UALink (Ultra Accelerator Link): A Full Guide to Open Scale-Up Fabric When NVLink Is Not Enough

NVLink still owns the densest GPU pods that ship today. That lock-in is exactly why hyperscalers, CPU vendors, and accelerator startups spent two years building an open alternative for scale-up.

This guide covers what Ultra Accelerator Link (UALink) is, how the consortium formed, what Spec 1.0 (April 2025) and Spec 2.0 (April 2026) actually ratified, how UALink sits next to NVLink, InfiniBand, and Ultra Ethernet, who is shipping silicon and IP, and the honest limits that keep NVIDIA fabrics on the same BOM through 2026 and into 2027.

Key Takeaways

• UALink is an open scale-up accelerator interconnect; Spec 1.0 enables 200G per lane and pods up to about 1,024 accelerators.
• The consortium launched as a May 2024 promoter group and incorporated in October 2024 with a hyperscaler-heavy board.
• April 2026 Spec 2.0 adds In-Network Compute, a UCIe 3.0-aligned Chiplet Spec, Manageability (gNMI/YANG/SAI/Redfish), and a split 200G DL/PL Spec.
• AMD Instinct, Astera Labs, Synopsys IP, and switch partners are on the critical path; full radix switch silicon still lags Spec paper.
• UALink is not a drop-in NVLink replacement in NVIDIA pods, and it is not a substitute for Ultra Ethernet or InfiniBand scale-out.

What UALink actually is

UALink is a scale-up interconnect: low-latency, high-bandwidth links among accelerators (and switches) inside an AI computing pod. The consortium positions it as an open alternative to proprietary GPU fabrics so multi-vendor XPUs can share a common load/store model across a rack-scale domain.

According to the UALink 200G 1.0 specification page and the April 8, 2025 BusinessWire release, Spec 1.0 defines communication between accelerators and switches, enables 200G per lane, and targets pods of up to 1,024 accelerators. The public white paper describes a load/store protocol over an Ethernet-family physical layer (IEEE 802.3dj-class signaling), with a four-lane station model and optional lower-rate paths including a separate 128G Data Link and Physical Layers specification.

Here's why that framing matters. Operators buying "open interconnect" are not buying a press release. They are buying a documented way for AMD Instinct GPUs, custom ASICs, and future XPUs to grow a coherent scale-up domain without licensing a single GPU vendor's switch stack.

Inside Deep Tech's earlier overview, NVLink, InfiniBand, and UALink: How AI GPUs Actually Talk, maps the three-layer split. This article goes deep on UALink alone: governance, Spec 1.0 vs Spec 2.0, vendor timelines, and where the standard still loses to shipping proprietary fabrics.

Consortium history: from promoter group to Spec 2.0

On May 30, 2024, AMD, Broadcom, Cisco, Google, Hewlett Packard Enterprise, Intel, Meta, and Microsoft announced the UALink Promoter Group. The charter was explicit: standardize a high-speed, low-latency scale-up interconnect for AI pods, aiming for Spec 1.0 with up to 1,024 accelerators and direct loads/stores between accelerator-attached memory.

The UALink Consortium incorporated in October 2024 (Beaverton, Oregon). By Spec 1.0 and Spec 2.0 releases, the board listed in consortium materials included Alibaba, AMD, Apple, Astera Labs, AWS, Cisco, Google, HPE, Intel, Meta, Microsoft, and Synopsys. At the Spec 1.0 announcement the consortium said it represented more than 85 member companies.

That roster is the point. UALink is not a boutique academic fabric. It is a hyperscaler- and CPU-vendor-backed attempt to keep scale-up from becoming a permanent single-vendor tax.

Spec 1.0 (April 2025): 200G lanes and thousand-accelerator pods

On April 8, 2025, the consortium ratified the UALink 200G 1.0 Specification. Public materials highlight:

• 200G per lane scale-up connections for accelerators and switches inside a pod.
• Support for pods up to about 1,024 accelerators (10-bit accelerator identifiers in the protocol model).
• A load/store programming model aimed at simple software across the pod.
• Ethernet-class PHY roots (802.3 family rates including 200GBASE-KR1/CR1 class paths), so the industry can reuse serdes and optics supply chains.
• A claimed ~93% effective peak bandwidth target for deterministic performance (consortium Spec 1.0 PR language; treat as a design goal, not a measured rack result).

A parallel UALink 128G DL/PL 1.0 path exists for implementations that need a lower per-lane rate. Spec 1.0 is the baseline buyers should still reference when a vendor claims "UALink ready" without naming Spec 2.0 features.

Spec 2.0 (April 2026): In-Network Compute, chiplets, and manageability

On April 7, 2026, the consortium published four related specifications. The BusinessWire primary release and the consortium's Spec 2.0 PR PDF define:

• UALink Common Specification 2.0: In-Network Compute for reductions, aggregations, synchronization, and data transformations inside the fabric path, aimed at lower latency and better bandwidth efficiency for distributed training and inference.
• UALink 200G Data Link / Physical Layer Specification 2.0: DL/PL split from Common so PHY rates can evolve without rewriting higher layers.
• UALink Chiplet Specification 1.0: interfaces, form factors, flow control, and chiplet management for integrating UALink into chiplet SoCs; fully compliant with UCIe 3.0.
• UALink Manageability Specification 1.0: centralized control and management planes using gNMI, YANG, SAI, and Redfish.

