GPU SmithCOMPUTE ENGINEERING
Back to Articles|Published on October 28, 2026|15 min read
Download PDF

GPUSmith Article

UALink vs NVLink: Open Scale-Up Fabrics Explained

UALink vs NVLink as of October 2026: specs, pod sizes, AMD Helios UALink over Ethernet, and which UALink switches are announced or shipping.

By GPU Smith

UALink vs NVLink: Open Scale-Up Fabrics Explained

Summary

  1. 01For 2026 deliveries, the article identifies NVLink racks and AMD's Ethernet-based UALoE racks as the real choices. No vendor has publicly confirmed shipping a native UALink switch; dated native-switch plans point to 2027.
  2. 02UALink is an open scale-up specification. Helios uses UALink over Ethernet on Broadcom Tomahawk 6 switches, and no UALoE document appears among the consortium's published specifications.
  3. 03The current NVLink 6 specification is 3 TB/s per GPU and 216 TB/s per NVL72. The NVLink Fusion page still carries the higher launch figures, so buyers should obtain the number for their exact SKU in writing.
  4. 04UALink's 93% effective peak bandwidth and approximately 1 microsecond round-trip latency are design targets. The article found no published hardware measurement, and neither headline figure establishes measured collective bandwidth.
  5. 05Switch suppliers, protocol documentation and measured collective performance matter for procurement. Native UALink interoperability remains unproven, and neither AMD nor HPE has said Helios UALoE trays can be swapped for native UALink switches.
Inside this article
  1. 01The short answer
  2. 02Scale-up vs scale-out in one page
  3. 03The UALink Consortium and its specifications
  4. 04What UALink 1.0 actually defines
  5. 05Reading the performance claims
  6. 06NVLink 6 and NVLink Fusion
  7. 07Where AMD stands: Infinity Fabric today, UALoE in Helios
  8. 08Side-by-side comparison (as of 2026-10-28)
  9. 09Availability reality check
  10. 10Procurement implications
  11. 11Decision guide
  12. 12Sources

The short answer

As of October 28, 2026, NVLink is a shipping scale-up fabric from a single vendor. UALink is an open specification, and no vendor has publicly confirmed shipping a native UALink switch. NVLink 6 connects 72 GPUs in a Vera Rubin NVL72 rack. UALink 1.0 defines 200G-per-lane links for pods of up to 1,024 accelerators [1]. Vendors that have given dates for native UALink switches point to 2027. AMD Helios, the first "UALink" rack you can order, runs UALink over Ethernet (UALoE) on Broadcom Tomahawk 6 switches, not on native UALink silicon [2].

So for 2026 deliveries the real choice is NVLink racks or AMD's Ethernet-based UALoE racks. Native UALink is a decision for 2027 and later.

72GPUs connected by NVLink 6 in a Vera Rubin NVL72 rack
1,024Accelerator addressing ceiling for UALink, rather than an announced system size
3 TB/sCurrent Rubin GPU NVLink bandwidth specification, bidirectional per GPU
93%UALink protocol payload divided by physical-layer capacity, a modeled design target

Scale-up vs scale-out in one page

A scale-up fabric joins accelerators into one memory domain. GPUs read, write and run atomics directly against each other's memory through purpose-built switches. UALink describes itself in these terms: "read, write, and atomic transactions" over "a simple load/store protocol" [1]. Tensor-parallel and mixture-of-experts (MoE) all-to-all traffic runs on this layer. NVIDIA cites MoE all-to-all as the reason for the NVL72 topology [3].

A scale-out fabric (InfiniBand, RoCE, Ultra Ethernet) connects those domains over RDMA across thousands of nodes. It is covered in our scale-out fabric comparison. UALink stops at the pod boundary: "inter-Pod communication and control is not specified by the UALink specification" [4]. Helios pairs UALoE for scale-up with Ultra Ethernet-aligned Ethernet for scale-out [5].

