Book a demo

Cut patent&paper research from weeks to hours with PatSnap Eureka AI!

Try now

Chiplet vs. monolithic design: yield, thermal tradeoffs

Chiplet Architecture vs. Monolithic Design — PatSnap Insights
Semiconductor Technology

Chiplet architecture has emerged as the dominant response to slowing Moore’s Law scaling, soaring mask costs, and yield constraints on large monolithic dies. But disaggregating a system-on-chip into multiple smaller dies introduces interconnect overheads, thermal complexity, and toolchain gaps that monolithic designs avoid entirely — and the patent record shows exactly where the competitive battle lines are drawn.

PatSnap Insights Team Innovation Intelligence Analysts 14 min read
Share
Reviewed by the PatSnap Insights editorial team ·

Why the Monolithic Model Is Running Out of Road

The monolithic system-on-chip — all functional blocks on a single semiconductor die — is constrained by three converging pressures: the reticle size limit imposed by photolithography, yield falloff as die area grows, and the economic impossibility of mixing process nodes on a single substrate. As process nodes advance, the 2020 literature record captured the core diagnosis plainly: “as process nodes move forward, dramatically rising cost, design cycle, and complexity are driving industry to focus on chiplets.”

858mm²
Maximum reticle-limit die area — TSMC’s monolithic-beyond-reticle patent targets chips exceeding this
20+
Intel chiplet patent records filed across US, WO, and IN jurisdictions from 2018–2026
2018–2026
Patent filing span analysed in this landscape — covering foundational to emerging chiplet innovations
2022
Year a formal quantitative cost model confirmed chiplet yield and cost advantages over monolithic design

The fault-tolerant monolithic architecture literature identified this constraint as early as 2000, noting explicitly that “lithographic errors have set limits on the complexity of circuitry that can be fabricated in one piece without fatal flaws.” The same yield argument — defect probability scaling with die area — now drives chiplet adoption at leading-edge nodes where mask costs alone make repeated respins of large monolithic dies economically untenable.

Chiplet architecture responds by fabricating functional blocks — compute cores, memory, I/O, accelerators — as discrete dies at the process node optimal for each function, then assembling them using advanced packaging: 2.5D silicon interposers, 3D stacking, and fan-out packaging. The assembly approximates the electrical behaviour of a monolithic device while retaining the yield and node-mixing advantages of disaggregation. IBM’s 1996 patent on minimising latency between processors and shared memory established the architectural motivation for disaggregation decades before chiplet terminology existed — making clear that the tension between integration density and access latency is not a new problem, only a newly urgent one.

What is chiplet architecture?

Chiplet architecture is the disaggregation of a monolithic system-on-chip (SoC) into multiple separately fabricated dies assembled into a common package. Each die (chiplet) can be manufactured at the process node optimal for its function — leading-edge logic for compute cores, mature node for I/O — and interconnected via advanced packaging technologies including 2.5D silicon interposers, 3D stacking, and fan-out packaging.

TSMC’s response — rather than abandoning monolithic integration — has been to extend it past the photolithography reticle size limit. The company’s 2023 US patent describes a single semiconductor substrate with chiplets connected by die-to-die conductive patterns bridging die-to-die spaces, targeting chip sizes greater than 858 mm² (26×33 mm) — a direct competitive answer from the monolithic side that blurs the line between the two approaches.

Figure 1 — Chiplet Architecture Patent Filing Timeline: Key Milestones 2018–2026
Chiplet architecture patent filing timeline showing key milestones from Intel, AMD, TSMC, Alibaba, Microsoft and Bosch between 2018 and 2026 2018 2020 2022 2023 2024 2025–26 Intel SoC Disaggregation Micron / Cost Model Lit. AMD / TSMC / Microsoft Alibaba 3D / Nanjing Auto Bosch / Intel per-chiplet PnP
Filing activity clusters into three phases: 2018–2020 (commercial foundations), 2021–2023 (maturation with AMD, TSMC, Microsoft entries), and 2024–2026 (automotive, per-chiplet power targeting, and fault tolerance).

