Silicon 02: IO chiplet

Lowest-power
PCIe & CXL
silicon.

PCIe Gen 6, CXL 3.2, SerDes (partner), and the UCIe D2D interface (partner): the complete IO subsystem for your chiplet. PCI-SIG certified. Designed without legacy overhead, at 1 GHz.

Low PowerMin Latency · Highest ThroughputHigh PerformanceOptimal AreaGround-up design: no legacy overheadPCI-SIG certified at 32 GT/sPPA-optimized
PCIe Gen 6CXL 3.2SerDes / PHY (partner)RC/EP dual modex1–x16UCIe D2D (partner)
PCI-SIG certified (32 GT/s, non-flit mode): Sept 2025
IO SILICONPCIe Gen 6 Controller32 GT/s · non-flit mode · RC and EP · x1–x16ATCATSDTIPCI-SIG ✓ certCXL 3.2 ControllerCXL.io · CXL.cache · CXL.mem · Type 1/2/3ATCATSDTI (shared)Back-invalMem poolingHost interfacesSFI · AXI4 · CPI · CHISerDes / PHYPartner-provided · high-speed physical layerncNoC local fabricIO control plane · SigIPUCIe D2D← to ComputeUCIe D2D→ to TransportPCIe / CXL lanes → host
01 ComputeUCIe D2D02 IOUCIe D2D03 TransportCXL C2CExternal AI / HPC systems
Core IP blocks

PCIe Gen 6 and CXL 3.2 —
designed from scratch.

No legacy architecture. No heritage PCIe Gen 3 foundations. SignatureIP's PCIe Gen 6 and CXL 3.2 controllers were designed from scratch: resulting in the lowest-power, highest-performance implementations available from any IP vendor.

SigIP · Interface

PCIe Gen 6 controller

Industry-first clean-sheet PCIe Gen 6 IP: certified by PCI-SIG in September 2025. Supports all link widths at the high bandwidth efficiency with modular power management architecture.

  • RC / EP dual mode: single IP, configurable at synthesis
  • All link widths: x1, x2, x4, x8, x16
  • 1 GHz operating frequency target
  • PAM4 signaling with FLIT mode support
  • TLP processing, flow control, error handling
  • SFI / AXI4 / CPI / CHI host interfaces
  • Advanced power management: L0s, L1, L1ss, L2
  • PCI-SIG certified (32 GT/s, non-flit mode) (Sept 2025)
Key differentiator
Lowest power PCIe Gen 6 on the market: no legacy overhead from prior generations. Modular power domains reduce active power by design, not by workaround.
SigIP · Interface

CXL 3.2 controller

CXL 3.2 solution with memory pooling and back-invalidation supporting memory expansion, memory pooling, and coherent device attachment: built on the same clean PCIe Gen 6 foundation, sharing partner-provided SerDes/PHY layers for area efficiency.

  • CXL.io: PCIe-compatible device enumeration
  • CXL.cache: device-to-host cache coherency
  • CXL.mem: coherent memory access protocol
  • Type 1, Type 2, and Type 3 device support
  • Multi-level memory pooling and switching
  • Back-invalidation support (CXL 3.x)
  • CXL Consortium member: active spec contributor
CXL chip-to-chip role
CXL from the IO chiplet reaches external AI servers, memory expanders, and accelerators. The Memory Expander silicon extends this with CXL 3.2 switching for multi-node scale-out topologies.
SigIP · PHY

SerDes / PHY (partner)

High-speed serializer/deserializer for PCIe Gen 6 and CXL 3.2 physical layer: multi-rate, PMA integration, validated across leading process nodes.

  • Multi-rate: 2.5 GT/s through 32 GT/s non-flit mode
  • PAM4 support for Gen 6 speeds
  • PMA integration with equalization
  • Process-node portable (TSMC, Samsung, Intel Foundry)
  • Shared across PCIe and CXL lanes
SigIP · UCIe / D2D

UCIe D2D interface (partner)

The IO chiplet sits in the middle of the UCIe chain: with one gateway connecting to the Compute chiplet and a second connecting to the Memory Expander silicon. Both gateways are SignatureIP IP, giving you single-vendor D2D coverage across the full package.

  • Two UCIe D2D interface (partner) instances in IO chiplet
  • Compute ↔ IO: UCIe D2D link
  • IO ↔ Transport: UCIe D2D link
  • UCIe standard-compliant adapter
  • Sub-nanosecond short-reach chip-to-chip latency
  • Validated end-to-end with Compute and Memory Expanders
CXL deep-dive

CXL 3.2: the
memory fabric layer.

CXL transforms the IO chiplet from a simple PCIe host into a coherent memory fabric endpoint. Type 1, 2, and 3 device support covers accelerators, smart memory expanders, and pooled memory in data-center and AI inference clusters.

PropertyValue
Spec versionCXL 3.0 / 3.2
Physical layerPCIe Gen 6 (shared)
Protocols supportedCXL.io · CXL.cache · CXL.mem
Device typesType 1 (accel), Type 2 (coherent accel), Type 3 (mem)
Memory poolingMulti-host shared memory (CXL 3.x)
Back-invalidationSupported (CXL 3.x back-invalidation snoop)
ConsortiumCXL Consortium member

CXL.io

PCIe-compatible discovery, configuration, and non-coherent DMA. Standard enumeration path for operating systems and hypervisors: lowest integration friction.

CXL.cache

Device-side cache coherency protocol: AI accelerators cache host memory lines, dramatically reducing redundant traffic over the link while maintaining coherency guarantees.

CXL.mem

Host-coherent memory access to attached memory expanders. Enables memory pooling and tiering architectures: critical for AI training and inference at scale where capacity exceeds on-package memory.

UCIe chip-to-chip

The IO chiplet is the
UCIe hub silicon.

The IO chiplet carries two UCIe D2D interface (partner)s: one on each side of the silicon. It is the central silicon in the three-chiplet package, bridging compute traffic from the Compute chiplet through to the Memory Expander silicon and out to chip-to-chip CXL links. All three UCIe D2D interface (partner)s across the full package are SignatureIP UCIe D2D IP.

← UCIe to ComputeIO siliconUCIe to Transport →
Both UCIe D2D interface (partner)s: UCIe D2D interface, short-reach D2D, partner-integrated on all dies.

UCIe interface specs

Gateway instances2 (one per adjacent silicon)
Interface typeUCIe D2D (UCIe D2D gateway (partner))
Left linkIO ↔ Compute chiplet
Right linkIO ↔ Memory Expander silicon
Package typeAdvanced packaging / standard pkg
SigIP coverageAll three silicon units: Compute, IO, Transport
Use cases

Where the IO chiplet fits.

01
Data center accelerator card
PCIe Gen 6 host interface connects the accelerator card to the server CPU. CXL.mem allows host-coherent memory expansion: eliminating separate memory copy operations.
02
AI inference chiplet SoC
IO chiplet provides the host connectivity layer while the Compute chiplet runs inference. CXL keeps model weights in coherent memory, reducing latency and bandwidth waste.
03
HPC memory fabric node
CXL 3.2 memory pooling enables shared memory across multiple nodes via the Memory Expander silicon's CXL switch. IO chiplet is the gateway from compute to the CXL fabric.