By: Marcin Niewiadomski (marcin.niewiadomski.delete@this.gmail.com), September 16, 2007 8:48 pm
Room: Moderated Discussions
David Kanter (dkanter@realworldtech.com) on 9/14/07 wrote:
---------------------------
>I'd put the memory controller with the GPU because it make sense to have the GPU
>use the system memory as the frame buffer.
>
Agree - it should be simpler when GPU manages all of the physical memory. Still CPU could be memory controller and GPU could have embedded memory dedicated for frame buffer (ATI Xenos used in XBox360 has 10MB eDRAM - through it is not perfect example, as it is also controller of whole XBox360 memory). Anyway - even current Intel's FSB (in theory) has enough bandwidth for low-end GPU solution (of course - GPU on FSB is really sick idea).
However the longer I think about it, less reasonable such idea looks for me.
>>Moving it to GPU-die looks like natural consequence - it could be even "cut-down"
>>northbridge with CSI ports. However memory controller in CPU-die would reduce latency
>>for CPU and GPU will rather not suffer too much about it (it rather need lots of
>>bandwidth).
>
>Right, but then the memory controller might wake up the CPU, or require a higher
>voltage. To do this efficiently you need separate PLLs, voltage planes and VRMs for the memory controller and cores.
>
I think that Barcelona has already such feature.
Best regards,
Marcin N.
---------------------------
>I'd put the memory controller with the GPU because it make sense to have the GPU
>use the system memory as the frame buffer.
>
Agree - it should be simpler when GPU manages all of the physical memory. Still CPU could be memory controller and GPU could have embedded memory dedicated for frame buffer (ATI Xenos used in XBox360 has 10MB eDRAM - through it is not perfect example, as it is also controller of whole XBox360 memory). Anyway - even current Intel's FSB (in theory) has enough bandwidth for low-end GPU solution (of course - GPU on FSB is really sick idea).
However the longer I think about it, less reasonable such idea looks for me.
>>Moving it to GPU-die looks like natural consequence - it could be even "cut-down"
>>northbridge with CSI ports. However memory controller in CPU-die would reduce latency
>>for CPU and GPU will rather not suffer too much about it (it rather need lots of
>>bandwidth).
>
>Right, but then the memory controller might wake up the CPU, or require a higher
>voltage. To do this efficiently you need separate PLLs, voltage planes and VRMs for the memory controller and cores.
>
I think that Barcelona has already such feature.
Best regards,
Marcin N.