High-Speed IO Architect - Digital - PCIe/UCIe/Ethernet Subsystems

• No endless meetings, no PowerPoint presentations to convince middle management., As the HSIO Architect, you will own the high-speed interface architecture across the various product families. You will drive decisions from SerDes architecture to UCIe die-to-die integration, working directly with the Head of Engineering, digital leads, and analog teams. You will also serve as the technical interface to key IP vendors (UCIe PHY, SerDes), foundry partners and anchor customers.

This is not an IP integration role. We need someone who can define the architecture, validate the tradeoffs, and guide execution through tape-out.

What You’ll Do

Architecture & Definition

Define high-speed IO architecture for PCIe/Ethernet-224G/448G scale-up fabric/CPO-optics to electrical interfaces

Define die-to-die interface architecture for UCIe integration: flit formats, credit-based flow control, sideband management, and latency targets

Architect the SerDes-to-photonics interface, retimer integration, and co-packaged optics (CPO) readiness

Own the signal integrity budget: TX/RX equalization, channel loss allocation, crosstalk margins, and jitter decomposition across the full channel

Implementation Oversight

Guide SerDes PHY selection and evaluation; assess vendor IP (custom vs. licensed), and establish integration requirements

Define and review UCIe PHY integration: bump map, power domain partitioning, analog/digital co-design requirements, and pad ring architecture

Collaborate with digital architecture lead on protocol bridge design - PCIe TLP * UCIe flit conversion, flow control, error handling

Drive DFT and compliance test architecture for PCIe certification and UCIe conformance testing

Establish PVT margin strategies and power management per-lane DVFS, shutdown sequences, and IVR chiplet coordination

Customer & Ecosystem Engagement

Serve as primary technical interface to anchor customers and engaging hyperscaler architecture teams (AWS, Google, Microsoft, Meta) to validate product definition against platform requirements

Roadmap & IP

Contribute to RTL reuse strategy - enabling derivative roadmap SKUs with minimal re-spin

Identify and scope patentable innovations in UCIe integration, adaptive equalization, and multi-protocol bridge architectures

If you’re reading this, you’re probably comfortable. You have a good job at a stable company with all the benefits., 15+ years in high-speed interface architecture; 5+ years at Principal level or higher in a fabless or IDM semiconductor environment

Deep expertise in PCIe architecture - from PHY layer through controller and protocol stack; direct tape-out experience strongly preferred

UCIe standard familiarity - flit-based transport, sideband protocol, and die-to-die integration in 2.5D/3D packages

Signal integrity expertise: S-parameter analysis, channel simulation (HSPICE/IBIS-AMI), eye diagram closure, and crosstalk budgeting

Experience with advanced nodes and advanced packaging (CoWoS, EMIB, InFO)

Track record of driving complex multi-party IP integrations from architecture through tape-out

Good to Have

Experience with 224G/400G+ SerDes architectures for scale-up AI fabric (NVLink, UALink, Ultra Ethernet, or proprietary)

Familiarity with electrical-optical co-design for CPO or near-package optics; EIC retimer or photonic interface experience

Prior work in chiplet-based multi-die architectures - bumping strategy, KGD handling, heterogeneous integration

PCIe-SIG membership or active contribution to UCIe Consortium working groups, Prior startup experience or comfort with early-stage ambiguity and fast-paced execution

The AI infrastructure market is exploding. Every hyperscaler, every cloud provider, every AI company is building custom silicon. But they all face the same problem: how do you connect hundreds of chips, deliver clean power at scale, and move terabits of data without melting the package?

That’s what we solve. TylSemi builds the chiplet infrastructure IP - the IO, power delivery, and interconnect building blocks - that makes AI/HPC systems actually work at scale.