Yield and Cost: Where Chiplets Win on the Numbers

Chiplet disaggregation reduces total system cost primarily through yield improvement: by replacing a single large die with multiple smaller dies, defect probability is distributed across individually testable units rather than concentrated in one high-area substrate. A 2022 quantitative cost model confirmed that “multi-chip integration reduces total cost of the VLSI system through yield improvement, chiplet and package reuse, and heterogeneity.” The cost benefit is design-dependent — largest for complex, high-transistor-count designs where monolithic die area would otherwise suppress yield to economically unviable levels.

A 2022 quantitative cost model confirmed that multi-chip chiplet integration reduces total VLSI system cost through yield improvement, chiplet and package reuse, and heterogeneous process-node mixing — with the benefit largest for complex designs where large monolithic die area would suppress yield to unviable levels.

Intel’s late-bind SKU architecture makes the yield argument operational. Rather than committing to a fixed die configuration at wafer fabrication, Intel’s patents describe chiplets tested for functional unit count and power metrics, then grouped into normalised “chunks” with pre-determined collective values — allowing base dies to support multiple product SKUs by swapping interchangeable logic and memory chiplets during assembly. The 2022 Intel patent on enabling product SKUs based on chiplet configurations adds independent power gating per chiplet: dynamic power-up and power-down per workload, a capability absent in monolithic designs where functional blocks share a power domain.

“As process nodes move forward, dramatically rising cost, design cycle, and complexity are driving industry to focus on chiplets.”

The reuse argument extends beyond silicon: chiplet modularity allows a single validated chiplet design to be deployed across multiple product families without respin, amortising non-recurring engineering (NRE) costs across a broader revenue base. This is a structural advantage that monolithic SoC design cannot replicate — each new product requires a full die respun at process cost.

Figure 2 — Chiplet vs. Monolithic Design: Key Tradeoff Dimensions
Comparative assessment of chiplet architecture versus monolithic design across yield, interconnect bandwidth, thermal complexity, toolchain maturity, process flexibility and cost at scale Low Mid High Yield High Low On-die BW Low High Thermal Cmplx High Low EDA Maturity Low High Node Flexibility High None Cost Advantage High Low Chiplet Architecture Monolithic Design
Chiplet architecture leads on yield, process-node flexibility, and cost at scale; monolithic design retains advantages in on-die bandwidth, EDA toolchain maturity, and thermal simplicity — based on patent evidence and literature from 2018–2026.

Map chiplet and monolithic patent filings across Intel, AMD, TSMC, and emerging assignees in PatSnap Eureka.

Explore Full Patent Data in PatSnap Eureka →

Interconnect Performance: The Central Tradeoff of Chiplet vs. Monolithic Design

Interconnect architecture is the decisive point of difference between chiplet and monolithic designs. Monolithic integration offers effectively unlimited on-die bandwidth between functional blocks with minimal latency; chiplet designs impose inter-die bandwidth constraints, signal latency, and power penalties that are absent when logic is co-located on a single substrate.

AMD’s 2023 chiplet processor patent explicitly quantifies that task transfer between chiplets to manage thermal thresholds incurs significant overhead — including task transfer delay and communication system bandwidth consumption — a direct latency cost of chiplet disaggregation that monolithic designs with on-die task migration across cores do not face.

The interconnect bottleneck is not static — it is the most active engineering battleground in the 2024–2026 patent record. Onesta IP’s 2026 US patent on GPU chiplets using high-bandwidth crosslinks claims near-linear performance scaling with chiplet count via passive crosslink interconnects providing coherent L3 memory access — directly addressing what has historically been the primary argument for monolithic GPU designs. The patent explicitly states “the performance of a processing system generally scales linearly based on the number of GPU chiplets utilized by nature of physical duplication.”

Packaging technology mediates the interconnect gap. According to IEEE research on millimetre-wave chip-to-chip interconnections, 2.5D and 3D stacking approaches can achieve significantly higher inter-chiplet bandwidth than off-package connections, bringing chiplet-to-chiplet communication closer to on-die performance for specific link configurations. The 2022 literature on using chiplet encapsulation for processing-in-memory functions notes that co-packaging compute and memory chiplets “alleviates the ‘storage wall’ problem” by placing memory bandwidth closer to compute — a bandwidth benefit unavailable in monolithic designs that cannot mix DRAM and logic on a single die.

Key finding: The interconnect standard layer is the competitive battleground

Intel and Onesta IP are actively patenting high-bandwidth crosslinks and UCIe/PCIe-based chiplet interconnects to close the bandwidth gap with monolithic designs. IP strategists should monitor the interconnect standard layer — UCIe, PCIe, and proprietary crosslinks — as the primary competitive differentiator in next-generation processor architectures.

