ScholarWorks@중앙대학교

초록

Rapid advancements in artificial intelligence, particularly large language models, have escalated the need for efficient parallel processing, straining traditional sequential instruction single data CPUs and single instruction multi-data GPUs due to sequential bottlenecks and data dependencies. This paper presents the Every One Period Parallel Processor (EOPPP), a novel multi-core multi-instruction multi-data (MIMD) architecture, designed to overcome these limitations. EOPPP achieves true parallelism by pre-arranging data, processing it within a single clock cycle, and enabling data reusability, validated through performance evaluations on a 128 -core FPGA prototype at 0.05 GHz. Comparative analyses against an Intel i7 6core CPU demonstrate superior efficiency in dependency-heavy Fibonacci sequence computations and high-throughput Finite Impulse Response (FIR) filters. These results highlight EOPPP potential to revolutionize computing by mitigating Amdahl’s Law constraints, reducing total cost of ownership, and simplifying parallel software development via its eFLOW language and automated compiler.

키워드