Memory bandwidth

Banked Memory. Another characteristic of many vector supercomputers is banked memory. The main memory is usually divided into a small number of electronically separate banks. A given memory bank can absorb or supply operands at a much slower rate than the rate at which the central processing unit (CPU) can produce or use data. If the data can be spread across multiple memory banks, the effective memory bandwidth, or rate at which memory can absorb or supply data, is increased. For example, if a single memory bank can supply one operand every 16 clock cycles, then 16 memory banks would enable the entire memory subsystem to deflver one operand per clock cycle, assuming that the data come sequentially from different memory banks. 

System Interconnect Expandability From the standpoint of expansion limitations, the shared memory system has problems in that the number of ports are fixed. Expanders can be used to alleviate this problem to some degree, but physical construction problems are ultimately met. Also, the memory bandwidth of the shared memory system is fixed and is relatively slow, thus limiting the degree of practical expansion. 

Figure 1 Peak floating-point operations per second (a) and memory bandwidth (b) for Intel CPUs and NVIDIA GPUs. Reproduced from [15].

D. Burger, J. Goodman. Memory bandwidth limitations of future microprocessors. In Proc. 23rd Int l Symposium on Computer Architecture, May 1996, pp. 78 - 89. 

It can be observed from Fig. 7.3 that the requirement for peak memory bandwidth for NVidia GPUs has increased from 0.53 GB/s to 35.2 GB/s, or a factor of 66.7, in a period of 10 years beginning from 1994 (NVidia introduced its first GPU, NV1, in 1994). This momentum has to be continued with the adoption of advanced graphic features such as the 128-bit floating-point color depth and DVD quality real-time computer games. 

Figure 7.3 Peak memory bandwidths of major NVidia GPUs...

Figure 7.5 Normalized clock rate vs. peak memory bandwidth of NVidia...

The shallowness and serrated behavior of tm make extensive optimization unnecessary. The precise location of this minimum depends on the relative computation times for the potential energy evaluations in the real and reciprocal spaces. These computation times can depend substantially on (i) the algorithms employed (type of neighbor list, direct or tabulated energy evaluation in the real space, implementation of the evaluation of the reciprocal term, single or parallel computation, etc.) and (ii) the architecture of the computer used (CPU computation speed, cache sizes, memory bandwidth, etc.), making such an optimization useful when changing code, compiler, or computer hardware. 

Like other multiscale methods, atomistic-continuum methods require an accurate treatment of the coupling between different domains. In addition to these difficulties, they pose serious challenges for performing extensive simulations. The physical processes described by continuum equations and particle-based models impose inherently distinct demands on the computer architerture. While continuum mechanics and hydrodynamics, typically dealing with regular meshes, are characterized by moderate computations with stractured communications, atomistic simulations are characterized by intense computations and intense interprocessor communications. As a result, large-scale simulations of this sort require a balanced computer architecture in terms of memory bandwidth and interconnea bandwidth. 

The color lookup table implements 3 look-up tables with 256 x 10 bit entries each. These can be used for different kinds of applications. For non-linear color transformations (called gamma correction), the color components R, G and B are processed independendy. The input code is used as a table index and the table entry as a 10-bit output color code. The alpha value inputs are by-passed (unchanged) through the color lookup table. The second application is to use the table as an indirect color palette (lookup index). To implement reduced memory bandwidth, the index is interpreted to a 10-bit color in the palette. 

---