WinterLabs
All labs
Silicon Lab

Chips, from a spec to signoff.

Silicon Lab is where Winter Labs builds hardware. We design certifiable RISC-V processor IP, neuro-inspired accelerators, and Xenon — an agentic silicon compiler that owns the fragmented open-source EDA toolchain and drives a design from a Markdown spec all the way to a place-and-routed, tape-out-ready layout.

RISC-V processor IPAgentic RTL-to-GDSII flowsOpen-source EDA (OpenLANE / OpenROAD)Functional safety & signoff
Silicon Lab logo
3
projects in this lab
130nm
tape-out-proven process targets
9
tracked milestones
Si
lab designation

3 projects underway

Xenon

In Development

The agentic silicon compiler.

Open-source EDA is now good enough to ship real silicon, but the toolchain is wide, fragmented and unforgiving — a single RISC-V core can need 15+ invocations across 8 tools, each with its own flags and log format. Xenon puts a team of Claude Code agents in charge of that toolchain: they read synthesis reports, parse timing violations, and edit RTL to fix them. A checkpointed step pipeline keeps every artifact reproducible, and an MCP server exposes the whole flow to agentic IDEs.

161
tests passing
10+
EDA tools orchestrated
RTL→GDS
flow

What makes it different

  • Agents read synthesis & timing reports, then edit RTL to close violations
  • Reproducible Docker sandbox — every run is checkpointed and diffable
  • MCP server exposes the flow to any agentic IDE
  • Targets tape-out-proven SkyWater 130 & GF 180 PDKs

Stack

VerilatorYosysOpenROADOpenLANEcocotbnextpnrGTKWave

Also in

Agentic EDARTL→GDSIIOpenLANEMCPClaude Code

AEGIS

In Development

Safety-certifiable RISC-V processor IP.

AEGIS (Adaptive Engine for Guarded Integrated Systems) is a three-domain SoC built for hard real-time control in safety-critical domains — airborne (DO-254 DAL-A) and automotive (ISO 26262 ASIL-D). A real-time control domain (RV32IMACF + Xdrone) runs at 240 MHz with TCLS 2oo3 lockstep voting; an application domain (RV64GCV, out-of-order) and a security domain (RV32E with PMP + root of trust) sit on the same safety interconnect. The design prioritises deterministic timing, fault detection, and certification traceability.

240 MHz
clock @ 130nm
12 cyc
IRQ latency
24/24
tests passing

What makes it different

  • TCLS 2oo3 lockstep voting with a dedicated safety monitor unit
  • 512 KB dual-bank scratchpad with SECDED ECC, 1-cycle read
  • 12-cycle guaranteed interrupt latency (49.9 ns @ 240 MHz)
  • Certification-traceable: DO-254 DAL-A / ISO 26262 ASIL-D

Stack

VerilogRISC-VSECDED ECCSkyWater 130
RISC-VRV32IMACFFunctional SafetyTCLS 2oo3Verilog

Azmuth

Research

A neuro-inspired RISC-V accelerator for edge AI.

Azmuth extends the standard RV32IMC instruction set with custom Xcew instructions for edge AI. An Expression-ML pipeline evaluates exp/ln/sub with a memoization cache and DAG-based common-subexpression elimination; a spiking-neural-network array uses time-to-first-spike encoding with STDP learning; and a ReRAM non-volatile-memory controller adds wear-leveling and ECC. Deterministic, constant-time policy execution guards against timing side-channels, while per-tile power orchestration handles sleep/wake and body-bias control.

RV32IMC
base ISA
Xcew
custom ISA
EML·SNN·NVM
AI blocks

What makes it different

  • Expression-ML pipeline with memoization + CSE via a DAG
  • Spiking neural network: LIF neurons, TTFS encoding, STDP learning
  • ReRAM NVM controller with wear-leveling and ECC
  • Constant-time policy execution + per-tile power orchestration

Stack

VerilogRISC-VReRAMSNN

Also in

RISC-VAI AcceleratorSNNEdge AIVerilog

Have a hard problem worth freezing time over?

Whether it's silicon, capital, or research at the edge of what's possible — we'd love to hear what you're building.