turbocache, shared and dedicated memory

Shared memory and Turbocache are just buzzwords. These technologies are not new. They're quite indeed quite old.

As you stated dedicated memory is physical memory owning only by the GPU. Only the GPU uses it and you only have access to it through the GPU (this is not entirely correct, but otherwise you would have to go into great detail).

I think everyone still knows AGP. The AG port already allowed shared memory when it was first introduced. Thats the AGP aperture size. It's first used when the GPU runs out of his own physical memory. It then transfers data to the main memory, making space for new stuff. Graphics APIs like OpenGL also allow it to upload data to the AGP, so you have a choice which memory you want to use.

Jam it back in, in the dark.

There is no card with solely dedicated memory. GPUs can always access main memory to swap data out (*). What you want is a card with a sane amount of physical (dedicated) memory.

(*)It can of course be disabled, like AGP could be completly disabled in the BIOS:

There's nowhere I can't reach.

Highly possible. Disallowing the GPU the access to main memory would hurt performance so I doubt they disable any of these features.

Always keep in mind that this shared memory / turbocache thing is only there to convince you to buy the product. Shared memory was always there but the vendors have realized that it makes a difference when they explicetly state that the product can use shared memory. Makes the customer believe he gets more for his money, although that's simply not the case.

Most amazing jew boots

Try to get as much physical (dedicated) memory as possible (as much as your budget allows). The shared memory amount isn't really important.

Short explanation:
You get good performance with your GPU when everything (geometry data and textures) fits into the dedicated memory on the card. As soon as there is no more space left on the card the driver/application is going to use the main memory. Now I should note that the card can't use ALL of the main memory, but only a fraction of it. But that doesn't really matter, because:
- As soon as data is 'swapped' out of the dedicated graphics memory the performance is going to collapse. I'm assuming the application that renders needs every bit of data it stores in the memory (both main and dedicated) to construct a frame. Which means it has to transfer data back and forth from and to the main memory to render the scene. Main memory is quite slow compared to the dedicated GPU memory. Now you're limited by your main memory speed and by the bus speed the data is transferred through. That results in a huge performance degradation.
- As soon as you notice that the FPS collapses you're going to reduce game settings, like geometry detail, texture resolution, decal amount, etc. - because you don't want crappy performance that makes it impossible to play.
- After adjusting settings everything fits into GPU memory again, so you can play with good performance

The whole thing is of course simplified. There a cases where using the main memory instead of the GPU memory is wise, but that's another story,

I am a dolphin, do you want me on your body?