연산-메모리 통합 : 시장 점유율 분석, 업계 동향 및 통계, 성장 예측(2026-2031년)

According to Mordor Intelligence, the compute-memory integration market size is projected to be USD 0.38 billion in 2025, USD 0.59 billion in 2026, and reach USD 3.45 billion by 2031, growing at a CAGR of 42.36% from 2026 to 2031.

This report is Segmented by Compute Type (Analog Compute-In-Memory Chips, and More), Memory Technology (SRAM-Based, DRAM/HBM-Based PIM and More), Deployment Environment (Edge Devices, Data Centers and Hyperscalers, and More), End User (Semiconductor Companies and Chip Designers, Cloud and Hyperscale Providers, and More), and Geography. The Market Forecasts are Provided in Terms of Value (USD)

Global Compute-Memory Integration Market Trends and Insights

Energy Efficiency Gains in Edge AI and On-Device Inferencing

The compute-memory integration market is being pulled forward by always-on AI use cases that need continuous operation at the sensor, laptop, and embedded device levels without the thermal budget of conventional accelerators. In these workloads, moving model weights between memory and compute remains the largest energy burden, and compute-in-memory reduces that burden by placing arithmetic closer to where weights are stored. A mixed-precision memristor-SRAM processor published in March 2025 reached 40.91 TFLOPS/W on ResNet-20 with less than 0.45% accuracy degradation, which showed that energy gains do not require a full tradeoff in usable model quality. EnCharge AI stated in February 2025 that its EN100 accelerator targets more than 200 TOPS at under 9W for client and edge inference, which fits the power range that mobile and fanless systems can actually absorb. The compute-memory integration market is therefore benefiting from a practical adoption path, because edge buyers often care more about sustained inference within fixed power envelopes than about peak benchmark throughput. This same pattern is encouraging hybrid designs, since many adopters want analog-class efficiency but still need digital support to keep accuracy stable across production conditions.

Memory Wall Pressure from Large Language Model Workloads

The compute-memory integration market is also advancing because large language model inference is now constrained more by memory movement than by raw arithmetic capacity. During autoregressive decoding, large models repeatedly fetch weights that are used once per token step, which keeps arithmetic intensity low and makes memory bandwidth the limiting factor in latency-sensitive serving. Frontiers in Science described the memory wall in 2025 as a cross-cutting hardware constraint for modern AI systems, which strengthened the architectural case for processing-in-memory beyond a narrow optimization argument. Samsung announced in February 2026 that it had begun commercial HBM4 shipment with 3.3TB/s bandwidth per stack and 40% better power efficiency than HBM3E, which shows how memory vendors are responding directly to this bottleneck. As a result, the compute-memory integration market is moving from a niche architecture discussion into a broader infrastructure decision around token cost, latency, and system energy. This shift also raises switching costs, because logic-enabled memory products are less interchangeable than earlier commodity memory generations.

Analog Precision Variability And Yield Risk In Compute-In-Memory Designs

The compute-memory integration market still faces a major technical limit in analog precision stability, especially when neural weights are stored as physical conductance states inside non-volatile cells. Process variation, temperature drift, endurance loss, and cell-to-cell spread can all introduce noise into multiply-accumulate operations, and those errors build across deeper neural networks. A 2025 review in npj Unconventional Computing described key analog in-memory error sources and noted that mitigation methods can consume a meaningful share of the energy savings that make the architecture attractive in the first place. An IEEE Transactions on Circuits and Systems I study in 2025 also found that precision limits become restrictive under operating conditions below the peak arithmetic throughput of a 28nm SRAM-based design, which shows that measured silicon behavior can differ sharply from idealized claims. The compute-memory integration market, therefore, continues to favor hybrid architectures in many commercial settings because they preserve part of the analog energy benefit while relying on digital stages to protect output quality. This issue also widens the gap between strong academic prototypes and repeatable high-volume manufacturing, which matters most when the target customer is automotive, defense, or enterprise infrastructure.

Other drivers and restraints analyzed in the detailed report include:

1. Rise of High Bandwidth Memory and Near-Memory Architectures in AI Servers
2. Custom Silicon Demand For Power-Constrained Autonomous And Industrial Systems
3. Immature EDA, Compiler, and Benchmark Ecosystem

For complete list of drivers and restraints, kindly check the Table Of Contents.

