Michael S (already5chosen.delete@this.yahoo.com) on November 11, 2014 9:23 am wrote:
> Huge memory capacity of Xeon-E5 is important for other applications, but for HPC capacity of Xeon-E3
> is probably sufficient. Taking as example, TSUBAME-KFC that you mentioned earlier, it has 2,560 GB
> of RAM = 64 GB per node = 32 GB per socket = 16 GB per GPU. Xeon-E3 can address 32 GB per socket.
>
> As to bandwidth, if we build system with 1 GPU per Xeon-E3 then it will have the
> same theoretical peak CPU-side bandwidth per GPU as TSUBAME-KFC, but slightly higher
> practically achievable bandwidth, due to lower latency of unbuffered DIMMs.
>
> I think, that it's not memory capacity or bandwidth that keeps Xeon-E3 away from Top500,
> but what I said in my previous post - bigger nodes are simply more economical.
>
> BTW, you say that XG1 can address vastly more memory compared to Xeon E3. I don't know if it is true in
> general, but XG1-based HP ProLiant m400 supports only 64 GB, which is twice more than Xeon-E3 and AMD Opteron
> X1150, the same as Intel Atom C2730 and many times less than even the most castrated Xeon-E5.
>
>
>

Agree both technical and economic factors ultimitately dictate if ARM adoption for GPU based supercomputers will actually happen or not. But the fact HPC solution builders like Cirrascale and E4 are clearly on board with XG1, give at least some credibility to the idea.

XG1 with 4 memory channels supports 512 GB according to http://arxiv.org/pdf/1410.3441.pdf

Ronald