Multicore, the Memory Wall, and Numerical Compression

Why is multicore processor utilization decreasing?  A worrying trend has emerged in supercomputing, which deploys thousands of multicore CPU and GPU sockets for big data applications, foreshadowing future problems with multicore. As a percentage of peak Mflops, today's supercomputers are less than 10% utilized. The reason is simple: input-output (I/O) has not nearly kept pace with multicore MIPS.  While cores per die scales with Moore's Law, interconnect to off-chip memory, system busses, networks, and disk drives does not.  Multicore CPUs and GPUs are hitting the memory wall.  We present a novel numerical compression algorithm for integers, floating-point values, still images, and video that increases DDRx, PCIe, Ethernet, and SAS/SATA bandwidth by 2x to 6x at wire speeds, with complete user control and without changing the final results of the computation.

Speaker: Al Wegener, Samplify Systems

Register at website.