city A, and in return city B gives investment and capital to City A. So if you have a single lane highway (AGP, or 8x, or whatever) connecting the cities; you can't get enough food and water to city B, and all of city B's citizens die. Which in turn, means they can't send back investment and capital to city A, so all of City A's citizens die.

Now that is a bit of an extreme example, and OP has a point here

Why does throughput matter so much, anyway? Once the GPU has the graphics loaded, isn’t it in the GPU’s memory?

Yes, a lot of the data is transferred to the GPU, but its not just a matter of then modifying vertices or tessellation data. First off, 1gb of memory that most GPU's currently have, is nothing. You have to remember said memory is split between 250 to 400 cores depending on the model of card.

Plus there is no compression (in a sense). A single texture that takes up 5kb in .jpg must be converted to raw data, and then mip-mapped 10-20 times depending on the card. Suddenly you've got 10 megs of ram gone for 512x512!

My point being, make the textures HDR, High resolution, include vertices, shadow, and every other bit of data and you'll see that there just isn't enough vram. You'll be swapping data back and forth from vram and regular ram more than actually drawing things to the screen.

And if that data swapping pipeline is only 100mb/s, the GPU is going to spend most of its time waiting, which is where you get the decrease in efficiency, fps, etc.

If you don't have to swap textures; like instead of playing video games you were say, bitcoin mineing or doing algorithmic computation, which only uses numbers and math; then you can safely run a 16x card in a 1x slot with no waiting or decrease in efficiency.

I hope this clears some stuff up .

 

[StrangleHold]

You guys are missing the point. He is saying why don't we run HD6990s in AGP ports? Because the port would be a bottle neck.

ie if someone says graphics card A is the fastest a pci-e 8x port can support... And that if you put the much faster graphics card B in, it wouldn’t perform any better than card A (in pci-e 8x), because the port is the limiting factor.

Where does he say a AGP slot? Looks like PCIe X8 slot to me.

He is basically talking about if a higherend card will be bottlenecked in a PCIe X8 slot. Or if you put a higherend card in a PCIe X8 slot, that it would be as slow as a mid range card in a X8 slot.

 

[kobaj]

Where does he say a AGP slot? Looks like PCIe X8 slot to me.

He is basically talking about if a higherend card will be bottlenecked in a PCIe X8 slot. Or if you put a higherend card in a PCIe X8 slot, that it would be as slow as a mid range card in a X8 slot.

I think his question is a bit ambiguous. I certainly understand your comment about how he's talking 8x, but as we've seen several times over, nothing has reached the capacity of 8x. I could have easily said "4x" or "1x" or "pci" instead of "AGP". My point was not to argue the 'x' of pci-e. It was to show why we don't run "super powerful card A, in slow slot Q".

 

[StrangleHold]

I think his question is a bit ambiguous. I certainly understand your comment about how he's talking 8x, but as we've seen several times over, nothing has reached the capacity of 8x. I could have easily said "4x" or "1x" or "pci" instead of "AGP". My point was not to argue the 'x' of pci-e. It was to show why we don't run "super powerful card A, in slow slot Q".

I see what your saying.