By: Kevin G (kevin.delete@this.cubitdesigns.com), April 25, 2013 6:48 am
Room: Moderated Discussions
Antti-Ville Tuunainen (avtuunainen.delete@this.gmail.com) on April 23, 2013 5:25 pm wrote:
> Mark Roulo (nothanks.delete@this.xxx.com) on April 23, 2013 3:20 pm wrote:
> > You have capacity limitations and memory bottlenecks at several places. So ...
> >
> > The Intel IGPU has a bandwidth advantage going to system DRAM because it shares the memory controller (and
> > access to the system DRAM) with the CPU. The discrete part is bandwidth limited because of PCIe.
> >
> > I get how more memory (1 GB vs 128 MB) at fast speeds (64 GB/sec +) is good, but I'd also think that access
> > to the rest of the system memory at faster speeds (25 GB/sec vs 5-10 GB/sec) would be good too.
> >
> > Any idea how these two factors compete? I don't game and have no idea ...
>
> All present-day games have been designed and programmed with the latency and bandwidth of the
> PCIe link in mind. I personally believe that bringing the GPU closer to the CPU will be a major
> win that will help much more than just raw power, but that is only *after* games can rely that
> they are indeed close. Right now, the PCIe link is just a common design constraint, meaning
> that bringing the GPU and CPU closer together buys you *nothing* on present games.
>
> PS4/the next XBOX should help with this, as they are APU designs that will teach game
> programmers how to exploit such architectures. For the shelf life of Haswell, I don't
> believe that a single game will exist that takes proper advantage of this.
The PCI-E link speed is not a limitation in modern PC architectures though. Only when running multiple video cards does moving from 8x to 16x lane widths show a significant increase in performance. Similarly, multiple GPU systems are where PCI-E 3.0 shows an advantage over PCI-E 2.0 for 8x widths. There is a minimum amount of bandwidth necessary but it is relatively low compared to what a full 16 lane PCI-E 3.0 bus can provide.
Logically you are correct though. Seeing the GPU has an integrated component with the same memory space will improve performance. Communication between the CPU and GPU can be done via pointers within the same address space with local memory.
> Mark Roulo (nothanks.delete@this.xxx.com) on April 23, 2013 3:20 pm wrote:
> > You have capacity limitations and memory bottlenecks at several places. So ...
> >
> > The Intel IGPU has a bandwidth advantage going to system DRAM because it shares the memory controller (and
> > access to the system DRAM) with the CPU. The discrete part is bandwidth limited because of PCIe.
> >
> > I get how more memory (1 GB vs 128 MB) at fast speeds (64 GB/sec +) is good, but I'd also think that access
> > to the rest of the system memory at faster speeds (25 GB/sec vs 5-10 GB/sec) would be good too.
> >
> > Any idea how these two factors compete? I don't game and have no idea ...
>
> All present-day games have been designed and programmed with the latency and bandwidth of the
> PCIe link in mind. I personally believe that bringing the GPU closer to the CPU will be a major
> win that will help much more than just raw power, but that is only *after* games can rely that
> they are indeed close. Right now, the PCIe link is just a common design constraint, meaning
> that bringing the GPU and CPU closer together buys you *nothing* on present games.
>
> PS4/the next XBOX should help with this, as they are APU designs that will teach game
> programmers how to exploit such architectures. For the shelf life of Haswell, I don't
> believe that a single game will exist that takes proper advantage of this.
The PCI-E link speed is not a limitation in modern PC architectures though. Only when running multiple video cards does moving from 8x to 16x lane widths show a significant increase in performance. Similarly, multiple GPU systems are where PCI-E 3.0 shows an advantage over PCI-E 2.0 for 8x widths. There is a minimum amount of bandwidth necessary but it is relatively low compared to what a full 16 lane PCI-E 3.0 bus can provide.
Logically you are correct though. Seeing the GPU has an integrated component with the same memory space will improve performance. Communication between the CPU and GPU can be done via pointers within the same address space with local memory.