InfiniBand vs Ethernet-AI Product Landscape — May 2026

The fabric war is no longer NVLink/InfiniBand vs "someday Ethernet." Since UEC 1.0 shipped (2025-06-11, ~1,000,000-endpoint transport target), the contest is between *proprietary low-latency switched fabrics* and *standards-track Ethernet-AI* — with optical I/O quietly redefining the scale-up layer underneath both. This module maps who actually ships at scale versus who is selling a roadmap.

Product Categories

• **Standards-track Ethernet-AI transport**: UEC 1.0-aligned scale-out fabrics targeting ~1M endpoints for training (Cornelis CN6000 [3e9e89c0], Meta DSF [7d6e6779/2a7d3150], HPE post-UEC [d906944d]).
• **Proprietary low-latency switched HPC fabric**: sub-microsecond, hardware adaptive routing, in-network collectives, IB-NDR-class — predates or sidesteps UEC (Cornelis CN5000 [199f4223], HPE Slingshot proprietary Ethernet variant [efda56e4]).
• **Converged SuperNIC + switch silicon**: single-die NIC/switch convergence at multi-Tbps with multi-port resiliency, plus CXL memory-fabric extension (Enfabrica ACF-S Millennium [3f45d2e2], EMFASYS [83940080]).
• **Co-packaged optical I/O for scale-up interconnect**: optical chiplet + external laser for chip-to-chip bandwidth beneath the fabric (Ayar Labs TeraPHY + SuperNova [2e6de52c]).

Comparison Table

| Product | Company | Quality Anchor (0-100) | Peak Proven Scale / Spec | Use Case | |---------|---------|------------------------|--------------------------|----------| | CN5000 | Cornelis Networks | 55 [199f4223] | sub-µs latency, 500k+ endpoints, in-network collectives | IB-NDR-class national-lab HPC scale-out | | CN6000 | Cornelis Networks | unscored, 2026 roadmap [3e9e89c0] | 800 Gbps, UEC 1.0 ~1M-endpoint target | Standards-track future scale-out (unshipped) | | ACF-S "Millennium" | Enfabrica | 50 [3f45d2e2] | ~3.2 Tbps converged SuperNIC+switch, multi-port resiliency | Resilient rail-optimized NIC fabric (no production) | | TeraPHY + SuperNova | Ayar Labs | 70 [2e6de52c] | optical I/O chiplet + external laser source | Scale-up chip-to-chip optical interconnect | | Slingshot | HPE | 75 [efda56e4] | El Capitan + Frontier exascale (2+ marquee refs) | Exascale Ethernet HPC fabric, shipping today | | Disaggregated Scheduled Fabric (RoCE) | Meta Platforms | ~75 [7d6e6779] | 24,576-GPU production Llama-class training | Hyperscale Ethernet training without IB lock-in | | Tofino | Intel | 12 [b97c209c] | discontinued 2023, P4 software ecosystem only | Programmable-switch silicon — avoid |

Differentiation Map

• **Best for exascale HPC + AI on a shared fabric**: HPE Slingshot — only fabric proven today at El Capitan and Frontier exascale scale [efda56e4].
• **Best for hyperscale 24k-GPU Llama-class training without NVIDIA IB lock-in**: Meta DSF — RoCE fabric shipping in production at 24,576 GPUs [7d6e6779].
• **Best for sub-µs national-lab scale-out with in-network collectives**: Cornelis CN5000 — IB-NDR-class with hardware adaptive routing, DOE references [199f4223][3a1cd12f].
• **Best for scale-up chip-to-chip optical bandwidth beneath any fabric**: Ayar Labs TeraPHY — co-packaged optics engaged by Intel/AMD/GlobalFoundries [2e6de52c].
• **Avoid if you need shipping first-party programmable-switch silicon**: Intel Tofino — line wound down 2023, P4 software only, no silicon roadmap [b97c209c].
• **Avoid if you need production-validated converged SuperNIC today**: Enfabrica ACF-S — architecturally ahead of BlueField-3/Thor2 but zero public production deployment [3f45d2e2][9da149bc].

Tool: Interactive Comparator

[Link to /tools/infiniband-ethernet-compare] — spec: filter the 7 rows above by quality anchor band, proven-scale type (endpoints / GPUs / Tbps), and standards-track vs proprietary axis; surface the data_point id behind every cell. Frontend wiring lands in Phase G.

Reverse-Hype Warnings

**Overpromised.** Enfabrica is the clearest gap between architecture and evidence: ACF-S "Millennium" claims ~3.2 Tbps leadership over BlueField-3 and Thor2 [3f45d2e2], and EMFASYS extends into ~18TB CXL pooled memory [83940080] — but *neither product has a public production deployment or shipped-on-time track record*, and the empty business-signals array caps the honest score near 40 [9da149bc]. Two unproven products is roadmap breadth as risk, not strength. Cornelis CN6000's UEC 1.0 alignment is sound positioning, but the 800 Gbps part is not expected until 2026 — the standards bet is unproven in production, a caveat not a delivered win [3e9e89c0]. Intel's networking story is the harshest reality check: Tofino is discontinued (~12) [b97c209c] and Gaudi+Ethernet publicly walked back its ~$500M 2024 revenue goal [ef97dd6f]; the only on-time execution (18A/Panther Lake at CES 2026, ~55) is a foundry proxy, explicitly *not* a fabric-product claim [a9d99071]. Meta's DSF, despite real 24k-GPU proof, still carries unresolved RoCE congestion-control and lossless-tuning risk at multi-GW Prometheus/Hyperion scale [326da885]. HPE Slingshot's proprietary Ethernet moat is simultaneously legitimized and eroded by UEC 1.0 standardization pressure from Spectrum-X, Tomahawk, and Arista [d906944d]. Ayar Labs faces a crowded co-packaged-optics field (Broadcom, Marvell, Celestial AI, Lightmatter, Avicena) — differentiation must be continuously defended [1e7d20c1].

**Underrated.** Cornelis CN5000 ships *genuine* delivered differentiation — sub-µs latency, hardware adaptive routing, fine-grained congestion control, 500k+ endpoints — at IB-NDR parity with selective leadership [199f4223], yet sits below HPE/Meta only because it lacks named hyperscaler case studies [3a1cd12f], not because the product is weaker. And Meta DSF is underrated relative to roadmap-promise vendors: it is a working alternative to InfiniBand *already in production at 24,576 GPUs* [7d6e6779], with Meta also shaping the UEC 1.0 baseline as a founding steering member [2a7d3150] — shipped scale beats announced Tbps.