Memory Architecture

Multi-Level PIM/NDP Architectures

Challenge: PIM core can be placed at bank level, off-chip buffer level and even inside the SSD controller, and the communication between different levels of PIM/NDP is challenging.

Year	Venue	Authors	Title	Tags	P	E	N
2026	arXiv	ICT	PAM: Processing Across Memory Hierarchy for Efficient KV-centric LLM Serving System	HBM-PIM/DDR-PIM/SSD-PIM heterogeneous 3-tier request scheduling; PAMattention via online softmax for distributed token-wise parallelism; importance-aware greedy KV scheduling for load balancing	4	4	3
2026	HPCA	ETHZ	Conduit: Programmer-Transparent Near-Data Processing Using Multiple Compute-Capable Resources in Solid State Drives	loop auto-vectorization to align with SSD page layout; instruction-granularity offloading via holistic cost function	2	3	2

In-Stroage (NAND Flash) Processing

Near-NAND Array Processing

Solution: bypass the NAND Flash controller chip and DRAM/SLC buffer, execute the tasks by near-array units.

Year	Venue	Authors	Title	Tags	P	E	N
2025	arXiv	Korea Univ.	Dissecting and Re-architecting 3D NAND Flash PIM Arrays for Efficient Single-Batch Token Generation in LLMs	re-architected 3D-NAND PIM array with H-tree network; QLC-SLC hybrid architecture for KV caching; static/dynamic MVM tiling and mapping	4	2	3
2025	ISCA	Seoul National	AiF: Accelerating On-Device LLM Inference Using In-Flash Processing	in-flash GEMV computation; charge-recycling read to skip precharge/discharge steps in flash memory	3	3	4

General Application Targeted Optimization

Solution: Intergrate the compute unit into the SSD controller to process the capacity-sensitive applications.

Year	Venue	Authors	Title	Tags	P	E	N
2024	HPCA	UCLA	BeaconGNN: Large-Scale GNN Acceleration with Out-of-Order Streaming In-Storage Computing	DirectGraph format for out-of-order sampling; die-level processing units; channel-level command router	4	2	3
2025	ISCA	ETHZ	REIS: A High-Performance and Energy-Efficient Retrieval System with In-Storage Processing	In-Storage processing	2	4	3
2025	ISCA	UCSD	In-Storage Acceleration of Retrieval Augmented Generation as a Service	metamorphic in-storage accelerator; Metadata Navigation Unit for dynamic data access	4	3	2
2025	arxiv	ETHZ	MARS: Processing-In-Memory Acceleration of Raw Signal Genome Analysis Inside the Storage Subsystem	PIM module inside the SSD controller; early signal quantization; read filtering	3	3	2
2025	ICCAD	SNU	LLM-on-the-Palm: Mobile LLM Inference with PIM-Enhanced NAND Flash Memory	a single MAC unit per plane; selective layer-wise mapping strategy offloading FC layers to PIM and attention to NPU; pipelined MAC and input broadcast via extended commands	2	2	2

LLM-Specific Optimizations

Solution: Store weights in flash memory as read-only to prevent failures caused by write operations.

Year	Venue	Authors	Title	Tags	P	E	N
2024	arXiv	ICT	Cambricon-LLM: A Chiplet-Based Hybrid Architecture for On-Device Inference of 70B LLM	chiplet-based NPU & NAND flash hybrid architecture; Hardware-aware tiling for NPU-flash workload distribution	4	3	2
2025	HPCA	THU	Lincoln: Real-Time 50~100B LLM Inference on Consumer Devices with LPDDR-Interfaced, Compute-Enabled Flash Memory	flash-on-LPDDR-interface for prefill phase; hybrid-bonding-based near-Flash computing for generation phase	3	4	3
2025	HPCA	PKU	InstAttention: In-Storage Attention Offloading for Cost-Effective Long-Context LLM Inference	offloading decoding-phase attention computation to computational SSDs; SparF Attention flash-aware sparse algorithm	4	2	2
2025	ASPLOS	ETHZ	CIPHERMATCH: Accelerating Homomorphic Encryption-Based String Matching via Memory-Efficient Data Packing and In-Flash Processing	memory-efficient data packing scheme for HE; in-NAND XOR & AND operations	3	4	2
2026	ASPLOS	Seoul National	A Cost-Effective Near-Storage Processing Solution for Offline Inference of Long-Context LLMs	delayed KV Cache writeback avoiding write amplification; cooperative X-cache for cooperative utilization	2	2	2

DIMM-PIMs

Challenge: Memory wall causing high latency of data transfer between CPU and memory.

Solution: Put the compute unit in the memory or near the memory to reduce the data transfer overhead.

General Application-Specific Optimization

Challenge: Existing NDP architecture are designed for general-purpose computing; not efficient for specific tasks like graph processing.

Year	Venue	Authors	Title	Tags	P	E	N
2022	ISCA	Micron	To PIM or Not for Emerging General Purpose Processing in DDR Memory Systems	vector engine inside NDP bank; intelligent code offload decision	2	3	2
2023	EuroSys	Univ. of Virginia	NearPM: A Near-Data Processing System for Storage-Class Applications	partitioned persist ordering for asynchronous CPU-NDP execution; delayed synchronization for multi-device consistency	4	3	3
2024	ISCA	Samsung	pSyncPIM: Partially Synchronous Execution of Sparse Matrix Operations for All-Bank PIM Architectures	partially synchronous PIM control; predicated execution; sparse matrix distribution & compaction	3	3	3
2025	ATC	RUC	Turbocharge ANNS on Real Processing-in-Memory by Enabling Fine-Grained Per-PIM-Core Scheduling	per-PU scheduling; persistent PIM kernel; per-PU dispatching with selective replication	3	4	4
2025	HPCA	UC Davis	NOVA: A Novel Vertex Management Architecture for Scalable Graph Processing	message-driven processors capable of executing algorithms; a direct-mapped cache with a write-back policy; support both asynchronous and bulk synchronous parallel execution models	3	3	3

DNN-Specific Optimization

Year	Venue	Authors	Title	Tags	P	E	N
2021	HPCA	Seoul National	GradPIM: A Practical Processing-in-DRAM Architecture for Gradient Descent	fixed-function PIM architecture for DNN gradient descent; non-invasive PIM operations using reserved DDR commands	3	3	2
2024	ASPLOS	PKU	PIM-DL: Expanding the Applicability of Commodity DRAM-PIMs for Deep Learning via Algorithm-System Co-Optimization	algorithm for DNN to look-up-table conversion; auto-tuner for optimizing LUT-NN mapping on DRAM-PIMs	3	4	3
2025	arXiv	National Tech Univ. of Athens	PIMfused: Near-Bank DRAM-PIM with Fused-layer Dataflow for CNN Data Transfer Optimization	hybrid dataflow combining fused-layer and layer-by-layer strategies	3	2	2
2026	HPCA	Seoul National	LoCaLUT: Harnessing Capacity-Computation Tradeoffs for LUT-Based Inference in DRAM-PIM	operation-packed LUT canonicalization via multiset indexing; reordering LUT for weight permutation remapping; LUT slice streaming for DRAM-buffer hierarchy	4	4	3
2026	HPCA	Seoul National	RoMe: Row Granularity Access Memory System for Large Language Models	row-granularity access interface for LLM streaming; virtual bank to eliminate bank groups and pseudo channels; logic-die command generator for C/A pin reduction and simpler MC	4	3	3

Graph-Specific Optimization

Challenge: Graph processing is fundamentally limited by memory bandwidth and requires frequent random accesses, which are not efficiently supported by non-interleaved, bank-level PIM architectures.

Year	Venue	Authors	Title	Tags	P	E	N
2022	PACT	PKU	GNNear: Accelerating Full-Batch Training of Graph Neural Networks with Near-Memory Processing	splitting reduce operations to NDP units; narrow-shard strategy for data reuse; hybrid graph partition strategy for load balancing	4	3	3
2025	HPCA	ZJU	GOPIM: GCN-Oriented Pipeline Optimization for PIM Accelerators	ML-based replica resource allocation for pipeline streamlining; interleaved mapping with adaptive selective vertex updating	2	3	3
2025	MICRO	Seoul National	FALA: Locality-Aware PIM-Host Cooperation for Graph Processing with Fine-Grained Column Access	8-byte-level granularity for fine-grained vertex access with HBM2-PIM; multiple non-contiguous column accesses within single activated DRAM row	2	2	2
2026	ASPLOS	Uppsala	CoGraf: Fully Accelerating Graph Applications with Fine-Grained PIM	tuple-based LLC for coalescing at flexible granularity; multi-column Fine-Grained PIM instructions to utilize row-level parallelism; predicated bank-parallel instructions for conditional apply-phase operations	3	3	3

LLM-Specific Optimization

Challenge: LLM inference is fundamentally bottlenecked by memory bandwidth; HBM is expensive and not scalable.

Year	Venue	Authors	Title	Tags	P	E	N
2024	npj Unconv. Comput.	UMich	PIM-GPT: a hybrid process in memory accelerator for autoregressive transformers	hybrid system to accelerate GPT inference; mapping scheme for data locality and workload distribution	3	2	2
2025	MICRO	KAIST	PIMBA: A Processing-in-Memory Acceleration for Post-Transformer Large Language Model Serving	Unified PIM acceleration for both transformer and post-transformer LLMs; access interleaving technique for shared State-update Processing Unit	4	2	3
2025	MICRO	Cornell	LongSight: Compute-Enabled Memory to Accelerate Large-Context LLMs via Sparse Attention	hybrid dense-sparse attention algorithm; KV cache offloading to CXL-PIM; sign-concordance filtering by iterative quantization	3	2	2
2025	IEEE LCA	Yonsei	RoPIM: A Processing-in-Memory Architecture for Accelerating Rotary Positional Embedding in Transformer Models	bank-level PIM accelerator for RoPE; row-level data mapping for single-token Q/K/V/S alignment; parallel data rearrangement via inverters	4	3	4

LLM Quantization & Capacity Optimization

Challenge: Bank-level PIM systems face tight capacity budgets and quantization overhead in long-context LLM serving.

