We are thinking about using Tesla GPUs for feedback control.
However, it seems that there is no established way for fast DA/AD
conversion of input and output data.
Has anyone ever done this before? What’s the best way to (quickly)
get analog data in and out of the GPU?
The solution that we are thinking about at the moment is to combine
a GFX unit with separate digitzers and analog output cards on a
CPCI bus that is connected to the GPU PCI via a bus extender.
However, the question is if the GPU offers a slave memory area on the
host computer bus (PCI/PCIe),so that the digitzer could write data
directly into the GFX (and vice versa for the analog output).
Does anyone know if that is possible?
Otherwise we probably have to go via the host PCs memory, i.e. the digitzer
pushes the data into the host PCs memory and the GPU actively pulls it.
In this case I would be interested if anyone can give me some numbers
on how fast the data could be transferred. The specifications talk a lot
about bandwith, but I’m more interested in latency. How much time would
it take to transfer e.g. 80 samples of 16 bits each from the host pc
into the GFX?
We are thinking about using Tesla GPUs for feedback control.
However, it seems that there is no established way for fast DA/AD
conversion of input and output data.
Has anyone ever done this before? What’s the best way to (quickly)
get analog data in and out of the GPU?
The solution that we are thinking about at the moment is to combine
a GFX unit with separate digitzers and analog output cards on a
CPCI bus that is connected to the GPU PCI via a bus extender.
However, the question is if the GPU offers a slave memory area on the
host computer bus (PCI/PCIe),so that the digitzer could write data
directly into the GFX (and vice versa for the analog output).
Does anyone know if that is possible?
Otherwise we probably have to go via the host PCs memory, i.e. the digitzer
pushes the data into the host PCs memory and the GPU actively pulls it.
In this case I would be interested if anyone can give me some numbers
on how fast the data could be transferred. The specifications talk a lot
about bandwith, but I’m more interested in latency. How much time would
it take to transfer e.g. 80 samples of 16 bits each from the host pc
into the GFX?
I know there is a technology called GPU-Direct which allows simulataneous DMA access to the same host memory area by a GPU and Infiniband network cards. Probably something similar could be used for your purposes. But I really dont know how hard it is to put together drivers/hardware changes to make that possible.
I know there is a technology called GPU-Direct which allows simulataneous DMA access to the same host memory area by a GPU and Infiniband network cards. Probably something similar could be used for your purposes. But I really dont know how hard it is to put together drivers/hardware changes to make that possible.