Cache coherency across chiplet boundaries is emerging as a distinct infrastructure problem. Wisconsin Alumni Research Foundation’s 2025 filing explicitly identifies the issue: chiplet assemblies lack the on-die coherency fabric that monolithic designs take for granted, requiring dedicated architectural solutions. Nanjing University’s 2025 heterogeneous chiplet compilation system using a dual-layer genetic algorithm addresses the related software gap — monolithic SoC design benefits from decades of mature EDA toolchains, while chiplet heterogeneity introduces compiler complexity that is only now beginning to be systematically solved.

Wisconsin Alumni Research Foundation’s 2025 US patent identifies cache coherency across chiplet boundaries as a problem requiring dedicated architectural solutions, noting that chiplet assemblies lack the on-die coherency fabric that monolithic processor designs take for granted.

Thermal and Power Management: Chiplet’s Hidden Engineering Cost

Thermal management in chiplet systems requires dedicated engineering investment that monolithic designs do not incur. Where a monolithic die presents a single, uniform thermal plane, a disaggregated multi-chiplet system generates localised hotspots at each die — and each chiplet’s power state can vary independently, creating dynamic thermal profiles that a single integrated thermal controller cannot handle efficiently.

Intel recognised this cost early: the company’s 2019 US patent on per-chiplet thermal control in a disaggregated multi-chiplet system — filed in the same window as its first SoC disaggregation patents — describes a system thermal engine that monitors each chiplet independently and reduces power to a single overheating chiplet without throttling the others. This capability, refined in the 2021 continuation, has no equivalent requirement in monolithic design flows. Micron Technology’s 2022 WO patent on dynamic power and thermal loading in a chiplet-based system explicitly contrasts the chiplet-based power management approach with the single-die SoC approach, underscoring the added system complexity.

Intel’s 2025 US patent on SOC disaggregation goes further, explicitly claiming differentiated per-chiplet power and performance targets: “the first logic chiplet has a first power and performance target and the second logic chiplet has a second power and performance target that is different from the first.” This fine-grained per-chiplet policy is impossible in monolithic designs where functional blocks share a power domain, but it introduces a management layer — hardware and firmware — that adds to design and validation cost. According to standards work tracked by JEDEC, power and signal integrity interfaces for multi-die packages remain an active standardisation area, further illustrating the engineering overhead chiplet teams must budget for.

Figure 3 — Per-Chiplet Thermal Management Architecture (Intel Patent Family)
Per-chiplet thermal management process showing how Intel’s system thermal engine monitors, isolates and throttles individual chiplets without affecting others in a disaggregated multi-chiplet processor system System Thermal Engine Monitors all chiplets Detect Hotspot Chiplet Per-chiplet thermal read Isolate Overheating Chiplet Target only Reduce Power Single Chiplet Others unaffected System Performance Maintained Fine-grained control
Intel’s per-chiplet thermal control system (2019, 2021 US patents) enables targeted power reduction to a single overheating chiplet without throttling others — a capability with no monolithic equivalent, but one that requires dedicated thermal management engineering not present in monolithic design flows.

The self-organising fault-tolerant chiplet approach from Zhong Cheng Hua Long Computer Technology (2023, CN) takes the power management argument further: a competitive scoring algorithm enables surviving chiplets to dynamically absorb the workload of a failed chiplet. In a monolithic design, a single fatal defect kills the entire die. In a well-architected chiplet system, fault isolation can be designed at the system level — converting a catastrophic failure mode into a graceful degradation. This architectural advantage comes at the engineering cost of designing, validating, and certifying the fault-tolerance logic itself.

Analyse chiplet thermal management and power architecture patents across Intel, Micron, and AMD with PatSnap Eureka.

Analyse Patents in PatSnap Eureka →

Application Domains: From AI Accelerators to Quantum Processors

Chiplet architecture is being adopted across a widening range of application domains, each with a different primary motivation for disaggregation from the monolithic baseline.