Year	Venue	Authors	Title	Tags	P	E	N
2025	IEEE LCA	POSTECH	Cost-Effective Extension of DRAM-PIM for Group-Wise LLM Quantization	scale cascading for simplifying dequantization; zero-offset removal for reduced circuit complexity	3	2	2
2026	HPCA	ISCT	AQPIM: Breaking the PIM Capacity Wall for LLMs with in-Memory Activation Quantization	lookup-based attention; importance-weighted K-means clustering; channel pre-sorting for subvector affinity	3	3	4

MoE-LLM Optimization

Challenge: MoE models often have higher Op/Byte ratios, making bank-level PIM easily compute-bound and limiting speedup.

Year	Venue	Authors	Title	Tags	P	E	N
2024	DAC	Seoul National	MoNDE: Mixture of Near-Data Experts for Large-Scale Sparse Models	activation movement strategy to replace costly parameter movement; dynamic GPU-MoNDE load balancing for hot/cold experts	4	4	2
2025	MICRO	Samsung	Duplex: A Device for Large Language Models with Mixture of Experts, Grouped Query Attention, and Continuous Batching	replace the GPU HBM memory die with HBM-PIM die; expert and attention co-processing for dynamic workload splitting within MoE/attn layers	4	4	4
2025	ICCAD	PKU	HD-MoE: Hybrid and Dynamic Parallelism for Mixture-of-Expert LLMs with 3D Near-Memory Processing	LP-based hybrid TP and EP mapping; bayesian optimization for topology-aware link balancing; online dynamic expert placement with predictive pre-broadcast	3	3	2

RAG-Specific Optimization

Challenge: Retrieving the top-k results from a vectorized database is also a memory-bound operation.

Year	Venue	Authors	Title	Tags	P	E	N
2025	ISCA	HUST	HeterRAG: Heterogeneous Processing-in-Memory Acceleration for Retrieval-augmented Generation	combine DIMM-PIM and HBM-PIM for acceleration; locality-aware retrieval and generation; fine-grained parallel pipelining	2	3	3
2025	MICRO	Yonsei	Accelerating Retrieval Augmented Language Model via PIM and PNM Integration	heterogeneous architecture integrating PIM for LLMs and PNM for retrievers; RALM scheduling strategy with selective batching and early generation	4	2	3

PIM KVCache Mapping

Challenge: Bank-level PIM processes one row at a time, needs data locality, making the data interleaving in HBM and host memory inefficient.

KV-Cache Mapping & Offloading

Year	Venue	Authors	Title	Tags	P	E	N
2026	ASPLOS	SJTU	REPA: Reconfigurable PIM for the Joint Acceleration of KV Cache Offloading and Processing	reconfigurable ReRAM-PIM for KV cache offload and in-situ processing; bulk-wise memory setting instructions for wordline parallelism; locality-aware data mapping and transfer overlapping	2	2	2
2026	ASPLOS	RPI	STARC: Selective Token Access with Remapping and Clustering for Efficient LLM Decoding on PIM Systems	semantic KV clustering for PIM row-level alignment; hardware-friendly cosine K-means via PIM primitives; incremental append-only remapping for KV cache sparsity	3	3	2

Memory Address Space

Challenge: Host pages need to enable interleaving to improve concurrent throughput, while PIM pages need to disable it to maintain better locality, creating a conflict.

Year	Venue	Authors	Title	Tags	P	E	N
2023	DAC	Georgia Tech	vPIM: Efficient Virtual Address Translation for Scalable Processing-in-Memory Architectures	network-contention-aware hashing to minimize cross-stack page table walks; pre-translation using repurposed PIM cores to move page table walks off the critical path	4	4	3
2024	ISCA	SJTU	UM-PIM: DRAM-based PIM with Uniform & Shared Memory Space	Uniform shared CPU-PIM memory; dual-track memory management; zero-copy data re-layout	3	3	4

Memory Allocation & Management

Challenge: Existing NDP architecture has numerous independent memory spaces; lacks unified management; and features inefficient memory allocation.

Year	Venue	Authors	Title	Tags	P	E	N
2024	ISCA	KAIST	PIM-malloc: A Fast and Scalable Dynamic Memory Allocator for Processing-In-Memory (PIM) Architectures	PIM-specific memory allocator; hierarchical memory allocation scheme; hardware metadata cache	4	2	3
2024	arXiv	ETHZ	PUMA: Efficient and Low-Cost Memory Allocation and Alignment Support for Processing-Using-Memory Architectures	aligned memory allocator for PUM; DRAM-aware memory allocation	2	3	2
2024	MICRO	KAIST	PIM-MMU: A Memory Management Unit for Accelerating Data Transfers in Commercial PIM Systems	data copy engine for host-PIM transfers; PIM-aware memory scheduler for MLP maximization; memory remapping unit for dual address mapping	2	4	3
2025	arXiv	Amazon	DL-PIM: Improving Data Locality in Processing-in-Memory Systems	subscription-based architecture to proactively move data; distributed address-indirection hardware lookup table	3	2	3

PIM ISA & Programming Model

Challenge: Existing PIM interfaces expose limited control and make it hard to express end-to-end execution inside memory.

Year	Venue	Authors	Title	Tags	P	E	N
2015	ISCA	Seoul National	PIM-Enabled Instructions: A Low-Overhead; Locality-Aware Processing-in-Memory Architecture	PIM-Enabled Instructions for ISA extension; PIM directory for atomicity and coherence; single-cache-block restriction	3	4	4
2020	ISCA	UCSB	iPIM: Programmable In-Memory Image Processing Accelerator Using Near-Bank Architecture	Single-Instruction-Multiple-Bank ISA; register allocation; instruction reordering	4	4	2
2026	HPCA	UIUC	The Memory Processing Unit: A Generalized Interface for End-to-End In-Memory Execution	ensemble execution model; thermal-aware round-robin VRF scheduler; comparison-based in-memory predication with arbitrary nesting	4	3	4

PIM Compiler & Data Layout

Challenge: Existing compilers are not optimized for locality-aware PIM architectures and require specialized programming models to fully utilize PIM capabilities.

Year	Venue	Authors	Title	Tags	P	E	N
2025	ISCA	POSTECH	ATIM: Autotuning Tensor Programs for Processing-in-DRAM	autotuning framework for DRAM PIM; search-based optimizing tensor compiler; balanced evolutionary search algorithm	3	3	4
2025	ISCA	ETHZ	OptiPIM: Optimizing Processing-in-Memory Acceleration Using Integer Linear Programming	layout-aware nested loop representation; Integer Linear Programming formulation for PIM mapping; analytical cost modeling for data layout enforcement	3	3	4
2026	HPCA	Hanyang	PIMphony: Overcoming Bandwidth and Capacity Inefficiency in PIM-based Long-Context LLM Inference System	token-centric PIM partitioning for high channel utilization; dynamic PIM command scheduling for out-of-order I/O-compute overlap; dynamic PIM access for dynamic virtual-to-physical translation	3	4	4

Evaluation & Simulators

Year	Venue	Authors	Title	Tags	P	E	N
2025	HPCA	THU	UniNDP: A Unified Compilation and Simulation Tool for Near DRAM Processing Architectures	unified NDP hardware abstraction; NDP compiler optimization; instruction-driven NDP simulator	3	5	2
2025	arXiv	ETHZ	EasyDRAM: An FPGA-based Infrastructure for Fast and Accurate End-to-End Evaluation of Emerging DRAM Techniques	FPGA-based DRAM evaluation framework; C++ high-level language for description; time scaling for accurate modeling	3	4	3
2026	arXiv	PKU	A Full-Stack Performance Evaluation Infrastructure for 3D-DRAM-based LLM Accelerators	ATLAS framework for hybrid bounding 3D DRAM NMP; hierarchical SPMD/MPMD programming model; grid-based transient thermal analyzer for temperature-constrained architecture exploration	4	4	5

Intra-DIMM Communication

Challenge: High latency of intra-DIMM (cross-bank) communication via host CPU forwarding.

Year	Venue	Authors	Title	Tags	P	E	N
2024	ISCA	THU	NDPBridge: Enabling Cross-Bank Coordination in Near-DRAM-Bank Processing Architectures	gather & scatter messages via buffer chip; task-based message-passing model; hierarchical, data-transfer-aware load balancing
2025	HPCA	Samsung	Piccolo: Large-Scale Graph Processing with Fine-Grained In-Memory Scatter-Gather	In-DRAM fine-grained scatter-gather via data bus offsets; fine-grained cache architecture using fg-tags; Standard DDR command interpretation for FIM control; Combined graph tiling with fine-grained memory access	3	3	4
2024	arXiv	Seoul National	PID-Comm: A Fast and Flexible Collective Communication Framework for Commodity Processing-in-DIMM Devices	Virtual hypercube PIM model; PE-assisted data reordering; in-register and cross-domain data modulation	3	4	3
2025	ISCA	KAIST	PIMnet: A Domain-Specific Network for Efficient Collective Communication in Scalable PIM	domain-specific PIM interconnect; hierarchical network for PIM packaging; PIM-controlled deterministic scheduling	2	4	3
2026	HPCA	ICT	COCOTree: A Computation-Capable Architecture for Collective Communication in Scalable PIM	hierarchical binary tree topology for inter-PE communication; in-network computation via computation-capable nodes; two-phase packet-based protocol with configuration-computation decoupling	4	3	2

Inter-DIMM Communication

Challenge: High latency of inter-DIMM (cross-DIMM) communication via host CPU forwarding.

