日本語
KEY TAKEAWAYS Chiplet thermal management depends on the complete heat path, not one material property. Select a substrate by comparing directional thermal conductivity, copper and via geometry, interface resistance, coefficient of thermal expansion, routing density, warpage, and cost. Then validate the package and PCB together with measured material data and a realistic power map.
A substrate can look excellent in a conductivity table and still fail in the package. Local heat flux, interfaces, copper discontinuities, and mechanical strain often control the result before bulk conductivity does.
This guide explains how substrate choice changes heat spreading in high-performance computing chiplets, how to screen four material platforms, and what engineers should verify before releasing a PCB or package design.
Why Substrate Material Changes the Chiplet Temperature Map
Substrate material changes the chiplet temperature map by controlling heat flow below and between dies while also setting the mechanical response to temperature gradients.
An HPC package can combine compute tiles, cache, input and output dies, high-bandwidth memory, and power devices. Each die has a different power density and temperature limit. A hot compute tile can warm neighboring memory even when both have adequate cooling in isolation, so chiplet thermal management must address local junction temperature and die-to-die coupling.
Thermal conductivity controls conduction through and along the substrate. Directional behavior matters because an organic build-up stack may spread heat through copper planes while conducting poorly through resin-rich dielectric. Coefficient of thermal expansion and elastic modulus determine how the same temperature field becomes warpage, bump strain, or delamination.
Rth = t divided by kA Temperature rise = heat flow multiplied by Rth
Here, t is thickness, k is thermal conductivity, A is effective area, and heat flow is the power crossing the layer. This equation screens concepts, but it omits spreading resistance, copper voids, neighboring sources, and contact resistance. Detailed chiplet thermal management still needs a compact network or three-dimensional simulation.
Where Heat Flows in an HPC Chiplet Package
Heat in a lid-cooled chiplet package follows a dominant upward path and a smaller downward path through the package substrate and PCB.
Die-to-lid path: Silicon, thermal interface material, the lid, and the heat sink or cold plate usually remove most package power. Die height variation and interface bond line thickness can outweigh a modest substrate upgrade.
Die to substrate path: Underfill, microbumps, redistribution layers, the interposer, and package substrate spread part of the heat toward cooler regions. This path matters near edge chiplets and localized hot spots.
Package to PCB path: Solder balls and land side copper move heat into motherboard planes and thermal vias. Use real convection or chassis contacts rather than treating the board as an unlimited heat sink.
Die-to-die path: A conductive interposer lowers lateral gradients but can carry more heat into a temperature sensitive die. Placement and workload influence chiplet thermal management as directly as material choice.
Build a power map before choosing a stack. A 120 W chiplet over 600 mm2 averages 20 W/cm2. If 60 W is concentrated in a 100 mm2 region, local heat flux reaches 60 W/cm2, which is the more useful simulation input.
How Organic Silicon Glass and Ceramic Substrates Compare
Organic, silicon, glass, and ceramic substrates solve different combinations of thermal, electrical, mechanical, and manufacturing constraints.
| Platform | Thermal Behavior | Best Reason to Select | Main Risk |
|---|---|---|---|
| Organic build-up | Dielectrics often measure about 0.3 to 1 W/mK. Copper and vias dominate spreading. | Flexible layers, mature PCB assembly, practical cost, strong power routing. | CTE mismatch with silicon, resin-rich choke points, copper-driven warpage. |
| Silicon interposer | Bulk silicon commonly measures about 120 to 150 W/mK. | Fine-pitch interconnect, dense routing, CTE alignment with silicon dies. | Cost, TSV constraints, heat coupling between adjacent chiplets. |
| Glass interposer | Bulk glass is roughly 1 W/mK, so metal carries much lateral heat. | Low electrical loss, dimensional stability, large format potential. | Weak intrinsic spreading, brittle handling, dependence on via metallurgy. |
| Aluminum nitride | Common grades are roughly 140 to 180 W/mK and electrically insulating. | Short heat paths for concentrated sources and harsh environments. | Brittleness, cost, routing limits, transition into organic boards. |
Organic substrates suit many large 2D and 2.5D packages when copper and vias act as thermal structures. Silicon fits very high interconnect density. Glass needs deliberate metal-based spreading. Aluminum nitride works where an insulating, high-conductivity path matters more than very fine routing.
What Determines the Right Substrate for Chiplet Thermal Management
The right substrate meets junction temperature, signal integrity, power integrity, reliability, and production yield together.
Local power density: List sustained power, transient peak, active area, maximum junction temperature, and allowable gradient for every die. Do not distribute total package power evenly without evidence.
Heat path ownership: State how much power the lid, cold plate, substrate, and PCB must remove. A board without airflow or chassis contact should not receive an optimistic load.
Directional data: Request in-plane and through-plane conductivity. Also request CTE in three axes below and above glass transition, moisture absorption, modulus, thickness tolerance, and test methods.
Electrical and thermal co-design: Continuous copper supports return current and heat spreading, while antipads, splits, and necks increase both electrical and thermal resistance.
Warpage and assembly: Copper balance, core thickness, lid stiffness, underfill modulus, reflow profile, and clamp load change coplanarity and interface thickness.
Manufacturing window: Via aspect ratio, lamination count, registration, plating, resin filling, and inspection access must remain repeatable beyond one prototype.
PCB Practices That Improve Heat Spreading
PCB heat spreading improves when copper and vias form continuous, manufacturable paths from package lands to a real thermal destination.
