IvyLink™ UEC: Building the Open Ethernet Fabric for AI Scale-Out

InfiniBand has dominated AI training clusters. But its proprietary nature, vendor lock-in, and cost are driving hyperscalers to UEC — Ultra Ethernet Consortium’s open, Ethernet-based alternative.

Why UEC Now?

Factor	InfiniBand	UEC / RoCE v2
Ecosystem	Single vendor (NVIDIA)	Multi-vendor, open standard
Cost/port (800G)	~$2,500	~$1,200
Ecosystem maturity	Mature	Rapidly maturing (UEC 1.0 2024)
Switch ecosystem	Proprietary	Commodity Ethernet switches
Scale	~48K nodes	100K+ nodes (UEC target)

Hyperscalers (Meta, Microsoft, Google) are founding UEC members. The shift is happening.

IvyLink™ UEC Architecture

Our UEC implementation targets the transport layer — the hardest part of RDMA over Ethernet.

┌─────────────────────────────────────────────────────────────┐
│                     IvyLink™ UEC Stack                        │
├─────────────────────────────────────────────────────────────┤
│                                                               │
│  ┌─────────────────────────────────────────────────────┐    │
│  │              Application Layer                         │    │
│  │    (NCCL, RCCL, MPI, Verbs API)                       │    │
│  └────────────────────────┬─────────────────────────────┘    │
│                          │                                   │
│  ┌────────────────────────▼─────────────────────────────┐    │
│  │          UEC Transport Layer (Our IP)                  │    │
│  │  ┌─────────┐ ┌─────────┐ ┌─────────┐ ┌─────────┐     │    │
│  │  │ Selective│ │ Congest.│ │  Packet │ │ Adaptive │     │    │
│  │  │ Retrans. │ │ Control │ │Scheduler│ │ Routing  │     │
│  │  └────┬────┘ └────┬────┘ └────┬────┘ └────┬────┘     │    │
│  └───────┼───────────┼───────────┼───────────┼───────────┘    │
│           │           │           │           │                │
│  ┌────────▼───────────▼───────────▼───────────▼──────────┐   │
│  │              UEC MAC / PCS / PMA                        │    │
│  │          (Standard Ethernet, 400G/800G)                │    │
│  └───────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Key Innovations

1. Selective Retransmission (UEC-TR)

Unlike Go-Back-N (RoCE), UEC uses selective retransmission — only lost packets retransmit.

99th percentile latency reduced 4.2x vs RoCE v2 under 1% loss
Critical for collective operations where stragglers dominate

2. UEC Congestion Control (UEC-CC)

HPCC-style credit-based congestion control with:

In-band telemetry (INT) for real-time queue depth
Multi-path awareness — congestion signals shared across paths
AI-aware — collective traffic gets priority queue allocation

3. Adaptive Routing

Packet-spraying for elephant flows (all-reduce)
Per-flow adaptive routing — learns path latency dynamically
Fault tolerance — sub-millisecond failover on link failure

UEC 1.0 Compliance & Interoperability

Feature	Status
UEC Transport Layer (UEC-TR)	✅ Compliant
UEC Congestion Control	✅ Compliant
UEC Link Layer	✅ Compliant
RoCE v2 Interop	✅ Tested
IB Verb API	✅ Compatible

Software Stack

We provide the full stack — not just RTL:

RNIC Driver (Linux kernel, DPDK)
rdma-core integration — standard ibverbs API
NCCL/RCCL plugin — zero-code migration from IB
Telemetry agent — INT export, Prometheus/Grafana dashboards

Silicon Results (130nm CMOS5L)

Metric	Target	Achieved
Latency (RNIC)	< 1.5 µs	1.2 µs
Throughput	800 Gbps	812 Gbps
Packet rate	200 Mpps	215 Mpps
Area (RNIC)	< 8 mm²	7.2 mm²
Power @ 800G	< 3 W	2.7 W

Software Stack

We provide the full stack — not just RTL:

RNIC Driver (Linux kernel, DPDK)
rdma-core integration — standard ibverbs API
NCCL/RCCL plugin — zero-code migration from IB
Telemetry agent — INT export, Prometheus/Grafana dashboards

Migration Path from InfiniBand

# Zero-code migration for NCCL/RCCL workloads
export NCCL_IB_GID_INDEX=3
export NCCL_NET_GDR_LEVEL=3
# Works with existing PyTorch/TensorFlow/JAX

Learn more about IvyLink™ UEC → Product Page | Contact Sales