Year	Venue	Authors	Title	Tags	P	E	N
2017	MEMSYS	UCLA	AIM: Accelerating Computational Genomics through Scalable and Noninvasive Accelerator-Interposed Memory	placing FPGA chip between DIMM and the conventional memory network; multi-drop bus for inter-accelerator communication	1	2	2
2023	ASPLOS	THU	ABNDP: Co-optimizing Data Access and Load Balance in Near-Data Processing	Traveller Cache; hybrid task scheduling; hybrid scheduling leveraging distributed cache	4	3	4
2023	HPCA	PKU	DIMM-Link: Enabling Efficient Inter-DIMM Communication for Near-Memory Processing	high-speed hardware link bridges between DIMMs; direct intra-group P2P communication & broadcast; hybrid routing mechanism for inter-group communication
2025	HPCA	SJTU	AsyncDIMM: Achieving Asynchronous Execution in DIMM-Based Near-Memory Processing	Offload-Schedule-Return mechanism; switch-recovery scheduling; explicit/implicit synchronization	2	4	3
2018	MICRO	UIUC	Application-Transparent Near-Memory Processing Architecture with Memory Channel Network	integrates a processor on a buffered DIMM; application-transparent near-memory processing; leverages memory channels for high-bandwidth/low-latency inter-processor communication	3	4	4

CPU-PIM Heterogeneous Systems

Challenge: High latency of concurrent host CPU and PIM operations via host CPU forwarding.

Year	Venue	Authors	Title	Tags	P	E	N
2024	IEEE CA	KAIST	Analysis of Data Transfer Bottlenecks in Commercial PIM Systems: A Study With UPMEM-PIM	runtime data transposition causing high CPU overhead; PIM-integrated system memory mapping impact	2	2	2
2025	ISCA	Univ. of Virginia	Membrane: Accelerating Database Analytics with Bank-Level DRAM-PIM Filtering	bank-level DRAM-PIM filtering; CPU-PIM cooperative query execution; denormalization for PIM-amenable filtering	3	3	2
2025	MICRO	Inha University	ComPASS: A Compatible PIM Protocol Architecture and Scheduling Solution for Processor-PIM Collaboration	PIM-ACT new memory command for multi-bank PIM operations; PIM request generator to offload host processor; static and adaptive throughput balancers for PIM and non-PIM request scheduling	4	2	2
2025	ASPLOS	SJTU	PUSHtap: PIM-based In-Memory HTAP with Unified Data Storage Format	PIM-specific HTAP storage data format; semi-interleaved data layout for CPU and PIM concurrent data access	2	3	3
2026	DAC	ICT	TriMoE: Augmenting GPU with AMX-Enabled CPU and DIMM-NDP for High-Throughput MoE Inference via Offloading	tri-domain offloading architecture coordinating GPU/AMX-enabled CPU/DIMM-NDP; bottleneck-aware greedy expert scheduling; prediction-driven expert relayout and rebalancing via DIMM-Link	2	4	3

NPU-PIM Heterogeneous Systems

Challenge: Data transfer between NPU and PIM needs to go through the Host, causing high latency.

Year	Venue	Authors	Title	Tags	P	E	N
2024	ASPLOS	KAIST	NeuPIMs: NPU-PIM Heterogeneous Acceleration for Batched LLM Inferencing	dual row buffer architecture; sub-batch interleaving; greedy min-load bin packing algorithm	3	4	3
2025	MICRO	SJTU	HEAT: NPU-NDP HEterogeneous Architecture for Transformer-Empowered Graph Neural Networks	topology-aware mixed-precision encoding for transformer; subgraph bundling and reordering for GNN memory efficiency; decoupled dataflow for NPU-NDP concurrent execution	2	3	3
2025	ICCAD	PKU	LP-Spec: Leveraging LPDDR PIM for Efficient LLM Mobile Speculative Inference with Architecture-Dataflow Co-Optimization	GEMM-enhanced hybrid LPDDR5 PIM; near-data memory controller for concurrent NPU-PIM execution & data reallocation; hardware-aware draft token pruning	3	4	2
2025	arXiv	Cornell	P3-LLM: An Integrated NPU-PIM Accelerator for LLM Inference Using Hybrid Numerical Formats	low-precision PIM compute unit with temporal data reuse; operator fusion for quantized dataflow to minimize dequantization overhead	4	3	2

GPU-PIM Heterogeneous Systems

Challenge: Weight, activation and KV Cache data mapping in conventional GPU-PIM systems is naive, causing space & bandwidth waste and load imbalance.

Year	Venue	Authors	Title	Tags	P	E	N
2024	DAC	Hunan Univ.	A Real-time Execution System of Multimodal Transformer through PIM-GPU Collaboration	dynamic strategy for PIM-GPU task offloading; variable-length-aware PIM allocation optimizer; extended TVM backend for PIM-GPU command generation	3	3	3
2025	HPCA	ICT	Make LLM Inference Affordable to Everyone: Augmenting GPU Memory with NDP-DIMM	activation sparsity-based hot(GPU)/cold(NDP) neuron partitioning; offline ILP + online predictor for neuron partition; window-based online remapping for GPU-NDP & NDP-NDP load balance	2	3	4
2025	arXiv	Hanyang Univ.	Scalable Processing-Near-Memory for 1M-Token LLM Inference: CXL-Enabled KV-Cache Management Beyond GPU Limits	hybrid TP-DP parallelism for GPU-PNM systems; token page selection; steady-token selection	2	3	4

Optimizations on UPMEM-PIM

Challenge: The original UMPEM API library is not well-suited for all workloads especially for those with cross-bank communication.

Year	Venue	Authors	Title	Tags	P	E	N
2023	arXiv	ETHZ	A Framework for High-throughput Sequence Alignment using Real Processing-in-Memory Systems	Alignment-in-Memory framework; hybrid WRAM-MRAM sketch data management for PIM	2	3	4
2025	arXiv	ETHZ	PIMDAL: Mitigating the Memory Bottleneck in Data Analytics using a Real Processing-in-Memory System	PIMDAL library on UPMEM PIM system for data analytics; scatter/gather-aware transfers for inter-PIM communication; Apache Arrow for host memory management	3	3	3

In-Cache-Computing

Year	Venue	Authors	Title	Tags	P	E	N
2025	arXiv	Torino	ARCANE: Adaptive RISC-V Cache Architecture for Near-memory Extensions	ARCANE in-cache NMC coprocessor architecture; software-defined matrix ISA for NMC abstraction; cache-integrated control runtime for NMC management	3	4	4

PIM & NDP Benchmarks

Challenge: Conventional parallel computing benchmarks are not suitable for PIM/NDP.

Benchmarks for Conventional Computing

Year	Venue	Authors	Title	Tags	P	E	N
2021	ATC	UBC	A Case Study of Processing-in-Memory in off-the-Shelf Systems	benchmark
2022	IEEE Access	ETH	Benchmarking a New Paradigm: Experimental Analysis and Characterization of a Real Processing-in-Memory System	benchmark suite PrIM
2024	CAL	KAIST	Analysis of Data Transfer Bottlenecks in Commercial PIM Systems: A Study With UPMEM-PIM	low MLP; manual data placement; unbalanced thread allocation and scheduling
2024	IEEE Access	Lisbon	NDPmulator: Enabling Full-System Simulation for Near-Data Accelerators From Caches to DRAM	simulator PiMulator based on Ramulator & gem5; full system support; multiple ISA support
2024	HPCA	KAIST	Pathfinding Future PIM Architectures by Demystifying a Commercial PIM Technology	simulator uPIMulator

Benchmarks for Quantum Computing

Year	Venue	Authors	Title	Tags	P	E	N
2025	ASPDAC	NUS	PIMutation: Exploring the Potential of PIM Architecture for Quantum Circuit Simulation	PIMutation framework for quantum circuit simulation; gate merging optimization; row swapping instead of matrix multiplication; vector partitioning for separable states; leveraging UPMEM PIM architecture

CXL-Based PIM

Challenge: No direct physical connectivity between the banks in the DIMM-based NDP architecture. Limited number of DDR channels causing poor scalability.

Solution: Introduce CXL-based interconnects to enable direct communication between memory banks; Use CXL memory pools and CXL switches to enable scalable NDP architecture.

Year	Venue	Authors	Title	Tags	P	E	N
2022	MICRO	UCSB	BEACON: Scalable Near-Data-Processing Accelerators for Genome Analysis near Memory Pool with the CXL Support	scalable hardware accelerator inside CXL switch or bank; lossless memory expansion for CXL memory pools
2024	HPCA	Samsung	An LPDDR-based CXL-PNM Platform for TCO-efficient Inference of Transformer-based Large Language Models	LPDDR5X-based CXL memory module; Processing-Near-Memory controller; software stack via direct access driver for transparent host-accelerator memory sharing	2	2	4
2024	ICS	Samsung	CLAY: CXL-based Scalable NDP Architecture Accelerating Embedding Layers	direct interconnect between DRAM clusters; dedicated memory address mapping scheme; Multi-CLAY system support through customized CXL switch
2024	MICRO	SK Hyrix	Low-overhead General-purpose Near-Data Processing in CXL Memory Expanders	CXL.mem protocol instead of CXL.io (DMA) for low-latency; lightweight threads to reduce address calculation overhead
2025	ISCA	Seoul National	COSMOS: A CXL-Based Full In-Memory System for Approximate Nearest Neighbor Search	CXL core-based ANNS task offload; rank-level parallel distance computation; adjacency-aware data placement algorithm	2	2	2
2025	ASPLOS	UMich	PIM Is All You Need: A CXL-Enabled GPU-Free System for Large Language Model Inference	hierarchical CXL PIM-PNM compute architecture; use die-shot to estimate area cost; multiple LLM parallelism policies	2	3	3

Chiplet-Based PIM

Challenge: The logic units of near-bank DRAM PIM are fabricated using the same process node as DRAM, which restricts their performance and efficiency.

Year	Venue	Authors	Title	Tags	P	E	N
2022	IEEE TCAD	WSU	SWAP: A Server-Scale Communication-Aware Chiplet-Based Manycore PIM Accelerator	coarse-to-fine multi-objective optimization algorithm for Network-On-Package design; communication-aware irregular NoP topology	3	2	2
2025	arXiv	Univ. of Virginia	Sangam: A Chiplet-Based DRAM-PIM Accelerator with CXL Integration for LLM Inferencing	decoupling logic and memory dies by chiplet; bank-level systolic arrays for flat-GEMM acceleration	4	3	4
2025	ICCAD	SEU	H3D-LLM: Heterogeneous 3D Chiplet Design for LLM Inference with Dynamic Task Scheduling and Memory-Aware Orchestration	Sparse-aware dynamic execution with precision-adaptive quantization; runtime-aware encoding and dynamic cluster allocation	3	3	2

3D-Stacked PIM

Challenge: There is no direct physical interconnection paths in DIMM-based, bank-level uniform NDP like UPMEM.

