Spzip Page

SpZip outperforms state-of-the-art accelerators by a geometric mean of 3.0× (up to 5.2×).

It uses a specialized Dataflow Configuration Language (DCL) to specify how data structures are traversed and generated, allowing it to adapt to diverse sparse formats (CSR, DCSR, COO, etc.). SpZip decouples the data structure traversal from the

Neighbor sets in a graph are rarely the same size. However, modern irregular applications (e

SpZip decouples the data structure traversal from the main processor core, allowing the system to "fetch and decompress" ahead of time, hiding the high latency of memory accesses. modern irregular applications (e.g.

Despite its capability, SpZip is designed to be inexpensive, adding only about 0.2% area overhead to each core.

Traditional compression methods excel at repetitive, sequential data. However, modern irregular applications (e.g., BFS, PageRank, graph algorithms) exhibit:

Data is scattered, making it hard to compress efficient, large contiguous blocks.