Segment Analysis

Analog compute-in-memory chips held 44.56% of the compute-memory integration market share in 2025, which reflected their earlier commercialization in edge inference and the depth of research around resistive memory-based matrix operations. The market reached this position because analog arrays can execute core operations inside memory cells and avoid much of the data movement that burdens conventional digital accelerators. That advantage is strongest in weight-stationary workloads where power savings matter more than absolute numerical determinism. The analog segment also benefited from the fact that many early adoption cases, such as embedded vision and anomaly detection, can tolerate carefully managed precision tradeoffs if the energy gain is large enough. In that sense, the compute-memory integration market first opened where battery life, thermal design, and sustained local inference mattered more than universal software portability.

Digital compute-in-memory chips continue to hold a complementary role because they offer deterministic behavior and better alignment with established software expectations, even if their energy gain is lower than analog alternatives. This makes digital approaches more suitable where traceability, validation, and numerical repeatability are required by safety or operational rules. Hybrid compute-in-memory chips are projected to grow at a 42.69% CAGR through 2031, and that trajectory shows how the compute-memory integration market is trying to balance analog efficiency with digital control. A mixed-precision processor published in Nature in 2025 demonstrated that partitioning neural layers across memristor, SRAM, and digital units can preserve accuracy while still delivering strong efficiency, which is the same design logic that many hybrid vendors are now following. IBM announced in June 2026 that its NanoStack sub-1nm architecture improved SRAM density by 40% versus 2nm nodes, which supports the longer-term case for denser hybrid macro integration at advanced nodes.

Flash, charge-based, and other technologies represented 42.38% of the compute-memory integration market size in 2025, which showed the value of process familiarity and broad manufacturing availability. This group benefits from the reuse of cell structures and fabrication knowledge that were originally built for high-density non-volatile storage and are now being adapted for in-array compute. The compute-memory industry has leaned on these options because they provide a practical bridge between new architecture goals and existing manufacturing habits. SRAM-based compute-in-memory remains important because it has the most transparent public silicon characterization for edge AI and gives developers a clearer view of accuracy, energy, and operating tradeoffs. Its limit is density, since SRAM takes more area per bit and is harder to scale for workloads that require larger on-chip weight storage.

DRAM and HBM-based processing-in-memory is projected to grow at a 42.99% CAGR through 2031, which makes it the fastest-moving memory technology path inside the market. Samsung said in February 2026 that its commercial HBM4 delivered 3.3TB/s per stack with 40% better power efficiency than HBM3E, which reinforces the appeal of memory-centric acceleration in AI infrastructure. SK hynix strengthened the same direction in June 2026 through its multi-year NVIDIA partnership, which linked future HBM roadmaps to AI supercomputers, consumer AI PCs, and robotics platforms. Micron added another signal in March 2026 with its 256GB LPDRAM SOCAMM2 launch, which targeted long-context LLM inference and showed that near-memory ideas are advancing across multiple DRAM categories at once. ReRAM, phase-change memory, and MRAM remain smaller in current revenue, but they retain strategic importance in applications where non-volatility, radiation tolerance, or specialized sensing and safety conditions justify a narrower but higher-value adoption path.

Complete Report Scope:

• By Compute Type
  • Analog Compute-in-Memory Chips
  • Digital Compute-in-Memory Chips
  • Hybrid Compute-in-Memory Chips
• By Memory Technology
  • SRAM-Based
  • DRAM/HBM-Based PIM
  • ReRAM-Based
  • Phase-Change Memory-Based
  • Magnetoresistive RAM-Based
• By Deployment Environment
  • Edge Devices
  • Data Centers and Hyperscalers
  • High-Performance Computing
  • Automotive and Transportation
  • Industrial Automation and IIoT
  • Consumer Electronics
  • Healthcare
  • Defense and Aerospace
• By End User
  • Semiconductor Companies and Chip Designers
  • Cloud and Hyperscale Providers
  • Server and AI Accelerator OEMs
  • Automotive OEMs and Tier-1 Suppliers
  • Industrial Automation Providers
  • Consumer Electronics OEMs
  • Other End Users (Government, Defense and Research Institutions)
• By Geography
  • North America
    • United States
    • Canada
    • Mexico
  • Europe
    • Germany
    • United Kingdom
    • France
    • Italy
    • Rest of Europe
  • Asia-Pacific
    • China
    • Japan
    • South Korea
    • India
    • Southeast Asia
    • Rest of Asia-Pacific
  • South America
  • Middle East and Africa