The UALink Consortium and its specifications

The consortium was incorporated in October 2024. Its board is Alibaba, AMD, Apple, Astera Labs, AWS, Cisco, Google, HPE, Intel, Meta, Microsoft and Synopsys; NVIDIA is not a member of the board [6]. As of October 28, 2026, the specification page lists six documents, all available for download [7]:

SpecificationReleasedWhat it adds
UALink 200G 1.02025-04-08Base protocol, 200G/lane, up to 1,024 accelerators
UALink 128G Data Link and Physical Layers 1.02025128G-per-lane option (the page's description reads "coming soon")
UALink Common Specification 2.02026-04-07In-Network Compute
UALink 200G Data Link and Physical Layers 2.02026-04-07Physical and data link layers split out, so new speeds don't require changes to the other specs
UALink Manageability 1.02026-04-07Control and management planes using gNMI, YANG, SAI and Redfish
UALink Chiplet 1.0 (listed as 1.01)2026-04-07Chiplet integration, "fully compliant with the UCIe 3.0 Specification"

The 2026 release dates and contents come from the 2.0 release [6]. The chair's roadmap post describes 1.0 as "at 200G and 128G per lane" and schedules UALink 3.0 for 2027: more bandwidth, reach "across racks and rows" and better reliability [8]. No "128G 2.0" document exists. Compliance and interoperability programs are still described as planned [6].

What UALink 1.0 actually defines

Source for all of the following: UALink 1.0 white paper, 2025-04.

  • Lane rate: up to 200 GT/s of data per lane, signaled at 212.5 GT/s to leave room for Ethernet FEC and encoding. A 106.25G option is also defined.
  • Links: x1, x2 or x4. A four-lane "Station" carries at most 800 Gbps in each direction.
  • PHY: based on IEEE 802.3 (200GBASE-KR1/CR1 through 800GBASE-KR4/CR4). The PMD is unchanged; lower-interleave FEC modes cut latency.
  • Flits: 64-byte transaction flits are packed into 640-byte data-link flits. Each fits exactly one RS(544,514) codeword and carries a 32-bit CRC and link-level replay.
  • Scale: switches "connect up to 1024 Accelerators" and can be partitioned into isolated Virtual Pods. UALinkSec adds encryption for confidential-computing tenants.

The 1,024 figure is an addressing ceiling. No announced system we found comes near it, and shipping racks stop at 72. Because UALink reuses the Ethernet PHY, it shares SerDes, cables and retimers with 200G Ethernet. Astera Labs, for example, sells retimers "for 224G Ethernet and UALink connectivity" [9]. Above the PHY, though, UALink is not Ethernet, and a UALink switch is not an Ethernet switch.

Putting 93% next to an NVLink TB/s figure compares a paper target with a marketing figure. Neither is a measured collective bandwidth.

Reading the performance claims

The consortium says UALink offers "the same raw speed as Ethernet with the latency of PCIe switches" and is "designed for deterministic performance achieving 93% effective peak bandwidth" [1]. Treat both as design targets:

  • A consortium-hosted 2026 white paper defines 93% as protocol payload divided by physical-layer capacity. It says the targets were "validated through architectural modeling and protocol simulation" and that "empirical verification across multi-vendor hardware is planned for 2026" [10]. We found no published hardware measurement.
  • The same paper gives "approximately 1 microsecond round-trip latency" over copper shorter than four meters. That is also a specification target.

Putting 93% next to an NVLink TB/s figure compares a paper target with a marketing figure. Neither is a measured collective bandwidth.

NVLink 6 and NVLink Fusion

NVLink is NVIDIA's proprietary GPU-to-GPU link, extended across the rack by NVLink Switch chips. Generational detail is in NVLink 5 vs NVLink 6. One point matters here: the current NVLink 6 figure is 3 TB/s per GPU and 216 TB/s per NVL72, but one NVIDIA page still shows the higher launch figure, because NVIDIA revised it after launch.

NVIDIA pagePer-GPU NVLink 672-GPU domain
NVLink ("Preliminary specifications; may be subject to change")3,000 GB/s, 36 links216 TB/s (NVL72)
Vera Rubin NVL72 spec table3 TB/s216 TB/s
NVLink Fusion (launch figure, not updated)"3.6 TB/s per XPU""260 TB/s ... (NVL72)"

The current figure checks out: 72 × 3 TB/s = 216 TB/s. The Fusion page's launch figure measures the same thing, bidirectional bandwidth per GPU in a Vera Rubin NVL72, so the difference is a revision, not a definition. NVIDIA announced 3.6 TB/s and 260 TB/s at CES 2026 [11], and the Vera Rubin NVL72 page still listed them in June 2026. The current product pages and the Vera Rubin datasheet (dated 2026-10-14) give 3 TB/s, 216 TB/s and "1.7x" over NVLink 5. The Fusion page and NVIDIA's developer blogs still carry the launch figure. Treat 3 TB/s as the current Rubin GPU specification, and get the number for your exact SKU in writing. For comparison, GB200 NVL72 is 1.8 TB/s per GPU and 130 TB/s per rack [12].