Solution: Put the logical, computational layer at the bottom of the die, and stack DRAM layers on top of it. Use TSVs to build thousands of physical paths between the logical and the DRAM layers.

Hybrid Bounding-Based PIM

Solution: Hybrid Bounding provides massively increased interconnect density and bandwidth by direct copper-to-copper connection.

Year	Venue	Authors	Title	Tags	P	E	N
2024	ISCA	THU	Exploiting Similarity Opportunities of Emerging Vision AI Models on Hybrid Bonding Architecture	clustering similarity effect architecture for hybrid bonding DRAM; hotspot content SRAM for parallel similarity detection; progressive sparsity detection and balance for computation skipping and workload redistribution	4	4	2
2025	ISCA	PKU	H2-LLM: Hardware-Dataflow Co-Exploration for Heterogeneous Hybrid-Bonding-based Low-Batch LLM Inference	operator-channel binding; computation-bandwidth trade-off; dataflow-based DSE	4	3	3
2025	MICRO	THU	3D-PATH: A Hierarchy LUT Processing-in-memory Accelerator with Thermal-aware Hybrid Bonding Integration	sparse-aware hierarchical slow-fast LUT design; multiplier-free floating-point operation by LUT; hotspot-aware hardware with self-throttling sense amplifier	4	2	2
2025	MICRO	Univ. of British Columbia	RayN: Ray Tracing Acceleration with Near-memory Computing	ray tracing units in 3D stacked DRAM logic layer; BLAS Breaking algorithm to partition BVH tree for load balancing; hybrid memory controller for concurrent GPU and near-memory access	4	3	2
2025	ICCAD	PKU	FENIX: Flexible and Efficient Hybrid HE/MPC Acceleration with Near-Memory Processing	fine-grained oblivious transfer partitioning to overlap HE and OT operations; batch-aware flexible encoding to reduce rotation overhead; near-bank NMP to offload memory-bound HE/OT primitives	3	2	2

HMC or HMC-like PIM

Challenge: No direct physical connectivity between the banks in the DIMM-based NDP architecture.

Solution: Use TSVs to provide TB/s level bandwidth in inter-bank communication & band-to-logic layer communication.

Year	Venue	Authors	Title	Tags	P	E	N
2013	PACT	KAIST	Memory-centric System Interconnect Design with Hybrid Memory Cubes	memory-centric network; distributor-based topology for reduced latency; non-minimal routing for higher throughput
2021	DAC	UCSD	MAT: Processing In-Memory Acceleration for Long-Sequence Attention	iterative tiled processing to reduce memory footprint of O(N^2) score matrices; late softmax update to enable pipelined attention; dynamic programming-based sample scheduling for optimal PIM resource utilization	3	2	2
2024	DAC	SNU	MoNDE: Mixture of Near-Data Experts for Large-Scale Sparse Models	NDP for MoE; activation movement; GPU-MoNDE load-balancing scheme
2024	ASPLOS	PKU	SpecPIM: Accelerating Speculative Inference on PIM-Enabled System via Architecture-Dataflow Co-Exploration	algorithmic and architectural heterogeneity; PIM resource allocation; multi-model collaboration workflow
2025	MICRO	UCSD	Stratum: System-Hardware Co-Design with Tiered Monolithic 3D-Stackable DRAM for Efficient MoE Serving	monolithic 3D-Stackable DRAM without TSV; in-memory tiering for vertical latency variation; topic-based expert usage prediction for MoE serving	4	3	3

HBM2-PIM

Solution: HBM2-PIM is the first commercial HBM-PIM product, and 4 out of 8 DRAM layers are PIM-enabled layers, while the other 4 layers are standard DRAM layers.

Year	Venue	Authors	Title	Tags	P	E	N
2021	ISCA	Samsung	Hardware Architecture and Software Stack for PIM Based on Commercial DRAM Technology Industrial Product	drop-in replacement for standard HBM2; bank-level parallelism using standard DRAM commands; address aligned mode to tolerate host-side command reordering	3	5	3
2022	Hot Chips	Samsung	Aquabolt-XL HBM2-PIM, LPDDR5-PIM With In-Memory Processing, and AXDIMM With Acceleration Buffer	HBM2-PIM with bank-level SIMD programmable computing units; Acceleration DIMM with acceleration buffers for rank-level parallelism	2	5	3
2023	MICRO	Yonsei	AESPA: Asynchronous Execution Scheme to Exploit Bank-Level Parallelism of Processing-in-Memory	Single-Instruction Long-Data execution model; asynchronous bank operation via long-data commands; column-major GEMV dataflow with shared accumulators	3	3	3

HBM3-PIM

Solution: HBM3-PIM further increases the capacity and the parallelism by stacking more DRAM layers and all the layers are the PIM-enabled layer.

Year	Venue	Authors	Title	Tags	P	E	N
2024	ASPLOS	Seoul National	AttAcc! Unleashing the Power of PIM for Batched Transformer-based Generative Model Inference	bank-level GEMV & buffer-die Softmax units; head-level pipelining; feedforward co-processing optimizations	4	4	4
2025	ICS	Korea Univ.	SparsePIM: An Efficient HBM-Based PIM Architecture for Sparse Matrix-Vector Multiplications	DRAM row-aligned format; bounded cap K-means clustering for load balancing; bank group accumulators	3	3	2

Benchmarks

Year	Venue	Authors	Title	Tags	P	E	N
2019	DAC	ETHZ	NAPEL: Near-Memory Computing Application Performance Prediction via Ensemble Learning	simulator Ramulator-PIM; tracefile from Ramulator & run on zsim
2021	CAL	UVA	MultiPIM: A Detailed and Configurable Multi-Stack Processing-In-Memory Simulator	simulator MultiPIM; multi-stack & virtual memory support; parallel offloading

PIM: Heterogeneous Architecture

Challenge: Different PIM architectures have different characteristics and performance trade-offs; communicating between different PIM architectures is challenging.

Year	Venue	Authors	Title	Tags	P	E	N
2025	arXiv	NUS	LEAP: LLM Inference on Scalable PIM-NoC Architecture with Balanced Dataflow and Fine-Grained Parallelism	data dynamicity-aware task assignment to PIM or NoC; fine-grained model partitioning and heuristically optimized spatial mapping strategy	3	4	3
2025	arXiv	THU	CompAir: Synergizing Complementary PIMs and In-Transit NoC Computation for Efficient LLM Acceleration	heterogeneous DRAM-PIM and SRAM-PIM architecture with hybrid bonding; in-transit NoC computation with Curry ALU; hierarchical ISA for hybrid PIM systems	3	4	2
2025	arXiv	BUAA	HPIM: Heterogeneous Processing-In-Memory-based Accelerator for Large Language Models Inference	SRAM-PIM/HBM-PIM heterogeneous LLM accelerator; compiler-based workload partitioning; intra-token SRAM-HBM pipeline	4	3	3

General CiM

Specific Application & Algorithm

Year	Venue	Authors	Title	Tags	P	E	N
2024	ISVLSI	USC	Multi-Objective Neural Architecture Search for In-Memory Computing	neural architecture search methodology; integration of Hyperopt, PyTorch and MNSIM
2024	arXiv	Intel	CiMNet: Towards Joint Optimization for DNN Architecture and Configuration for Compute-In-Memory Hardware	framework that jointly searches for optimal sub-networks and hardware configurations for CiM architectures; multi-objective evolutionary search method	4	2	4
2025	AICAS	UVA	Optimizing and Exploring System Performance in Compact Processing-in-Memory-based Chips	Pipeline Method for Compact PIM Designs; Dynamic Duplication Method (DDM); Maximum NN Size Estimation & Deployment in Compact PIM Design
2025	NeurIPS	IBM	Analog Foundation Models	synthetic-data distillation for analog LLM adaptation; iterative weight clipping for high-SNR conductance mapping; static DAC/ADC range learning; per-channel hardware-noise injection	4	3	4

Modeling & Simulation

Challenge: Need fast and accurate estimators for area, latency, energy, and system-level behavior across varied CiM hardware designs.

Year	Venue	Authors	Title	Tags	P	E	N
2018	TCAD	ASU	NeuroSim: A Circuit-Level Macro Model for Benchmarking Neuro-Inspired Architectures in Online Learning	estimate the circuit-level performance of neuro-inspired architectures; estimates the area, latency, dynamic energy, and leakage power; Support both SRAM and eNVM; tested on 2-layer MLP NN, MNIST
2020	TCAD	ZJU	Eva-CiM: A System-Level Performance and Energy Evaluation Framework for Computing-in-Memory Architectures	models for capturing memory access and dependency-aware ISA traces; models for quantifying interactions between the host CPU and the CiM module
2022	ICCAD	Purdue	Design Space and Memory Technology Co-Exploration for In-Memory Computing Based Machine Learning Accelerators	simulation framework to evaluate the systemlevel performance of IMC architecture; area-aware weight mapping strategy	4	3	2
2024	ISPASS	MIT	CiMLoop: A Flexible, Accurate, and Fast Compute-In-Memory Modeling Tool	flexible specification to describe CiM systems; accurate model/fast statistical model of data-value-dependent component energy
2025	ASPDAC	HKUST	MICSim: A Modular Simulator for Mixed-signal Compute-in-Memory based AI Accelerator	modulared Neurosim; data statistic-based average-mode instead of trace-based mode	4	3	2

Robustness-Aware Modeling & Training

Challenge: Device and peripheral nonidealities make it hard to predict and recover deployment accuracy on analog CiM hardware.

