Skip to content

AiT Partner · UCIe & Chiplet Interconnect IP

UCIe is standardized.
The IP that carries it is not.

Two UCIe blocks on the same process node can behave nothing alike once real packages, thermals and workloads set in. InPsy guarantees the link that everything else now depends on.

COMPUTE DIE I/O DIE UCIe beachfront · sea of wires

The moment confidence gets tested

Early in the program, the interconnect looks like a checkbox

Teams pick an IP provider, agree on bandwidth, latency and power assumptions, and move on. Then the system reality evolves. Packaging constraints shift, thermal budgets tighten, routing complexity grows — and the question quietly changes from whether the design works to whether the link behaves predictably under real system pressure.

The conversation shifts from “Does it support UCIe?” to “How much confidence does this give us when the system gets complicated?”

01

Slice quality ≠ system quality

A component that passes standalone can still miss once it lands in your ASIC and package. The margin that survives real routing and thermal is what actually ships.

02

The validation falls on you

Component-only vendors hand off implementation, packaging, SI/PI and design verification to the system team — exactly when schedule margin has disappeared.

03

Interconnect becomes a risk variable

More chiplets, more heterogeneous tiles, more aggressive targets. The die-to-die fabric stops being plumbing and becomes a first-order, system-level risk.

Where InPsy changes the outcome

A full ownership guarantee, not a component drop

InPsy takes end-to-end ownership of device-to-device quality — guaranteeing that your D2D link operates as proposed inside your ASIC and your package, not just on a datasheet.

END-TO-END D2D GUARANTEE

The whole block, not the slice

Implementation of multiple UCIe blocks plus PLL, clock tree and glue logic — delivered with IP- and package-level SI/PI reports and full design-verification. Where other vendors hand off the hard integration, InPsy owns it.

SILICON-PROVEN TRACK RECORD

First, repeatedly

First to silicon-proven UCIe at 16G (2022) and 32G (2023), plus 64G, 3D face-to-face SoIC and SoW-X. First to mass production on a Tier-1 US flagship AI accelerator, with 15+ UCIe programs projected to reach MP through 2029.

PERFORMANCE HEADROOM
1E-23

Bit error rate reachable in silicon — outperforming the UCIe 3.0 standard. Energy efficiency under 0.2 pJ/bit for UCIe-A, beating the specification by 50%+.

CUSTOM, NOT CATALOG

Tailored to your product

Bandwidth, beachfront-vs-depth-vs-reach, power budget, metal stack and package type tuned to your design — with an embedded MCU enabling firmware-based revisions after tape-out.

Fixed Tx / Rx per lane

The conventional approach: each lane is dedicated to one direction, decided at design time. Predictable, but it spends beachfront and power to reach a bandwidth target.

InPsy SBD — every lane transmits and receives at once

Simultaneous Bi-Directional architecture lets the clock tree and PLL run at half the frequency for the same throughput — three breakthroughs in one macro:

+100%bandwidth across the same beachfront
−50%beachfront for the same bandwidth
−40–50%power — 2× bandwidth at only 1.5× power

The partner

InPsytech — the Egis Group silicon-IP subsidiary

A high-speed interconnect specialist pioneering high-complexity silicon IP on the most advanced process nodes, with value-added integration and SI/PI services that carry customers from evaluation to tape-out.

Core technology
UCIe PHY + Controller, high-speed D2D & data-transfer IP
Process reach
2 / 3 / 4 / 5 / 6 / 7 / 12 / 22nm across leading foundries
Foundries
TSMC · Samsung · Intel · UMC
Team
~153 engineers · Hsinchu, Taiwan

Ecosystem: ARM Total Design · Samsung SAFE™ IP Alliance · Intel Foundry Accelerator IP Alliance  ·  recognized for outstanding partnership by a leading AI-accelerator program.

The AiT play

Where InPsy sits in the EDA 3.0 lifecycle

AiT’s thesis is that as silicon disaggregates, the bottleneck stops being any single tool and becomes orchestration across the lifecycle — from intent to yield. InPsy anchors the architecture and physical-interconnect layers, where die-to-die decisions get locked in.

L1
Intent & Requirements
L2
Architecture & Modeling — chiplet partitioning, D2D bandwidth & reach tradeoffsInPsy
L3
RTL, Firmware & Physical IP — UCIe PHY + Controller, SI/PI, packageInPsy
L4
Verification
L5
Yield & Physical
L6
AI Orchestration

Read the full EDA 3.0 thesis →

The link everything depends on shouldn’t be the thing you find out about at tape-out

If your team is working through UCIe partitioning or evaluating interconnect IP, that’s the conversation. Thirty minutes, no slides required.

Book a die-to-die briefing
AiT represents InPsytech across North America.  ·  The Watchtower Brief  ·  inpsy.com  ·  © 2026 AI Tech Sales