Conventional compute-in-memory (CIM) deployment flows usually assume a fixed crossbar geometry, although DNN layers often have different channel counts and matrix shapes. The resulting shape mismatch leaves part of the array capacity unused and can increase the number of split-and-transfer operations during inference. This paper presents FlexiCIM, a compiler framework for neural network deployment on multi-core CIM systems. FlexiCIM groups physical cores into virtual computing units (VCUs), maps reshaped layer weights under CIM capacity constraints, and schedules dependent tasks while accounting for communication overhead. The framework also uses an evolutionary search procedure to select VCU partitions for latency-first or throughput-first objectives. Experiments on five CNN models show that FlexiCIM achieves an average utilization of 75.5% under the default setting. In the latency-first setting, FlexiCIM reduces latency by up to 54.8% and obtains the lowest normalized energy among the evaluated designs (1.00 vs. 2.02/1.37/1.60 for Fixed-L/M/S). In the throughput-first setting, FlexiCIM provides the highest throughput on all five models, with $1.07\times-1.38\times$ speedup over the best fixed-size baseline. These results indicate that compiler-managed VCU partitioning and capacity-aware mapping improve deployment efficiency on multi-core CIM architectures.
Kaiwen Deng, Sifan Sun, Hanjie Liu et al.· IEEE Non-Volatile Memory Sys...· 0 citations
ReRAM-based in-memory computing (IMC) architectures enable efficient neural-network inference on edge devices, due to their high density and non-volatility. However, ReRAM is prone to the stuck-at faults (SAFs) which would distort weight mappings and reduce model accuracy substantially. To tackle this problem, a high-efficiency fault-resilient (HE-FR) framework is introduced in the paper, which couples layered-precision quantization based on the generalized gaussian distribution cumulative distribution function (GGD-CDF), preserving weight distributions with the minimal information loss. Further, a fault-aware weight re-decomposition method is proposed. Specifically, a Q-agent-based method is proposed for sparse SAFs, which employs offline reinforcement learning to construct a globally optimized hash table with reduced compilation complexity, while incorporating both weight matching and cell-state stability into the reward function to suppress electro-stress. For dense SAFs, we exploit a greedy search over the remaining cells to provide rapid and accurate mappings. Experimental results show that GGD-CDF quantization improves accuracy by 2.18% over Float32. Under fabrication faults, HE-FR surpasses a fault-free method by 12.97% and accelerates compilation by $197\times $ with only 0.024% picojoule-level energy overhead. In electro-stress evaluations, HE-FR reduces performance degradation by 46.27%. It demonstrates that the proposed framework has advantages in robustness, efficiency, and reliability.
Junlong Lin, Wenhua Wu, Mingtao Wang et al.· IEEE Transactions on Circuit...· 0 citations
