By: David Kanter (dkanter.delete@this.realworldtech.com), September 14, 2007 9:50 pm
Room: Moderated Discussions
Marcin Niewiadomski (marcin.niewiadomski@gmail.com) on 8/29/07 wrote:
---------------------------
>Hi David,
>
>The article is simply excellent. Thank You for very good >and hard work.
Thank you very much!
>I'd add few words regarding the speculation part. Most probably we can expect first
>Intel attempt to put CPU and GPU in one package as multi-die solution (like Presler,
>Kentsfield, etc. - same is expected for AMD Fusion) and the think which make me
>wonder is whenever it is better to integrate memory controler with CPU-die or GPU-die?
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.
>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.
>The drawback of such approach is that we will have quite a lot of traffic
>between CPU and GPU. Both approaches seem to possible - even a mixed variant (1
>channel for CPU, 1 for GPU), but I don't think that it will go that way.
>
>Once again - thank You for excellent article - it was >pleasure to read.
Absolutely!
DK
---------------------------
>Hi David,
>
>The article is simply excellent. Thank You for very good >and hard work.
Thank you very much!
>I'd add few words regarding the speculation part. Most probably we can expect first
>Intel attempt to put CPU and GPU in one package as multi-die solution (like Presler,
>Kentsfield, etc. - same is expected for AMD Fusion) and the think which make me
>wonder is whenever it is better to integrate memory controler with CPU-die or GPU-die?
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.
>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.
>The drawback of such approach is that we will have quite a lot of traffic
>between CPU and GPU. Both approaches seem to possible - even a mixed variant (1
>channel for CPU, 1 for GPU), but I don't think that it will go that way.
>
>Once again - thank You for excellent article - it was >pleasure to read.
Absolutely!
DK