Energy Efficient Matrix Computations Through Homomorphic Compression

Scientific computing is one of the most energy-intensive areas of computer science. Petabytes of data are processed and stored for a single experiment while super-computers consume tens of megawatts. These last years, new lossy compression techniques dedicated to scientific data have been introduced, such as zfp and sz. They allow one to dras-tically reduce the size of the data but they require additional computations for compression and decompression. Recently, a new homomorphic compressor, named blaz, has been developed which makes it possible to perform matrix operations directly among the compression data. Important gains are obtained since decompression and re-compression operations are avoided and since less operations are needed to compute among the compressed data. In this article, we show that using blazfor linear algebra operations among matrices may reduce the energy consumption by a factor 10 compared to standard operations (without any compression) and by a factor 100 compared to the zfp. Beside measures on sequences of matrix operations, a case study in data analysis in presented.