Continuous copper: Avoid unnecessary splits, dense antipads, and narrow necks below hot dies. Preserve return current while giving heat enough cross-sectional area.
Thermal vias with a destination: Connect arrays to broad planes, exposed metal, a chassis contact, or another cooled structure. Vias ending in an isolated island offer little benefit.
Balanced copper: Compare density across opposing layers and panel quadrants. Better balance reduces lamination distortion and assembly warpage.
Qualified filled vias: Filled and capped vias improve planar routing and thermal continuity, but stacked structures raise plating and reliability demands. Specify fill, cap, coupons, and acceptance criteria.
Sensitive die placement: Moving memory, analog, or photonic functions away from the main hot spot can outperform a package wide material change. Recheck link constraints before placement is frozen.
A fabricator can identify conflicts between thermal intent and process capability here. PCBgogo lists custom stack-up review, controlled impedance, blind and buried vias, resin-filled vias in pad, copper-filled microvias, back drilling, and high layer count PCB production. Sending the stack, copper distribution, via plan, and power map for DFM review allows these features to be checked before release.
How to Validate the Thermal Design
Validation must correlate simulation with a physical test vehicle at several power conditions.
Define boundaries: Record coolant or air inlet temperature, flow, heat sink resistance, clamp load, ambient condition, and board mounting.
Model interfaces: Include thermal interface material, underfill, microbumps, solder balls, solder mask, adhesive, oxide layers, and contact resistance.
Build measurement access: Use calibrated diode or resistance sensors where possible. Infrared work requires controlled emissivity, and thermocouples must not disturb the heat path.
Correlate several loads: Compare idle, a middle load, and worst-case power. A correction factor that changes with load often signals missing temperature dependence or changing contact.
Stress the assembly: Use thermal cycling, power cycling, storage, and post-test inspection as required. Track thermal resistance as well as electrical continuity.
What Recent Research Adds to Substrate Decisions
Recent research shows that local heat extraction and source proximity can matter more than a broad increase in substrate conductivity at extreme heat flux.
The 2023 IEEE Heterogeneous Integration Roadmap thermal chapter reports a 44.4 percent reduction in junction-to-inlet thermal resistance for backside microfluidic cooling of a processor dissipating up to 215 W under its reported flow-normalized comparison with a conventional cold plate. The lesson is to shorten the path from the dominant source to the coolant when a lid and substrate cannot spread the flux.
A 2024 IEEE International Conference on Electronic Packaging Technology paper modeled diamond on chip on interposer integration in a realistic 2.5D package. It reported more than a 20 percent reduction in maximum temperature at power density above 1.5 W/mm2. Diamond is specialized, but the result supports targeted spreaders near the highest flux region.
The 2021 IEEE Heterogeneous Integration Roadmap HPC chapter described water cold plates for package power in the 300 to 400 W range and discussed thermal vias, conformal lids, and heat shields. These are roadmap values, not universal limits. Cooling still depends on junction target, heat flux, pressure drop, contact resistance, and facility conditions.
Questions to Ask Before Fabrication Release
Release readiness depends on whether supplier and design team can answer the same thermal and manufacturing questions with traceable data.
Material evidence: Are conductivity, CTE, Tg, modulus, moisture absorption, and thickness tied to named test methods and production lots?
Via construction: Which filled, capped, stacked, or staggered structures are qualified for the exact dielectric thickness?
Copper distribution: Has density been reviewed by layer and panel location, and are balancing features electrically safe?
Inspection: Will cross sections, X-ray inspection, thermal coupons, impedance coupons, warpage data, and traceability be supplied where required?
Assembly: Are reflow, package warpage, coplanarity, void criteria, interface bond line, and clamp load defined?
Model handoff: Do final material lots and measured thicknesses return to the thermal, signal, power, and mechanical models?
Close any missing answers before volume release. Chiplet thermal management problems cost less to correct in stack review than after thermal excursion, intermittent bump failure, or warped assembly appears in qualification.
Frequently Asked Questions About Chiplet Thermal Management
Is a silicon interposer always cooler than an organic substrate?
No. Silicon has much higher bulk conductivity than organic dielectric, but temperature also depends on thickness, metal pattern, die placement, interfaces, and top side cooling. Silicon can also carry more heat into a neighboring sensitive die.
Can thermal vias replace a cold plate?
No. Thermal vias improve the downward path into package and PCB copper, but high-power compute tiles usually need efficient top side cooling. Use vias to complement the primary cooler.
How much heat should the PCB remove?
The answer comes from system boundaries. A board with large planes, airflow, or a chassis contact can remove useful heat, while an enclosed board may remove little. Model the actual mounting condition.
Why can a conductive material increase reliability risk?
Conductivity does not describe CTE, modulus, fracture behavior, or adhesion. A material can lower temperature yet raise bump strain, brittle fracture risk, or warpage.
When should testing begin?
Testing should begin with a simplified thermal vehicle before the package is frozen. Early vehicles verify interface resistance, copper and via assumptions, sensor placement, and model correlation while changes remain affordable.
Conclusion
Reliable chiplet thermal management starts with the real power map and complete heat path, then selects a substrate that balances spreading, routing, mechanical compatibility, manufacturability, and cost. Use supplier-specific data, preserve continuous copper and useful via paths, and correlate the package plus PCB model with a physical test vehicle. Send the stack and thermal constraints for DFM review as soon as the interconnect architecture is stable.