High-Performance Computing and AI Acceleration

The largest application cluster in the 2018–2026 patent record is HPC and AI workloads, where data movement between processor and memory — the “storage wall” — is the primary bottleneck. Chiplet-based processing-in-memory architectures co-package compute and memory chiplets to directly address this: “Chiplet combines processor cores and memory chips with advanced packaging technologies, such as 2.5D, 3D, and fan-out packaging… alleviates the ‘storage wall’ problem.” Alibaba’s 3D stacked processing system patents (2022 WO, 2023 US) target “high memory capacity, high memory bandwidth, low power consumption and small form factor” — requirements that monolithic SoC designs cannot simultaneously optimise at leading-edge nodes.

GPU and Parallel Compute Scaling

Intel’s largest filing cluster targets GPU disaggregation. The 2023 US modular GPU architecture patent describes a graphics processor with an active base die fabric and chiplet-based graphics processing resource blocks. Onesta IP’s 2026 patent claims that performance “scales linearly based on the number of GPU chiplets utilized by nature of physical duplication” — addressing the traditional monolithic argument that GPU performance depends on uniform on-die memory access. According to WIPO‘s global technology trends tracking, GPU-adjacent semiconductor packaging has been among the fastest-growing patent filing domains over the 2020–2025 period.

Automotive and Safety-Critical Systems

Robert Bosch’s 2026 DE patent integrating AI hardware accelerator, CPU, internal GPU, and working memory as distinct chiplets for automotive technical systems marks the expansion of chiplet architecture into safety-critical industrial markets. A 2024 CN filing from Nanjing University of Information Science and Technology targets autonomous driving applications, implementing a redundant dual-compute-system architecture with TSV interconnects for fault tolerance — explicitly addressing the high-safety requirements of autonomous driving systems, where monolithic designs are limited by their inability to isolate faults at the functional block level.

Quantum Computing

Chiplet scalability is being applied to quantum processor design for the same reason it was applied to classical processors: yield constraints from device variation. The 2022 literature on scaling superconducting quantum computers with chiplet architectures found that “scaling QCs by building ever larger individual chips with more qubits is hampered by device variation” due to frequency collisions in transmon qubits — the same yield-driven motivation as classical chiplets. The chiplet model offers a path to scaling qubit counts without requiring defect-free large substrates, a constraint that is even more demanding in quantum fabrication than in classical CMOS. Research published by institutions including Nature has documented the quantum coherence challenges that make large monolithic quantum chips difficult to manufacture reliably.

Flexible Hardware Acceleration

Chiplet-based hardware accelerators directly address the rigidity of monolithic GPU, DSP, and ASIC designs. Zhong Cheng Hua Long’s 2023 CN patent notes that “traditional single-type hardware accelerators cannot meet existing computing needs” and positions chiplet technology as a new hardware accelerator architecture enabling dynamic task scheduling across heterogeneous accelerator chiplets — resource utilisation improvements that fixed-allocation monolithic designs cannot achieve.

Patent Landscape and Strategic Implications for IP Teams

The patent landscape for chiplet architecture vs. monolithic design is concentrated, geographically expanding, and increasingly contested — with implications for freedom-to-operate analysis that engineering and IP teams cannot afford to treat as secondary.

Intel Corporation holds more than 20 distinct chiplet patent records across US, WO, and IN jurisdictions filed between 2018 and 2026, covering SoC disaggregation, late-bind SKU configurability, thermal management, modular GPU architecture, and memory chiplet packaging — the largest single-assignee chiplet IP position in the 2018–2026 patent dataset.

Intel’s dominance in chiplet disaggregation IP across US, WO, and IN jurisdictions represents a platform-level IP position rather than a point solution. Any organisation designing chiplet-based processors — particularly GPU or parallel processor architectures — must conduct freedom-to-operate analysis against Intel’s disaggregation patent family. The late-bind SKU architecture, per-chiplet thermal control, and modular GPU patents each represent independent claim structures that could affect downstream product designs.

AMD’s focused cluster on chiplet-level performance information and task scheduling (2023 US, 2023 WO, 2024 IN) presents a narrower but strategically important IP position at the performance management layer. TSMC’s monolithic-beyond-reticle patents (2023, 2025 US) represent a distinct defensive position — a middle path that may be more accessible to organisations seeking to avoid Intel’s disaggregation claims while still exceeding reticle-limit die sizes.

