GDDR5 memory for HPC/server applications

[magellan]

Would there be any HPC/server apps that could take advantage of a system that had the GDDR5 memory bandwidth of the PS4? This is assuming you could get systems w/a larger memory footprint than the PS4.

I know the HPC apps that are run on the Rocks and Xcat SGE clusters I've worked on have huge, multi-terabyte datasets, so I'd imagine low latency memory access might help with these type of applications (molecular research using NAMD being the one I'm somewhat familiar with).

I wonder if AMD might not try leveraging this GDDR5 tech in their server market?

 

[EarthDog]

Maybe. But in most cases, servers are not memory bandwidth starved in the first place, but are way more likely to be than a home PC, that is for sure.

Perhaps more speed would work better than lower latency? Remember with memory, 1600MHz CL7 is around equal to say 2133 CL9 in terms of speed, right? So in order to reach the GDDR5 speeds, latency is actually higher. I also think it is quite cost prohibitive compared to DDR3.

 

[magellan]

Maybe. But in most cases, servers are not memory bandwidth starved in the first place, but are way more likely to be than a home PC, that is for sure.

Perhaps more speed would work better than lower latency? Remember with memory, 1600MHz CL7 is around equal to say 2133 CL9 in terms of speed, right? So in order to reach the GDDR5 speeds, latency is actually higher. I also think it is quite cost prohibitive compared to DDR3.

1600Mhz = .626ns cycle time, 2122Mhz = .469ns cycle time. 9 * .469ns = 4.22ns of latency. 7 * .626ns = 4.38ns. Is CL7 commonplace for 1600Mhz? I've never had any DDR3 memory that ran CL7 at stock.

Channel interleaving and bank interleaving help to minimize latencies.

The author of the SB-E overclocking guide did some informal testing of
the latencies and bandwidth of memory running at various speeds, he
found that optimized, tightened memory timings at 1600Mhz weren't sufficient to overcome the additional clock speed of unoptimized 1800Mhz memory in terms of both latency and bandwidth.

While I see your point that the timings are much looser for GDDR5 memory, the period for each wait state is much smaller. The actual difference in latencies could be minimal.