Geography Analysis

North America held 66.58% of the compute-memory integration market share in 2025, which reflected the region's concentration of LLM developers, hyperscale operators, and fabless semiconductor designers. The market remained strongest there because architecture development, software optimization, and early commercial buying power are closely linked in the same regional ecosystem. North America also benefits from the presence of vendors across the stack, from startup analog compute firms to large memory and logic companies with advanced packaging ambitions. IBM reinforced this ecosystem in June 2026 when it introduced its NanoStack sub-1nm architecture and highlighted its relevance for future SRAM density and energy efficiency improvements. The region is therefore likely to remain the reference market for early compute-memory product validation, especially where buyers need close interaction between chip design, software, and system integration.

Asia-Pacific is projected to grow at a 50.74% CAGR through 2031, which makes it the fastest-growing regional block in the compute-memory integration market. The region's strength comes from its leadership in advanced memory fabrication, HBM development, and the packaging capabilities that near-memory compute increasingly requires. Samsung's commercial HBM4 shipment and SK hynix's expanded NVIDIA collaboration both showed in 2026 that South Korea remains central to the future supply path for memory-centric AI hardware. China also demonstrated technical momentum through a 2025 Acta Physica Sinica paper on a 3D NAND compute-in-memory architecture for GPT-2-124M inference, which pointed to growing local capability in production-oriented system simulation.

Europe held a smaller share of the market in the current period, but its automotive and industrial demand profile gives it influence beyond its present revenue base. The strongest regional pull is likely to come from applications that need functional safety, long lifecycle support, and energy-efficient local inference, which keeps automotive electronics especially relevant. STMicroelectronics supported that direction with its Stellar P3E introduction for automotive edge intelligence, and Honda's collaboration with Mythic also signaled how vehicle programs can accelerate compute-memory qualification when power efficiency becomes central. South America and the Middle East and Africa remain early-stage areas for the market, and their near-term adoption is more likely to emerge through targeted sovereign AI, defense, and secure on-device inference programs than through broad commercial rollout.

1. Intel Corporation
2. Samsung Electronics Co., Ltd.
3. SK hynix Inc.
4. Micron Technology, Inc.
5. NVIDIA Corporation
6. Advanced Micro Devices, Inc.
7. Qualcomm Incorporated
8. IBM Corporation
9. Cerebras Systems, Inc.
10. Mythic, Inc.
11. EnCharge AI, Inc.
12. TetraMem Inc.
13. Rain AI, Inc.
14. Tenstorrent Inc.
15. NXP Semiconductors N.V.
16. Graphcore Limited
17. Marvell Technology, Inc.
18. Cadence Design Systems, Inc.
19. Synopsys, Inc.
20. Taiwan Semiconductor Manufacturing Company Limited

Additional Benefits:

• The market estimate (ME) sheet in Excel format
• 3 months of analyst support

TABLE OF CONTENTS

1 INTRODUCTION

• 1.1 Study Assumptions and Market Definition
• 1.2 Scope of the Study

2 RESEARCH METHODOLOGY

3 EXECUTIVE SUMMARY

4 MARKET LANDSCAPE

• 4.1 Market Overview
• 4.2 Market Drivers
  • 4.2.1 Energy Efficiency Gains in Edge AI and On-Device Inferencing
  • 4.2.2 Memory Wall Pressure From Large Language Model Workloads
  • 4.2.3 Rise of High Bandwidth Memory and Near-Memory Architectures in AI Servers
  • 4.2.4 Custom Silicon Demand for Power-Constrained Autonomous and Industrial Systems
  • 4.2.5 Software Co-Optimization for Quantized AI Inference at the Edge
  • 4.2.6 Compute Density Requirements in Data Centers Facing Power and Cooling Limits
• 4.3 Market Restraints
  • 4.3.1 Analog Precision Variability and Yield Risk in Compute-in-Memory Designs
  • 4.3.2 Immature EDA, Compiler, and Benchmark Ecosystem
  • 4.3.3 High Integration Cost and Long Qualification Cycles
  • 4.3.4 Limited Standardization Across Memory Types, Interfaces, and Programming Models
• 4.4 Supply Chain Analysis
• 4.5 Technological Outlook
• 4.6 Porter's Five Forces Analysis
  • 4.6.1 Threat of New Entrants
  • 4.6.2 Bargaining Power of Suppliers
  • 4.6.3 Bargaining Power of Buyers
  • 4.6.4 Threat of Substitutes
  • 4.6.5 Industry Rivalry
• 4.7 Capacity and Ecosystem Analysis