That package is the real 2026 story. Spec 1.0 proved the industry could write an open scale-up protocol. Spec 2.0 is where the consortium admits that pods are systems: they need collective offload, modular silicon packaging, and operators who can manage the fabric with the same tooling they already use for Ethernet switches.

For packaging context, see Inside Deep Tech's Chiplets and UCIe Full Guide. UCIe alignment is not marketing fluff; it is how UALink IP is supposed to land next to compute dies without a one-off interposer protocol.

Scale-up vs scale-out: UALink next to NVLink, InfiniBand, and Ultra Ethernet

Mixing these labels is how BOMs get expensive. Keep the columns separate.

UALink competes with NVLink for the scale-up column. It does not replace Ultra Ethernet or InfiniBand on the scale-out backend. Large systems typically need both: a tight accelerator domain plus a node-to-node fabric.

AMD's Linux bring-up work for UALink over Ethernet (UALoE) on Helios-class Instinct platforms (reported via Phoronix coverage of AMDGPU patches in August 2026) shows a pragmatic transitional path: reuse Ethernet PHYs and NICs while the dedicated switch ecosystem matures. Treat UALoE as a deployment bridge, not as proof that "UALink equals Ethernet networking."

Host attach remains a different problem. PCIe 6.0 still feeds NICs, storage, and many accelerator hosts even when the GPU-to-GPU path is NVLink or UALink.

Ultra Ethernet: A Full Guide to AI Scale-Out Fabric After InfiniBand

UEC Spec 1.0.3, UET transport, profiles, and congestion control for the scale-out column that sits beside UALink.

Inside Deep Tech

Who is shipping or developing UALink paths

AMD Instinct

AMD is a founding promoter and board member. Public roadmap commentary and supply-chain reporting place UALink on the Instinct MI400-generation path, with larger MI500-class rack-scale pods (including UAL256-style configurations discussed in 2027 planning coverage) depending on partner switch ASICs. AMD does not claim to ship a full proprietary NVSwitch clone; it depends on the open ecosystem for radix.

Software is moving in parallel. AMD's UALink / UALoE kernel patch series for Helios and Instinct MI455X-class bring-up is the visible open-source signal that the company intends production software paths, not only slides.

Astera Labs and switch partners

Astera Labs, a board member, has publicly described a portfolio plan around UALink switches, retimers, and controllers alongside its Scorpio-X fabric work and Aries retimers, aimed at AMD Instinct and custom XPU racks. Other switch names appear in ecosystem statements (for example XConn in Synopsys collaboration materials). The honest constraint: merchant UALink switch silicon availability in 2026 remains the long pole. Mesh or small topologies can ship earlier than full 1,024-endpoint switched pods.

Synopsys IP and design enablement

On December 11, 2024, Synopsys announced UALink controller, PHY, and verification IP, with a stated availability target in the second half of 2025. That IP layer is how custom ASIC teams and GPU vendors actually put UALink ports on dies. Without it, Spec PDFs stay PDFs.

Marvell and other custom silicon houses show up in industry commentary as potential implementers; cite only what they publish under their own names when you need SKU-level claims.

Honest limits: when NVLink still wins

Named downsides you should put on the whiteboard:

• Not a drop-in NVLink replacement inside NVIDIA pods. CUDA libraries, NVSwitch topologies, and NVIDIA networking SKUs remain a vertical stack.
• Switch silicon lag: Spec 1.0 and Spec 2.0 do not ship radix by themselves. 2026 deployments may be limited to smaller meshes until partner ASICs land at volume.
• Software and ecosystem maturity: collectives, profiling, and operations tooling still trail the NVLink/NVSwitch experience curve.
• UALoE and Ethernet-PHY reuse help schedule risk, but they do not magically equal dedicated UALink switch latency and bandwidth profiles.
• Open ISA or open interconnect does not mean free silicon. IP, PHYs, and switches remain commercial products (see also Inside Deep Tech's RISC-V for AI guide for the same pattern on the compute side).

When NVLink still wins: NVIDIA-heavy fleets, jobs that already saturate NVSwitch domains, teams that cannot wait on partner switch schedules, and any procurement that prices time-to-train above interconnect openness.

NVLink, InfiniBand, and UALink: A Full Guide to How AI GPUs Actually Talk

The three-layer interconnect map this Full Guide extends with Spec 2.0 depth.

Inside Deep Tech

What operators should do Monday

• Ask vendors which UALink revision they claim (1.0 vs Common 2.0 + Chiplet + Manageability + DL/PL 2.0) and which switch ASIC generation is on the BOM.
• Separate scale-up (UALink/NVLink) from scale-out (UEC/IB/RoCE) on every rack drawing.
• If you run multi-vendor XPUs, put UALink compliance and UCIe chiplet attach on the RFQ; if you run all-NVIDIA pods, treat UALink as a 2027 option, not a 2026 migration.
• Track Manageability Spec 1.0 against your existing gNMI/YANG/Redfish toolchain so fabric ops do not invent a second plane.
• Read primary sources: ualinkconsortium.org specs, the April 2026 BusinessWire Spec 2.0 release, and vendor IP/switch roadmaps dated within the last 12 months.

The industry did not invent UALink because NVLink stopped working. It invented UALink because too many AI systems will not be single-vendor forever. The teams that win will treat Spec 2.0 as a compliance checklist with failure modes, not as a slogan for "open and therefore free."

FAQ