Year	Venue	Authors	Title	Tags	P	E	N
2019	IEDM	Georgia Tech	DNN+NeuroSim: An End-to-End Benchmarking Framework for Compute-in-Memory Accelerators with Versatile Device Technologies	a python wrapper to interface NeuroSim; for inference only
2023	Nat. Commun.	IBM	Hardware-aware training for large-scale and diverse deep learning inference workloads using in-memory computing-based accelerators	hardware-aware retraining with PCM-noise injection; learned input/output dynamic-range scaling; standardized AIMC crossbar model; topology robustness sensitivity analysis	4	3	4

Compiler

Challenge: Compiler for CIM is not well studied. Existing compilers are either for specific architecture or not efficient.

Year	Venue	Authors	Title	Tags	P	E	N
2023	TACO	HUST	A Compilation Tool for Computation Offloading in ReRAM-based CIM Architectures	compilation tool to migrate legacy programs to CPU/CIM heterogeneous architectures; a model to quantify the performance gain
2023	DAC	CAS	PIMCOMP: A Universal Compilation Framework for Crossbar-based PIM DNN Accelerators	compiler based on Crossbar/IMA/Tile/Chip hierarchy; low latency and high throughput mode; genetic algorithm to optimize weight replication and core mapping; scheduling algorithms for complex DNN
2024	ASPLOS	CAS	CIM-MLC: A Multi-level Compilation Stack for Computing-In-Memory Accelerators	compilation stack for various CIM accelerators; multi-level DNN scheduling approach
2024	DATE	RWTH Aachen University	CLSA-CIM: A Cross-Layer Scheduling Approach for Computing-in-Memory Architectures	algorithm to decide which parts of NN are duplicated to reduce inference latency; cross layer scheduling on tiled CIM architectures
2024	TCAD	NJU	A Compilation Framework for SRAM Computing-in-Memory Systems With Optimized Weight Mapping and Error	input/output side parallelism (IOSP); partition-based MAQE(duplicating MSB storage)	4	2	3

CIM: DRAM

Solution: Rather than placing logic units into DRAM; modify the physical structure of DRAM/eDRAM to enable in-memory computing.

DRAM CIM: General Architecture

Year	Venue	Authors	Title	Tags	P	E	N
2021	ICCD	ASU	CIDAN: Computing in DRAM with Artificial Neurons	Threshold Logic Processing Element (TLPE) for in-memory computation; Four-bank activation window; Configurable threshold functions; Energy-efficient bitwise operations; Integration with DRAM architecture
2022	HPCA	UCSD	TransPIM: A Memory-based Acceleration via Software-Hardware Co-Design for Transformer	token-based dataflow for general Transformer-based models; ring-based data broadcast in modified HBM	4	2	4
2025	arXiv	UTokyo	MVDRAM: Enabling GeMV Execution in Unmodified DRAM for Low-Bit LLM Acceleration	GeMV operations for end-to-end low-bit LLM inference using unmodified DRAM; processor-DRAM co-design; on-the-fly vector encoding; horizontal matrix layout	4	4	3
2025	arXiv	Purdue	HALO: Memory-Centric Heterogeneous Accelerator with 2.5D Integration for Low-Batch LLM Inference	heterogeneous CiD/CiM accelerator; phase-aware mapping strategy	3	2	2

DRAM CIM: eDRAM-Based Design & Optimization

Year	Venue	Authors	Title	Tags	P	E	N
2024	A-SSCC	UNIST	A 273.48 TOPS/W and 1.58 Mb/mm2 Analog-Digital Hybrid CIM Processor with Transpose Ternary-eDRAM Bitcell	analog DRAM CIM for partial sum and digital adder	1	4	2
2025	arXiv	KAIST	RED: Energy Optimization Framework for eDRAM-based PIM with Reconfigurable Voltage Swing and Retention-aware Scheduling	RED framework for energy optimization; reconfigurable eDRAM design; retention-aware scheduling; trade-off analysis between RBL voltage swing, sense amplifier power, and retention time; refresh skipping and sense amplifier power gating

CIM: SRAM

Challenge: Memory wall causing high latency of data transfer between CPU and memory; DIMM-based NDP causing high energy consumption; area overhead and low performance efficiency.

Solution: Generally modify the physical structure of SRAM to enable in-memory computing; rather than placing logic units into SRAM.

SRAM CIM: Charge-Domain Architecture

Challenge: Charge-domain SRAM CIM improves energy efficiency, but multi-bit analog operation is constrained by ADC precision limits and error amplification during reconstruction.

Solution: Rework charge-domain encoding, accumulation, or readout structures so analog SRAM CIM can preserve accuracy without giving up its efficiency advantage.

Year	Venue	Authors	Title	Tags	P	E	N
2026	HPCA	ICT	Cambricon-CIM: Enabling Energy-Efficient and Error-Resilient Analog CIM Acceleration via Reformation of Coding Bases	minimal non-binary coding bases for multi-bit slicing; runtime vector centering for dynamic-range compression; LUT-based base-length selection for charge-domain SRAM CIM	4	2	3
2024	ESSCIRC	THU	A 65nm 8b-Activation 8b-Weight SRAM-Based Charge-Domain Computing-in-Memory Macro Using A Fully-Parallel Analog Adder Network and A Single-ADC Interface	SRAM-based CD-CiM architecture; charge-domain analog adder tree; ReLU-optimized ADC	4	4	4
2018	JSSC	MIT	CONV-SRAM: An Energy-Efficient SRAM With In-Memory Dot-Product Computation for Low-Power Convolutional Neural Networks	SRAM-embedded convolution (dot-product) computation architecture for BNN; support multi-bit input-output

SRAM CIM: General Architecture

Year	Venue	Authors	Title	Tags	P	E	N
2024	ISCAS	NYCU	CIMR-V: An End-to-End SRAM-based CIM Accelerator with RISC-V for AI Edge Device	incorporates CIM layer fusion, convolution/max pooling pipeline, and weight fusion; weight fusion: pipelining the CIM convolution and weight loading
2021	ISSCC	TSMC	An 89TOPS/W and 16.3TOPS/mm2 All-Digital SRAM-Based Full-Precision Compute-In Memory Macro in 22nm for Machine-Learning Edge Applications	programmable bit-widths for both input and weights; SRAM and CIM mode	2	5	1
2021	JSSC	KAIST	Z-PIM: A Sparsity-Aware Processing-in-Memory Architecture With Fully Variable Weight Bit-Precision for Energy-Efficient Deep Neural Networks	bit-serial operation to support variable weight bit-precision; data mapping and computation flow for sparsity handling	3	4	4

SRAM CIM: Reconfigurable Macro

Year	Venue	Authors	Title	Tags	P	E	N
2021	JSSC	UCSB	Colonnade: A Reconfigurable SRAM-Based Digital Bit-Serial Compute-In-Memory Macro for Processing Neural Networks	reconfigurable column MAC; sparse pipelining	4	4	4
2023	TCSI	UCSB	A 1-16b Reconfigurable 80Kb 7T SRAM-Based Digital Near-Memory Computing Macro for Processing Neural Networks	radix-4 booth encoding; bit-serial reconfigurable logic	4	4	3
2024	TCSI	CAS	A 1-8b Reconfigurable Digital SRAM Compute-in-Memory Macro for Processing Neural Networks	Decompose-Accumulate-and-Shift(DAS); Reconfigurable Digital Arithmetic Unit (RDAU); input-sparsity driven clock gating	4	4	3

SRAM CIM: Specific Use or Application

Year	Venue	Authors	Title	Tags	P	E	N
2023	TCAS-I	UIC	MC-CIM: Compute-in-Memory With Monte-Carlo Dropouts for Bayesian Edge Intelligence	SRAM-based CIM macros to accelerate Monte-Carlo dropout; compute reuse between consecutive iterations
2024	DAC	GWU	Addition is Most You Need: Efficient Floating-Point SRAM Compute-in-Memory by Harnessing Mantissa Addition	decomposing FP mantissa multiplication into sub-ADD and sub-MUL; hybrid-domain SRAM CIM architecture	3	3	2
2025	A-SSCC	Georgia Tech	A 28nm 1.80Mb/mm2 Digital/Analog Hybrid SRAM-CIM Macro Using 2D-Weighted Capacitor Array for Complex Number Mac Operations	Hybrid DCIM/ACIM SRAM; lightweight correction schemes; complex CIM-SRAM units	2	4	2
2025	arXiv	GWU	Unicorn-CIM: Uncovering the Vulnerability and Improving the Resilience of High-Precision Compute-in-Memory	SRAM-CIM for FP DNNs; a fault-injection framework for FP DNNs; a ECC scheme for FP DNNs	3	2	3
2025	ISCAS	KAUST	Reconfigurable Precision INT4-8/FP8 Digital Compute-in-Memory Macro for AI Acceleration	parallel-input approach; mantissa parallel-alignment technique	3	2	2

SRAM CIM: Hardware-Software Co-Design

Year	Venue	Authors	Title	Tags	P	E	N
2022	TCAD	NTHU	MARS: Multi-macro Architecture SRAM CIM-Based Accelerator with Co-designed Compressed Neural Networks	sparsity algorithm designed for SRAM CiM; quantization algorithm with BN fusion	3	3	2
2023	TCAD	UCSB	SDP: Co-Designing Algorithm, Dataflow, and Architecture for In-SRAM Sparse NN Acceleration	double-broadcast hybridgrained pruning method; bit-serial booth inSRAM (BBS) multiplication dataflow	3	3	2
2024	TCAD	BUAA	DDC-PIM: Efficient Algorithm/Architecture Co-Design for Doubling Data Capacity of SRAM-Based Processing-in-Memory	doubling the equivalent data capacity of SRAM-based PIM; FCC algorithm to obtain bitwise complementary filters	4	4	2
2024	TCASAI	Purdue	Algorithm Hardware Co-Design for ADC-Less Compute In-Memory Accelerator	reduce ADC overhead in analog CiM architectures; Quantization-Aware Training; Partial Sum Quantization; ADC-Less hybrid analog-digital CiM hardware architecture HCiM	3	3
2025	TCAD	BUAA	Efficient SRAM-PIM Co-design by Joint Exploration of Value-Level and Bit-Level Sparsity	hybrid-grained pruning algorithm; customized Dyadic Block PIM (DB-PIM) architecture	4	3	2

SRAM CIM: Simulator & Modeling

Year	Venue	Authors	Title	Tags	P	E	N
2020	ISCAS	JCU	MemTorch: A Simulation Framework for Deep Memristive Cross-Bar Architectures	supports both GPUs and CPUs; integrates directly with PyTorch; simulate non-idealities of memristive devices within cross-bar, tested on VGG-16, CIFAR-10
2021	TCAD	Geogia Tech	DNN+NeuroSim V2.0: An End-to-End Benchmarking Framework for Compute-in-Memory Accelerators for On-Chip Training	non-ideal device properties of NVMS' effect for on-chip training	3	3	2
2025	DAC	BUAA	CIMFlow: An Integrated Framework for Systematic Design and Evaluation of Digital CIM Architectures	workflow for implementing and evaluating DNN workloads on digital CIM architectures; CIM-specific ISA design; compilation flow built on the MLIR infrastructure	4	2	3

SRAM CIM: Transformer Accelerator

Challenge: Transformer architecture is widely used in NLP and CV tasks. Existing SRAM CIM architectures are not suitable for transformer acceleration.