Chinese assignees — including Huawei, Nanjing University, Zhong Cheng Hua Long, and Wuxi Advanced Technology Research Institute — are building a parallel domestic IP position. The 2023 CN filing on PCIe-based chiplet interconnect explicitly notes that “under current domestic process conditions, both [monolithic approaches] are difficult to achieve,” signalling that domestic process node constraints are a direct driver of Chinese chiplet IP development. This creates a parallel IP ecosystem that organisations operating in Chinese markets must assess independently of the US-dominated landscape. The China National Intellectual Property Administration has documented rising semiconductor patent filings volumes consistent with this trend.

Figure 4 — Patent Filing Concentration by Assignee: Chiplet Architecture 2018–2026
Relative chiplet patent filing volumes by assignee from 2018 to 2026, showing Intel Corporation as the dominant filer with over 20 records, followed by AMD, TSMC, Microsoft, Alibaba, Chinese assignees collectively, and Bosch 5 10 15 20+ Intel 20+ AMD 6 TSMC 4 Microsoft 3 Alibaba 2 CN cluster 5+ Approximate patent record count (2018–2026 dataset)
Intel’s filing volume in chiplet architecture IP is dominant across jurisdictions; Chinese assignees collectively represent a growing parallel domestic IP ecosystem driven partly by process node access constraints. Figures are approximate counts from the retrieved patent dataset.

Three strategic implications follow from this landscape. First, the EDA and compiler toolchain gap is an underappreciated chiplet tradeoff: monolithic SoC design benefits from decades of mature EDA tools, while Nanjing University’s 2025 compiler patent and Wisconsin Alumni Research Foundation’s 2025 cache coherency patent signal that chiplet-specific toolchain infrastructure remains an open problem requiring long-lead investment. Second, engineering teams should model break-even die area before committing to disaggregation — the cost model advantage is real but design-dependent. Third, organisations entering chiplet architecture for the first time should treat per-chiplet thermal engineering and cache coherency infrastructure as first-class design requirements, not afterthoughts, based on the depth of patent activity in both domains. Links to PatSnap’s own patent analytics platform and technology landscape analysis tools provide structured methods for conducting this type of freedom-to-operate and whitespace analysis.

Frequently asked questions

Chiplet architecture vs. monolithic design — key questions answered

The central tradeoff is yield and cost versus interconnect performance. Chiplet architecture improves yield by replacing a single large die with multiple smaller dies, each fabricated at the process node optimal for its function. However, inter-chiplet communication introduces latency, bandwidth constraints, and power overheads that are absent in monolithic designs where all functional blocks share on-die interconnects with effectively unlimited bandwidth between them.

Defect probability scales with die area. A large monolithic die accumulates more defects and thus has lower yield than multiple smaller dies. A 2022 quantitative cost model confirmed that multi-chip integration reduces total VLSI system cost through yield improvement, chiplet and package reuse, and heterogeneity. The benefit is largest for complex, high-transistor-count designs where monolithic die area would otherwise suppress yield to economically unviable levels.

Chiplet interconnect approaches visible in the 2018–2026 patent record include UCIe (Universal Chiplet Interconnect Express), PCIe-based interfaces, proprietary high-bandwidth crosslinks, and advanced packaging substrates such as 2.5D silicon interposers, 3D stacking, and fan-out packaging. Onesta IP’s 2026 US patent claims near-linear performance scaling with chiplet count via passive crosslink interconnects providing coherent L3 memory access — directly addressing the interconnect bandwidth gap with monolithic GPU designs.

Monolithic designs benefit from uniform thermal management across a single die. Chiplet systems introduce localised hotspots requiring per-chiplet thermal policies. Intel holds multiple active patents on per-chiplet thermal control (2019, 2021, US) describing a system thermal engine that monitors each chiplet independently and can reduce power to a single overheating chiplet without throttling others. This capability requires dedicated hardware and firmware engineering that monolithic design flows do not need to budget for.

Based on patent filings spanning 2018–2026, Intel Corporation dominates by filing volume with more than 20 distinct patent records across US, WO, and IN jurisdictions covering SoC disaggregation, thermal management, modular GPU, and memory chiplet packaging. AMD, TSMC, Microsoft, Alibaba, and a growing cluster of Chinese assignees including Huawei, Nanjing University, and Zhong Cheng Hua Long Computer Technology also hold significant chiplet-related IP.

