chiplet interconnect

**Chiplet Interconnect Standards and Architecture** encompasses the **physical interface, protocol, and packaging technologies that enable multiple semiconductor dies (chiplets) to communicate within a single package** — with UCIe (Universal Chiplet Interconnect Express) emerging as the industry standard for die-to-die communication, defining electrical specifications, protocol layers, and packaging requirements to enable a plug-and-play chiplet ecosystem. **Why Chiplet Interconnects Matter:** The chiplet model disaggregates monolithic SoCs into smaller, specialized dies (compute, I/O, memory, accelerator) that are assembled in a package. This requires die-to-die (D2D) links that are: - **High bandwidth**: >1 TB/s aggregate for AI accelerators - **Low latency**: <2ns for cache-coherent communication - **Energy efficient**: <0.5 pJ/bit (100× better than off-package links) - **Standardized**: Enable mixing chiplets from different vendors/processes **UCIe (Universal Chiplet Interconnect Express):** UCIe 1.0 (2022) and UCIe 2.0 (2024) define a layered architecture: ```svg UCIe Stack:┌─────────────────────────────┐ Application Protocol Layer PCIe/CXL/custom streaming├─────────────────────────────┤ Die-to-Die Adapter Layer ARQ retry, CRC, link training├─────────────────────────────┤ Physical Layer Electrical signaling, clocking├─────────────────────────────┤ Packaging Layer Bump pitch, package type└─────────────────────────────┘ ``` **UCIe Physical Layer Options:** | Package Type | Bump Pitch | Data Rate | BW Density | Reach | |-------------|-----------|-----------|------------|-------| | Standard (organic) | 100-130μm | 4-32 GT/s | ~28 GB/s/mm | <10mm | | Advanced (Si interposer) | 25-55μm | 4-32 GT/s | ~165 GB/s/mm | <2mm | Advanced packaging with 25μm bump pitch provides ~6× the bandwidth density of standard packaging. **Protocol Options:** - **PCIe streaming**: For standard I/O communication (NIC chiplets, storage controllers) - **CXL**: For cache-coherent memory expansion and memory pooling chiplets - **Custom/Raw**: Proprietary protocols for vendor-specific high-bandwidth communication (e.g., AMD's Infinity Fabric, Intel's EMIB-connected tiles) **Existing Proprietary D2D Links:** | Interface | Company | BW/Link | Latency | Application | |-----------|---------|---------|---------|-------------| | Infinity Fabric | AMD | 600 GB/s | ~2ns | MI300X chiplet mesh | | EMIB | Intel | >100 GB/s | <5ns | Meteor Lake, Ponte Vecchio | | NVLink-C2C | NVIDIA | 900 GB/s | ~5ns | Grace-Hopper | | Lipincon | TSMC | 1.6 TB/s | <1ns | CoWoS chiplets | | BoW (Bunch of Wires) | OCP standard | Variable | ~3ns | Open standard | **Signal Integrity Challenges:** D2D links at 16-32 GT/s across microbumps face: **crosstalk** between closely spaced signals (~25μm pitch), **power supply noise** coupling through shared substrate, **impedance discontinuities** at bump transitions, and **thermal effects** on signal propagation. Solutions include: shielding ground lines between signal lanes, equalization (CTLE + limited DFE), and careful power distribution network design on the interposer. **Chiplet interconnect standardization through UCIe is the technical foundation enabling a heterogeneous chiplet ecosystem** — allowing the semiconductor industry to transition from monolithic SoC design to a modular, multi-vendor chiplet assembly paradigm where compute, memory, I/O, and accelerator dies from different companies and process nodes can be combined in a single package.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account