Paul A. Clayton (paaronclayton.delete@this.gmail.com) on July 27, 2012 10:31 am wrote:
> jp (jipe4153.delete@this.gmail.com) on July 27, 2012 9:47 am wrote:
> > aaron
> spink (aaronspink.delete@this.notearthlink.net) on July 27, 2012 9:36 am
> >
> wrote:
> [snip]
> >> GPUs have high local bandwidth but rather poor global
>
> >> bandwidth. Not to mention rather limited capacity.
> >
> >
> Global bandwidth on GPUs is much higher than on any other CPU ( 5-6 times
> >
> higher), you seem to have no backing for your claim.
>
> Are there GPUs with more
> than 16 lanes of PCIe?
>
> One 3.2 GHz QPI link might have comparable bandwidth
> to 16 lanes of PCIe 3.0 (6.4 GT/s * 20 lanes vs. 8 GT/s * 16), the higher-end
> Intel processors have more than one QPI link.
>
> I assume you call "global" the
> what Aaron Spink calls "local", perhaps using the the OpenCL terminology of
> global memory and local memory. Aaron Spink comes from a CPU/NUMA system
> background (IIRC he worked on Alpha EV7 coherence) where global memory refers to
> the entire system memory and local memory refers to the memory attached directly
> to a single node.

That's true, we may well be talking about different things.

Local memory in OpenCL speak is what is known as shared memory in CUDA, which can be thought of as a fast on-chip user managed cache.

With PCIe 3.0 cards the max PCI express bandwidth is now 16 GB/s.