Yes. A 2022 paper on scaling superconducting quantum computers with chiplet architectures found that scaling quantum computers by building ever-larger individual chips is hampered by device variation due to frequency collisions in transmon qubits — the same yield-driven motivation as classical chiplets. The chiplet scalability model is being applied to quantum processor design as a path to scaling qubit counts without requiring defect-free large substrates.

Still have questions about chiplet vs. monolithic processor architecture? Let PatSnap Eureka answer them for you.

Ask PatSnap Eureka for a Deeper Answer →

References

  1. Chiplet Heterogeneous Integration Technology—Status and Challenges (Literature, 2020)
  2. A quantitative cost model and multi-chiplet architecture exploration (Literature, 2022)
  3. Using Chiplet Encapsulation Technology to Achieve Processing-in-Memory Functions (Literature, 2022)
  4. Scaling Superconducting Quantum Computers with Chiplet Architectures (Literature, 2022)
  5. Architecture of Computing System based on Chiplet (Literature, 2022)
  6. The Advances, Challenges and Future Possibilities of Millimeter-Wave Chip-to-Chip Interconnections for Multi-Chip Systems (Literature, 2018)
  7. Disaggregation of system-on-chip (SOC) architecture — Intel Corporation (US, 2025)
  8. Disaggregation of system-on-chip (SOC) architecture — Intel Corporation (US, 2024)
  9. Chiplet architecture for late bind SKU fungibility — Intel Corporation (US, 2023)
  10. Chiplet architecture chunking for uniformity across multiple chiplet configurations — Intel Corporation (US, 2023)
  11. Enabling product SKUs based on chiplet configurations — Intel Corporation (US, 2022)
  12. Per chiplet thermal control in a disaggregated multi-chiplet system — Intel Corporation (US, 2021)
  13. Per chiplet thermal control in a disaggregated multi-chiplet system — Intel Corporation (US, 2019)
  14. Dynamic power and thermal loading in a chiplet-based system — Micron Technology, Inc. (WO, 2022)
  15. Chiplet-Level Performance Information for Configuring Chiplets in a Processor — Advanced Micro Devices, Inc. (US, 2023)
  16. Method for low-cost, high-bandwidth monolithic system integration beyond reticle limit — TSMC (US, 2023)
  17. Method for low-cost, high-bandwidth monolithic system integration beyond reticle limit — TSMC (US, 2025)
  18. Three-dimensional stacked processing systems — Alibaba Group Holding Limited (US, 2023)
  19. Homogeneous chiplets configurable as a two-dimensional system or a three-dimensional system — Microsoft Technology Licensing, LLC (US, 2023)
  20. Chiplet architecture for a technical system — Robert Bosch GmbH (DE, 2026)
  21. GPU chiplets using high bandwidth crosslinks — Onesta IP, LLC (US, 2026)
  22. Cache synchronization for chiplet accelerators — Wisconsin Alumni Research Foundation (US, 2025)
  23. Heterogeneous Chiplet Compilation System Based on Dual-Layer Genetic Algorithm — Nanjing University (CN, 2025)
  24. A method and chip for implementing self-organizing chiplets — Zhong Cheng Hua Long Computer Technology Co., Ltd. (CN, 2023)
  25. Chip architecture and related methods — Huawei Technologies Co., Ltd. (CN, 2025)
  26. Chiplet architecture-based chip and control method — Nanjing University of Information Science and Technology (CN, 2024)
  27. Modular GPU architecture for clients and servers — Intel Corporation (US, 2023)
  28. WIPO — World Intellectual Property Organization: Global Technology Trends
  29. IEEE — Institute of Electrical and Electronics Engineers: Chip-to-Chip Interconnect Research
  30. JEDEC — Solid State Technology Association: Multi-Die Package Standards
  31. Nature — Quantum Computing and Semiconductor Research Publications
  32. China National Intellectual Property Administration (CNIPA)

All data and statistics in this article are sourced from the references above and from PatSnap‘s proprietary innovation intelligence platform. This landscape is derived from a targeted set of patent and literature records spanning 2018–2026 and represents a snapshot of innovation signals within that dataset only.

Your Agentic AI Partner
for Smarter Innovation

PatSnap fuses the world’s largest proprietary innovation dataset with cutting-edge AI to
supercharge R&D, IP strategy, materials science, and drug discovery.

Book a demo