5 MARKET SIZE AND GROWTH FORECASTS (VALUE)

• 5.1 By Compute Type
  • 5.1.1 Analog Compute-in-Memory Chips
  • 5.1.2 Digital Compute-in-Memory Chips
  • 5.1.3 Hybrid Compute-in-Memory Chips
• 5.2 By Memory Technology
  • 5.2.1 SRAM-Based
  • 5.2.2 DRAM/HBM-Based PIM
  • 5.2.3 ReRAM-Based
  • 5.2.4 Phase-Change Memory-Based
  • 5.2.5 Magnetoresistive RAM-Based
• 5.3 By Deployment Environment
  • 5.3.1 Edge Devices
  • 5.3.2 Data Centers and Hyperscalers
  • 5.3.3 High-Performance Computing
  • 5.3.4 Automotive and Transportation
  • 5.3.5 Industrial Automation and IIoT
  • 5.3.6 Consumer Electronics
  • 5.3.7 Healthcare
  • 5.3.8 Defense and Aerospace
• 5.4 By End User
  • 5.4.1 Semiconductor Companies and Chip Designers
  • 5.4.2 Cloud and Hyperscale Providers
  • 5.4.3 Server and AI Accelerator OEMs
  • 5.4.4 Automotive OEMs and Tier-1 Suppliers
  • 5.4.5 Industrial Automation Providers
  • 5.4.6 Consumer Electronics OEMs
  • 5.4.7 Other End Users (Government, Defense and Research Institutions)
• 5.5 By Geography
  • 5.5.1 North America
    • 5.5.1.1 United States
    • 5.5.1.2 Canada
    • 5.5.1.3 Mexico
  • 5.5.2 Europe
    • 5.5.2.1 Germany
    • 5.5.2.2 United Kingdom
    • 5.5.2.3 France
    • 5.5.2.4 Italy
    • 5.5.2.5 Rest of Europe
  • 5.5.3 Asia-Pacific
    • 5.5.3.1 China
    • 5.5.3.2 Japan
    • 5.5.3.3 South Korea
    • 5.5.3.4 India
    • 5.5.3.5 Southeast Asia
    • 5.5.3.6 Rest of Asia-Pacific
  • 5.5.4 South America
  • 5.5.5 Middle East and Africa

6 COMPETITIVE LANDSCAPE

• 6.1 Market Concentration
• 6.2 Strategic Moves
• 6.3 Market Share Analysis
• 6.4 Company Profiles (includes Global Level Overview, Market Level Overview, Core Segments, Financials as available, Strategic Information, Market Rank/Share, Products and Services, Recent Developments)
  • 6.4.1 Intel Corporation
  • 6.4.2 Samsung Electronics Co., Ltd.
  • 6.4.3 SK hynix Inc.
  • 6.4.4 Micron Technology, Inc.
  • 6.4.5 NVIDIA Corporation
  • 6.4.6 Advanced Micro Devices, Inc.
  • 6.4.7 Qualcomm Incorporated
  • 6.4.8 IBM Corporation
  • 6.4.9 Cerebras Systems, Inc.
  • 6.4.10 Mythic, Inc.
  • 6.4.11 EnCharge AI, Inc.
  • 6.4.12 TetraMem Inc.
  • 6.4.13 Rain AI, Inc.
  • 6.4.14 Tenstorrent Inc.
  • 6.4.15 NXP Semiconductors N.V.
  • 6.4.16 Graphcore Limited
  • 6.4.17 Marvell Technology, Inc.
  • 6.4.18 Cadence Design Systems, Inc.
  • 6.4.19 Synopsys, Inc.
  • 6.4.20 Taiwan Semiconductor Manufacturing Company Limited

7 MARKET OPPORTUNITIES AND FUTURE OUTLOOK

• 7.1 White-space and Unmet-Need Assessment