Year	Venue	Authors	Title	Tags	P	E	N
2025	DATE	PKU	Leveraging Compute-in-Memory for Efficient Generative Model Inference in TPUs	architecture model and simulator for CIM-based TPUs; designed for LLM inference	4	2	4
2023	arXiv	Keio	An 818-TOPS/W CSNR-31dB SQNR-45dB 10-bit Capacitor-Reconfiguring Computing-in-Memory Macro with Software-Analog Co-Design for Transformers	Capacitor-Reconfiguring analog CIM architecture	1	4	3
2025	arXiv	Purdue	Hardware-Software Co-Design for Accelerating Transformer Inference Leveraging Compute-in-Memory	SRAM based softmax-friendly CIM architecture for transformer; finer-granularity pipelining strategy	4	3	2
2025	arXiv	PKU	Leveraging Compute-in-Memory for Efficient Generative Model Inference in TPUs	Energy-efficient CIM core integration in TPUs (replace the original MXU); CIM-MXU with systolic data path; Array dimension scaling for CIM-MXU; Area-efficient CIM macro design; Mapping engine for generative model inference
2024	JSSC	THU	MulTCIM: Digital Computing-in-Memory-Based Multimodal Transformer Accelerator With Attention-Token-Bit Hybrid Sparsity	long reuse elimination scheduler (LRES) to dynamically reshape the attention matrix; runtime token pruner (RTP) to remove insignificant tokens; modal-adaptive CIM network (MACN) to dynamically divide CIM cores into Pipeline; effective-bits-balanced CIM (EBBCIM) macro architecture	5	4	3

SRAM CIM: Design Space Exploration

Year	Venue	Authors	Title	Tags	P	E	N
2026	arXiv	SoutheastU	CIM-Tuner: Balancing the Compute and Storage Capacity of SRAM-CIM Accelerator via Hardware-mapping Co-exploration	RL-driven CIM HW/SW co-exploration; SRAM-CIM compute-storage balancing	3	3	2

CIM: RRAM

Challenge: RRAM devices are non-volatile and have high density; suitable for CIM applications. However; RRAM devices have non-ideal effects that can cause significant performance degradation.

RRAM CiM: Simulator

Year	Venue	Authors	Title	Tags	P	E	N
2018	TCAD	THU	MNSIM: Simulation Platform for Memristor-Based Neuromorphic Computing System	reference design for largescale neuromorphic accelerator and can also be customized; behavior-level computing accuracy model
2023	TCAD	THU	MNSIM 2.0: A Behavior-Level Modeling Tool for Processing-In-Memory Architectures	integrated PIM-oriented NN model training and quantization flow; unified PIM memory array model; support for mixed-precision NN operations
2024	DATE	UCAS	PIMSIM-NN: An ISA-based Simulation Framework for Processing-in-Memory Accelerators	event-driven simulation approach; can evaluate the optimizations of software and hardware independently

RRAM CiM: Architecture

Year	Venue	Authors	Title	Tags	P	E	N
2019	ASPLOS	Purdue & HP	PUMA: A Programmable Ultra-efficient Memristor-based Accelerator for Machine Learning Inference	Programmable and general-purpose ReRAM based ML Accelerator; Supports an instruction set; Has potential for DNN training; Provides simulator that accepts model
2018	ICRC	Purdue & HP	Hardware-Software Co-Design for an Analog-Digital Accelerator for Machine Learning	compiler to translate model to ISA; ONNX interpreter to support models in common DL frame work; simulator to evaluate performance
2024	VLSI-SoC	RWTH Aachen University	Architecture-Compiler Co-design for ReRAM-Based Multi-core CIM Architectures	inference latency predictions and analysis of the crossbar utilization for CNN
2024	arXiv	CAS	A Fully Hardware Implemented Accelerator Design in ReRAM Analog Computing without ADCs	Based on Stochastic Binary Neural Networks; Winner-Take-All (WTA) strategy; Hardware implemented sigmoid and softmax	4	3	4

RRAM CiM: Signed and High-Precision Arithmetic

Challenge: RRAM crossbars struggle with signed weights and high-precision analog representation, so numeric formats must be reshaped to preserve accuracy without excessive peripheral cost.

Year	Venue	Authors	Title	Tags	P	E	N
2021	ISCA	Northeastern et al.	FORMS: Fine-grained Polarized ReRAM-based In-situ Computation for Mixed-signal DNN Accelerator	ADMM-based fragment polarization for same-sign ReRAM columns; fine-grained sub-array computation; input zero-skipping for small fragments	4	2	4
2024	Science	USC	Programming memristor arrays with arbitrarily high precision for analog computing	represent high-precision numbers using multiple relatively low-precision analog devices;using RRAM CIM to solve PDEs	5	4	3

RRAM CiM: Architecture optimization

Year	Venue	Authors	Title	Tags	P	E	N
2023	ISCA	MIT	RAELLA: Reforming the Arithmetic for Efficient, Low-Resolution, and Low-Loss Analog PIM: No Retraining Required!	per-layer optimized weight slicing; center+offset weight encoding	3	3	3
2023	TETCI	TU Delft	Accurate and Energy-Efficient Bit-Slicing for RRAM-Based Neural Networks	unbalanced bit-slicing scheme for higher accuracy; holistic solution using 2's compliment
2024	MICRO	HUST	DRCTL: A Disorder-Resistant Computation Translation Layer Enhancing the Lifetime and Performance of Memristive CIM Architecture	address conversion method for dynamic scheduling; hierarchical wear-leveling (HWL) strategy for reliability improvement; data layout-aware selective remapping (LASR) to improve communication locality and reduce latency
2024	TC	SJTU	ERA-BS: Boosting the Efficiency of ReRAM-Based PIM Accelerator With Fine-Grained Bit-Level Sparsity	bit-level sparsity in both weights and activations; bit-flip scheme; dynamic activation sparsity exploitation scheme

RRAM CIM: Weight Mapping

Year	Venue	Authors	Title	Tags	P	E	N
2021	ICCAD	SJTU	Bit-Transformer: Transforming Bit-level Sparsity into Higher Performance in ReRAM-based Accelerator	bit-wise weight clustering; weight bit-sparsity utilization	3	3	2
2021	ICCD	SJTU	SME: ReRAM-based Sparse-Multiplication-Engine to Squeeze-Out Bit Sparsity of Neural Network	bit-plane slicing and squeezing; significance-compensated input streaming	4	3	2

RRAM CiM: Design Space Exploration

Year	Venue	Authors	Title	Tags	P	E	N
2025	arXiv	RWTH Aachen	Optimizing Binary and Ternary Neural Network Inference on RRAM Crossbars using CIM-Explorer	Tensor Virtual Machine (TVM)-based compiler; implementation of different mapping techniques; DSE flow to analyze the impact of parameters	3	3	3

RRAM CiM: Modeling

Year	Venue	Authors	Title	Tags	P	E	N
2022	ICCD	UNIST	Accurate Prediction of ReRAM Crossbar Performance Under I-V Nonlinearity and IR Drop	IRP-Net (IR Drop Prediction Network); iterative refinement	2	3	3
2023	Nature	TetraMem	Thousands of conductance levels in memristors integrated on CMOS	2048 conductance levels (11-bit); linear weight update protocol; bayesian hyperparameter optimization for inference	4	5	3
2024	AICAS	RWTH Aachen University	A Calibratable Model for Fast Energy Estimation of MVM Operations on RRAM Crossbars	system energy model for MVM on ReRAM crossbars; methodology to study the effect of the selection transistor and wire parasitics in 1T1R crossbar arrays
2024	arXiv	MIT	Modeling Analog-Digital-Converter Energy and Area for Compute-In-Memory Accelerator Design	architecture-level model that estimates ADC energy and area	4	3	3
2024	Nat. Commun.	KAUST	Hardware implementation of memristor-based artificial neural networks	automated SPICE netlist generation; Dual-Side Connection (DSC) scheme; partial in-situ training loop	3	3	2

RRAM CiM: Training optimization

Year	Venue	Authors	Title	Tags	P	E	N
2021	TCAD	SJTU	ITT-RNA: Imperfection Tolerable Training for RRAM-Crossbar-Based Deep Neural-Network Accelerator	prevent the large-weight synapses from being mapped to the imperfect memristor cells; off-device training algorithm to alleviate the accumulation of errors across multiple layers; bit-wise mechanism to compensate the resistance variations	3	3	2
2023	arXiv	UND	U-SWIM: Universal Selective Write-Verify for Computing-in-Memory Neural Accelerators	only do write-verify for important weights; based on weight second derivatives as a guide	3	3	3
2023	Adv. Mater.	UMich	Bulk‐Switching Memristor‐Based Compute‐In‐Memory Module for Deep Neural Network Training	Bulk-ReRAM based digital-CIM hybrid architecture for training; CIM for forward, digital for backward	4	4	1
2024	APIN	SWU	Multi-optimization scheme for in-situ training of memristor neural network based on contrastive learning	optimizations to the deployment method, loss function and gradient calculation; compensation measures for non-ideal effects
2025	TNNLS	SNU	Efficient Hybrid Training Method for Neuromorphic Hardware Using Analog Nonvolatile Memory	Hybrid offline-online training method

RRAM CiM: Float-Point processing

Challenge: Raw RRAM devices are not suitable for floating-point operations; while floating point data is common in DNNs (e.g. FP32).

