Differences
This shows you the differences between two versions of the page.
| Both sides previous revisionPrevious revision | |||
| seminars [2025-08-12] – [XPLUTO: a theoretical model and algorithms for bulk-parallel in-memory processing] Martin Ziegler | seminars [2025-08-13] (current) – [XpLUTo: Modelling Bulk Parallel Processing in RAM via Lookup Tables] Martin Ziegler | ||
|---|---|---|---|
| Line 20: | Line 20: | ||
| * Nguyên Trần Bảo (HCMUT and KAIST) | * Nguyên Trần Bảo (HCMUT and KAIST) | ||
| + | Processing-in-memory (PIM) has been investigated for its ability to | ||
| + | perform bulk data operations while eliminating data movement, which is a major | ||
| + | performance bottleneck. However, existing designs, regardless of how minimal, | ||
| + | still require modifications to the physical memory circuitry. Moreover, each | ||
| + | proposed operation introduces different primitives, inherently hindering the development | ||
| + | of a general design capable of supporting all operations. In this work, | ||
| + | we propose XPLUTO, a new parallel architecture model that leverages the capabilities | ||
| + | of PIM. Our key observation is that, in the worst case, any complex | ||
| + | operation can be implemented via a lookup table (precomputation and query), | ||
| + | which can be viewed as a SIMD (single-instruction multiple-data) operation. | ||
| + | Based on this insight, we focus on designing algorithms built upon SIMD operations, | ||
| + | with asymptotic costs estimated according to lookup table performance. | ||
| + | So far, XPLUTO has demonstrated the ability to emulate various problems, | ||
| + | including sorting, addition, and prefix operations. | ||