Here the idea being conveyed is that chip designers always tend to over-provision and be conservative when developing fixed-function components.
Here for ex:If they had under-provisioned the rasterizer by giving it 1% rather than the 1.2% required by it, then assuming it ran at 100% when it was correctly provisioned it will be running at around 83.3%(100/1.2) on under-provisioning.
Now if all the other components depended on these then all of them would be slowed down proportionally and hence one single under-provisioned component can slow down the entire chip. So the usual practice is to over-provision which might diminish a few advantages of the fixed functionality device.
This comment was marked helpful 0 times.
eatnow
Is it possible for the software to detect when hardware rasterization is not enough (and is the bottleneck), and so have computation "spill over" to the general purpose cores?
This comment was marked helpful 0 times.
bwasti
I wouldn't be surprised if that was done in hardware, if at all. As far as I know, GPUs offer little control to the programmer over which chips are used.
Here the idea being conveyed is that chip designers always tend to over-provision and be conservative when developing fixed-function components.
Here for ex:If they had under-provisioned the rasterizer by giving it 1% rather than the 1.2% required by it, then assuming it ran at 100% when it was correctly provisioned it will be running at around 83.3%(100/1.2) on under-provisioning.
Now if all the other components depended on these then all of them would be slowed down proportionally and hence one single under-provisioned component can slow down the entire chip. So the usual practice is to over-provision which might diminish a few advantages of the fixed functionality device.
This comment was marked helpful 0 times.
Is it possible for the software to detect when hardware rasterization is not enough (and is the bottleneck), and so have computation "spill over" to the general purpose cores?
This comment was marked helpful 0 times.
I wouldn't be surprised if that was done in hardware, if at all. As far as I know, GPUs offer little control to the programmer over which chips are used.
This comment was marked helpful 0 times.