Year	Venue	Authors	Title	Tags	P	E	N
2023	SC	UCLA	ReFloat: Low-Cost Floating-Point Processing in ReRAM for Accelerating Iterative Linear Solvers	data format and accelerator architecture
2024	DATE	UESTC	AFPR-CIM: An Analog-Domain Floating-Point RRAM -based Compute- In- Memory Architecture with Dynamic Range Adaptive FP-ADC	all-analog domain CIM architecture for FP8 calculations; adaptive dynamic range FP-ADC & FP-DAC
2025	arXiv	GWU	A Hybrid-Domain Floating-Point Compute-in-Memory Architecture for Efficient Acceleration of High-Precision Deep Neural Networks	SRAM based hybrid-domain FP CIM architecture; detailed circuit schematics and physical layouts

RRAM CiM: Convolutional Layer

Challenge: Convolutional layer is the most compute-intensive layer in CNNs. RRAM CIM architecture is quite suitable for convolutional layer operations but face challenges related to non-ideal effects and performance degradation.

Year	Venue	Authors	Title	Tags	P	E	N
2020	Nature	THU	Fully hardware-implemented memristor convolutional neural network	fabrication of high-yield, high-performance and uniform memristor crossbar arrays; hybrid-training method; replication of multiple identical kernels for processing different inputs in parallel
2019	TED	PKU	Convolutional Neural Networks Based on RRAM Devices for Image Recognition and Online Learning Tasks	RRAM-based hardware implementation of CNN; expand kernel to the size of image
2025	TVLSI	NBU	A 578-TOPS/W RRAM-Based Binary Convolutional Neural Network Macro for Tiny AI Edge Devices	ReRAM XNOR cell; BCNN CIM macro with FPGA as the control core	4	4	3

RRAM CiM: Mapping for CNN

Challenge: Efficient mapping of CNN layers onto RRAM CIM architecture is crucial for performance.

Year	Venue	Authors	Title	Tags	P	E	N
2020	TCAS-I	Georgia Tech	Optimizing Weight Mapping and Data Flow for Convolutional Neural Networks on Processing-in-Memory Architectures	weight mapping to avoid multiple access to input; pipeline architecture for conv layer calculation
2021	TCAD	SJTU	Efficient and Robust RRAM-Based Convolutional Weight Mapping With Shifted and Duplicated Kernel	shift and duplicate kernel (SDK) convolutional weight mapping architecture; parallel-window size allocation algorithm; kernel synchronization method
2023	VLSI-SoC	Aachen	Mapping of CNNs on multi-core RRAM-based CIM architectures	architecture optimized for communication; compiler algorithms for conv2D layer; cycle-accurate simulator
2023	TODAES	UCAS	Mathematical Framework for Optimizing Crossbar Allocation for ReRAM-based CNN Accelerators	formulate a crossbar allocation problem for ReRAM-based CNN accelerators; dynamic programming based solver; models the performance considering allocation problem
2025	IEEE Access	UTehran	SCiMA: A Systolic CiM-Based Accelerator With a New Weight Mapping for CNNs—A Virtual Framework Approach	kernel-major inter-crossbar weight mapping (KM-InterCWM) for convolution layers; structured pruning techniques; system-level virtual framework	4	2	2

RRAM CIM: Transformer Accelerator

Challenge: RRAM's cross-bar architecture is suitable for matrix operations.

Year	Venue	Authors	Title	Tags	P	E	N
2023	VLSI	Purdue	X-Former: In-Memory Acceleration of Transformers	in-memory accelerate attention layers; intralayer sequence blocking dataflow; provides a simulator
2024	TODAES	HUST	A Cascaded ReRAM-based Crossbar Architecture for Transformer Neural Network Acceleration	cascaded crossbar arrays that uses transimpedance amplifiers; data mapping scheme to store signed operands; ADC virtualization scheme
2023	VLSI	HUST	An RRAM-Based Computing-in-Memory Architecture and Its Application in Accelerating Transformer Inference	RRAM-based in-memory floating-point computation architecture (RIME); pipelined implementations of MatMul and softmax	3	3	4
2020	ICCAD	Duke	ReTransformer: ReRAM-based processing-in-memory architecture for transformer acceleration	MatMul does matrix decomposition in scaled dot-product attention; in-memory logic techniques for softmax; sub-matrix pipeline	4	3	3
2022	TCAD	KAIST	A Framework for Accelerating Transformer-Based Language Model on ReRAM-Based Architecture	window self-attention and window-size search algorithm; ReRAM hardware design optimized for this algorithm	4	2	3

RRAM CIM: Transformer Robustness & Hybrid Optimization

Challenge: Attention-specific robustness and hybrid-memory design add extra architectural constraints beyond baseline Transformer acceleration.

Year	Venue	Authors	Title	Tags	P	E	N
2020	ICCD	LSU	ATT: A Fault-Tolerant ReRAM Accelerator for Attention-based Neural Networks	ReRAM-based accelerator with pipeline for AttNNs; heuristic redundancy algorithm	3	2	2
2025	ISCA	UCSD	Hybrid SLC-MLC RRAM Mixed-Signal Processing-in-Memory Architecture for Transformer Acceleration via Gradient Redistribution	architectural and circuit-level hardware designs supporting importance-based data flow with hybrid SLC-MLC ReRAM; gradient redistribution technique	3	2	4

RRAM CiM: Special Usage

Year	Venue	Authors	Title	Tags	P	E	N
2019	Adv. Funct. Mater.	HUST	Functional Demonstration of a Memristive Arithmetic Logic Unit (MemALU) for In‐Memory Computing	non-volatile Boolean logic using RRAM crossbar;reconfigurable boolean logic gates	3	4	3
2024	TRETS	UFRGS	Reprogrammable Non-Linear Circuits Using ReRAM for NN Accelerators	perform typical non-linear operations using ReRAM	4	3	4
2026	Nat. Elec.	UM	Memristive cellular neural networks for fast in-pixel computing	in-pixel memristive cellular neural networks(CeNN)	4	4	3
2026	DAC	VillanovaU	CQ-CiM: Hardware-Aware Embedding Shaping for Robust CiM-Based Retrieval	codebook isolation architecture; hardware noise injection; asymmetric distance computation pipeline	3	3	3

RRAM CIM: Batchnorm

Year	Venue	Authors	Title	Tags	P	E	N
2019	ASPDAC	POSTECH	In-memory batch-normalization for resistive memory based binary neural network hardware	in-memory batchnormalization schemes; integrate BN layers on crossbar
2023	GLSVLSI	Yale	Examining the Role and Limits of Batchnorm Optimization to Mitigate Diverse Hardware-noise in In-memory Computing	non-idealities; circuit-level parasitic resistances and device-level non-idealities; crossbar-aware fine-tuning of batchnorm parameters

RRAM CiM: Matrix Equation Solver

Year	Venue	Authors	Title	Tags	P	E	N
2024	DATE	PKU	BlockAMC: Scalable In-Memory Analog Matrix Computing for Solving Linear Systems	Novel scalable algorithm for matrix equation solving; reconfigurable BlockAMC macros design	3	3	3
2025	Sci.Adv.	HUST	Fully analog iteration for solving matrix equations with in-memory computing	Analog Iteration with Digital Refinement solver	4	4	3
2025	Nat.Elec.	PKU	Precise and scalable analogue matrix equation solving using resistive random-access memory chips	Mixed-Precision Iterative Algorithm for High-Precision Analogue Computing; Scalable Hardware Implementation with BlockAMC algorithm	3	5	4

CIM: Hybrid Architecture

Solution: Use hybrid architecture (like SRAM + RRAM) to overcome the limitations of single device (e.g. RRAM's non-ideal effects).

Hybrid CIM: General-Purpose Heterogeneous Systems

Challenge: A single CIM substrate rarely supports complete applications efficiently; these works coordinate multiple compute domains or expose a broader programming interface so hybrid CIM can execute full workloads instead of isolated kernels.

Year	Venue	Authors	Title	Tags	P	E	N
2023	GLSVLSI	USC	Heterogeneous Integration of In-Memory Analog Computing Architectures with Tensor Processing Units	hybrid TPU-IMAC architecture; TPU for conv, CIM for fc
2023	NANOARCH	HUST	Heterogeneous Instruction Set Architecture for RRAM-enabled In-memory Computing	General ISA for RRAM CiM & digital heterogeneous architecture; a tile-processing unit-array three-level architecture
2025	ASPLOS	CAS	PAPI: Exploiting Dynamic Parallelism in Large Language Model Decoding with a Processing-In-Memory-Enabled Computing System	dynamic parallelism-aware task scheduling for llm decoding; online kernel characterization for heterogeneous architectures; hybrid PIM units for compute-bound and memory-bound kernels
2025	DAC	Chung-Ang Univ.	HH-PIM: Dynamic Optimization of Power and Performance with Heterogeneous-Hybrid PIM for Edge AI Devices	heterogeneous-hybrid PIM with HP/LP modules and MRAM/SRAM; dynamic data placement algorithm for energy optimization; dual PIM controller design	3	4	2
2026	ASPLOS	UIUC	DARTH-PUM: A Hybrid Processing-Using-Memory Architecture	hybrid analog-digital ReRAM PUM architecture; ACE-DCE coordinating hardware for full-kernel in-memory execution; instruction injection unit for shift-add amortization; vACore abstraction for variable-width operands	4	2	3

Hybrid CIM: Accuracy and Precision Recovery

Challenge: Analog or mixed-signal CIM loses efficiency when noise, device variation, or limited precision force conservative design; these works add digital or SRAM assistance to recover accuracy while preserving most of the CIM speed/energy benefits.

Year	Venue	Authors	Title	Tags	P	E	N
2024	Science	NTHU	Fusion of memristor and digital compute-in-memory processing for energy-efficient edge computing	Fusion of ReRAM and SRAM CiM; ReRAM SLC & MLC Hybrid; Current quantization; Weight shifting with compensation
2024	IPDPS	Georgia Tech	Harmonica: Hybrid Accelerator to Overcome Imperfections of Mixed-signal DNN Accelerators	select and transfer imperfectionsensitive weights to digital accelerator; hybrid quantization(weights on analog part is more quantized)
2024	ASP-DAC	Keio	OSA-HCIM: On-The-Fly Saliency-Aware Hybrid SRAM CIM with Dynamic Precision Configuration	On-the-fly Saliency-Aware precision configuration scheme; Hybrid CIM Array for DCIM and ACIM using split-port SRAM
2025	arXiv	AaltoU	Acore-CIM: build accurate and reliable mixed-signal CIM cores with RISC-V controlled self-calibration	reliability-focused MAC cell; proof-of-concept SoC composed of a CIM core and a RISC-V control processor; automated Built-In Self-Calibration (BISC) routine	3	3	4
2026	DATE	ESI & IBM	HILAL: Hessian-Informed Layer Allocation for Heterogeneous Analog–Digital Inference	Hessian-informed analog/digital layer allocation; k-means robust/sensitive layer clustering; layer-wise analog fine-tuning with noise injection	3	3	3

Hybrid CIM: Macro and Dataflow Co-Design

Challenge: Combining multiple memory/computing media in one accelerator introduces restore, density, and dataflow mismatches; these papers redesign the macro or mapping/dataflow stack so the hybrid substrate is actually usable.

Year	Venue	Authors	Title	Tags	P	E	N
2023	ICCAD	SJTU	TL-nvSRAM-CIM: Ultra-High-Density Three-Level ReRAM-Assisted Computing-in-nvSRAM with DC-Power Free Restore and Ternary MAC Operations	DCpower-free weight-restore from ReRAM; ternary SRAM-CIM mechanism with differential computing scheme
2025	TCAD	HKUST	Configurable Dataflow and Adaptive Mapping Optimization for Hybrid ReRAM and SRAM Compute-in-Memory Accelerator	Hybrid Macro Unit (HMU); adaptive mapping optimization; configurable dataflow control	4	3	3
2025	Nature	TSMC	A mixed-precision memristor and SRAM compute-in-memory AI processor	layer based INT-FP hybrid architure; kernel-based mix-CIM (SRAM/ReRAM/digital hybrid architecture)	5	5	2

Hybrid CIM: Transformer and Attention Acceleration

Challenge: Transformer workloads mix dense analog-friendly matrix operations with irregular, sparse, or precision-sensitive attention steps, motivating hybrid CIM architectures that split roles across analog and digital domains.

Year	Venue	Authors	Title	Tags	P	E	N
2023	arXiv	HP	RACE-IT: A Reconfigurable Analog CAM-Crossbar Engine for In-Memory Transformer Acceleration	Compute Analog Content Addressable Memory (Compute-ACAM) structure; accelerator based on crossbars and Compute-ACAMs; encoding-based optimization	3	3	4
2024	VLSI	FDU	HARDSEA: Hybrid Analog-ReRAM Clustering and Digital-SRAM In-Memory Computing Accelerator for Dynamic Sparse Self-Attention in Transformer	product-quantization-based sparse self-attention algorithm; ADC-free ReRAM-CIM macro; ReRAM-CIM for front-end attention sparsification, SRAM-CIM for back-end sparse attention	4	3	3
2024	DAC	SJTU	HEIRS: Hybrid Three-Dimension RRAM- and SRAM-CIM Architecture for Multi-task Transformer Acceleration	Hybrid Distributive Accumulation; Local Recovery Unit (LRU) in 3D	3	3	3
2024	ESSERC	UCSD	An Analog and Digital Hybrid Attention Accelerator for Transformers with Charge-based In-memory Computing	analog CIM for low-score tokens, digital processor for high	3	4	2

Hybrid CIM: LLM Adaptation and Deployment

Challenge: LLM deployment on hybrid CIM needs software-hardware partitioning that keeps large static weights dense and energy-efficient while protecting task-specific or precision-sensitive paths.

Year	Venue	Authors	Title	Tags	P	E	N
2025	arXiv	South Carolina	PIM-LLM: A High-Throughput Hybrid PIM Architecture for 1-bit LLMs	hybrid PIM-Digital architecture; analog PIM for low-precision MatMul; digital systolic array for high-precision matMul	4	3	1
2026	TODAES	HKU	HaLoRA: Hardware-aware Low-Rank Adaptation for Large Language Models Based on Hybrid Compute-in-Memory Architecture	pretrained weights on RRAM-CIM and LoRA branches on SRAM-CIM; noise-aware LoRA regularization for RRAM non-ideality; hybrid CIM energy reduction for LLM inference	4	3	3

CIM: Quantization

Challenge: Limited by the precision & area & power trade-off of the ADC; certain CIM devices like RRAM are not suitable for high-precision computation (e.g. FP32). Quantization is needed to reduce the precision of the data.

CIM: Quantization: Partial Sum Quantization

Year	Venue	Authors	Title	Tags	P	E	N
2023	ISLPED	Purdue	Partial-Sum Quantization for Near ADC-Less Compute-In-Memory Accelerators	ADC-Less and near ADC-Less CiM accelerators; CiM hardware aware DNN quantization methodology
2025	DATE	SKKU	Column-wise Quantization of Weights and Partial Sums for Accurate and Efficient Compute-In-Memory Accelerators	granularity alignment logic; learned step-size quantization(LSQ)	3	2	3

CIM Quantization: For Analog CIM

Year	Venue	Authors	Title	Tags	P	E	N
2022	TCAS-I	Georgia Tech	BitS-Net: Bit-Sparse Deep Neural Network for Energy-Efficient RRAM-Based Compute-In-Memory	bit-sparsity quantization; bias-shifted MVM; hardware-aware loss function	3	2	3
2023	AICAS	TU Delft	Mapping-aware Biased Training for Accurate Memristor-based Neural Networks	favorability constraint analysis to find important weight values; mapping-aware biased training to restrict weight values to low variance RRAM states	3	4	2
2024	TCAD	BUAA	CIMQ: A Hardware-Efficient Quantization Framework for Computing-In-Memory-Based Neural Network Accelerators	bit-level sparsity induced activation quantization; quantizing partial sums to decrease required resolution of ADCs; arraywise quantization granularity
2024	TCAD	BUAA	CIM²PQ: An Arraywise and Hardware-Friendly Mixed Precision Quantization Method for Analog Computing-In-Memory	mixed precision quantization method based on evolutionary algorithm; arraywise quantization granularity; evaluation method to obtain the performance of strategy on the CIM
2024	ICCAD	TU Delft	Hardware-Aware Quantization for Accurate Memristor-Based Neural Networks	analysis of fixed-point quantization impact on conductance variation; weight quantization tuning technique; approach to reduce the residual error	3	2	3

CIM Quantization: For all CIM

Year	Venue	Authors	Title	Tags	P	E	N
2018	CVPR	Google	Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference	integer-only inference arithmetic; quantizes both weights and activations as 8-bit integers, bias 32-bit; provides both quantized inference framework and training frame work
2023	ICCD	SJTU	PSQ: An Automatic Search Framework for Data-Free Quantization on PIM-based Architecture	post-training quantization framework without retraining; hardware-aware block reassembly
2025	arXiv	UHK	Binary Weight Multi-Bit Activation Quantization for Compute-in-Memory CNN Accelerators	a quantization framework that considers CIM's mixed-signal constraints; closed-form layer-specific weight binarization method; differentiable function for uniform multi-bit quantization	3	2	2

CIM: Digital CIM

Year	Venue	Authors	Title	Tags	P	E	N
2025	ISCAS	CAS	StreamDCIM: A Tile-based Streaming Digital CIM Accelerator with Mixed-stationary Cross-forwarding Dataflow for Multimodal Transformer	tile-based reconfigurable CIM macro microarchitecture; mixed-stationary cross-forwarding dataflow; ping-pong-like finegrained compute-rewriting pipeline

NVM

Year	Venue	Authors	Title	Tags	P	E	N
2018	Nat. Elec.	IBM	Mixed-precision in-memory computing	selectively-precise iterative refinement; PCM-based computational memory; drift-resilient computing	5	4	4
2020	GLSVLSI	UND	Benchmarking Computing-in-Memory for Design Space Exploration	uniform benchmarking of CiM designs based on different memory technologies	3	3	2
2024	ISCAS	UMCP	On-Chip Adaptation for Reducing Mismatch in Analog Non-Volatile Device Based Neural Networks	float-gate transistors based; hot-electron injection to address the issue of mismatch and variation
2023	DATE	UniBo	End-to-End DNN Inference on a Massively Parallel Analog In Memory Computing Architecture	many-core heterogeneous architecture; general-purpose system based on RISC-V cores and nvAIMC cores; based on Phase-Change Memory(PCM);
2025	DAC	ZJU	VQT-CiM: Accelerating Vector Quantization Enhanced Transformer with Ferroelectric Compute-in-Memory	static-dynamic VMM Conversion; look-up table accelerated RVQ(Residual Vector Quantization); product vector quantization (PVQ)	3	3	3

Prefetching

Challenge: Speculative prefetch requests can cause undesirable effects on the system (e.g., increased memory bandwidth consumption, cache pollution, memory access interference).

Year	Venue	Authors	Title	Tags	P	E	N
2021	MICRO	ETHZ	Pythia: A Customizable Hardware Prefetching Framework Using Online Reinforcement Learning	formulating prefetching as a reinforcement learning problem; holistic learning from multiple program features and system feedback; customizable prefetching objective via configuration registers	3	3	2
2025	MICRO	NUDT	Elevating Temporal Prefetching Through Instruction Correlation	critical instruction detection based on miss contribution; coverage-based classification for metadata utility; adaptive metadata cache partitioning via controller	3	4	4