  • Domain size: NVLink 6 domains are listed as "8 | 72" [3]. The Fusion page cites "future roadmap domain sizes up to 1,152", a roadmap figure in the same way that UALink's 1,024 is a specification ceiling.
  • In-network compute: every NVLink Switch has SHARP engines for reductions and multicast (same source). UALink added In-Network Compute on paper in April 2026.
  • Availability: "Production shipments of Vera Rubin are set to begin starting this fall" [13]. The product page now reads "Available Now".

NVLink Fusion lets hyperscalers "deploy custom XPUs and CPUs into NVIDIA's world-leading AI infrastructure platform" on MGX racks [14]. It opens the endpoint, not the fabric: NVIDIA still supplies the switch. Fusion-specific checks are in NVLink Fusion due diligence.

Where AMD stands: Infinity Fabric today, UALoE in Helios

Until Helios, AMD's scale-up domain was an 8-GPU Infinity Fabric node; see NVLink vs Infinity Fabric. AMD describes Helios this way: "Four scale-up cartridges leverage UALink over Ethernet (UALoE) to seamlessly connect up to 72 GPUs ... with up to an industry leading 260 TB/s aggregate scale-up bandwidth." Each switch tray holds "Two 512 lane 200G UALoE switch ASICs with 216 UALoE (x2) active links per ASIC", rated at "21.6 TB/s (Bi-Dir) bandwidth per switch" [15]). HPE builds the switching with Broadcom's Tomahawk 6, using six HPE Juniper Networking QFX5252 scale-up Ethernet switch trays per rack ([2][16].

HPE calls UALoE an "open ... standard", but no UALoE document appears among the consortium's published specifications [7]. Ask which document defines the UALoE you are buying.

Worked example: do the Helios numbers add up?

The calculation uses only AMD and HPE figures plus two assumptions of ours. First, "216 UALoE (x2) active links" means 216 two-lane links. Second, each QFX5252 tray is one two-ASIC switch tray. AMD's "four scale-up cartridges" is not defined. Four two-ASIC trays would give only 172.8 TB/s, so "cartridge" probably does not mean switch tray.

  1. Active lanes per ASIC: 216 × 2 = 432 (out of 512).
  2. One direction: 432 × 200 Gb/s = 86.4 Tb/s = 10.8 TB/s.
  3. Both directions: 2 × 10.8 = 21.6 TB/s. This matches AMD's per-switch figure.
  4. ASICs per rack: 6 trays × 2 = 12.
  5. Rack total: 12 × 21.6 = 259.2 TB/s, rounded to "up to 260 TB/s".
  6. Per GPU: 259.2 / 72 = 3.6 TB/s bidirectional, or 72 lanes of 200 Gb/s per GPU (72 × 72 = 5,184 = 12 × 432).

So 260 TB/s is the sum of raw lane rates. It matches NVIDIA's CES 2026 launch figure, which is still on the Fusion page, and is 20% above the 216 TB/s in NVIDIA's current Vera Rubin specification.

Native UALink at AMD. As of October 28, 2026, we found no AMD primary statement committing a GPU to native UALink switches. AMD says "Next-generation AMD Instinct MI500 Series GPUs are coming in 2027" with "next-gen Pensando 'Como' and 'Monza' networking", but does not name the scale-up protocol [17]. On the Q1 2026 earnings call (third-party transcript; Astera's written results releases give no UALink switch date), Astera Labs' CEO said "Both Amazon as well as AMD have said that their ASIC and GPU will launch sometime in 2027, and we will certainly be prepared to intercept that launch with our UALink switch" [18]. That is a supplier describing its customers' plans, not a commitment from AMD.

Side-by-side comparison (as of 2026-10-28)

NVLink 6Native UALinkUALoE (Helios)
GovernanceNVIDIA proprietary; Fusion licenses endpointsOpen consortium specCalled an open standard; no consortium document found
Lane rateNot published200G (212.5 GT/s signaling); 128G option200G on Ethernet ASICs
Per-GPU3 TB/s (current spec; 3.6 TB/s at CES 2026 launch)Implementation-definedUp to 3.6 TB/s (derived)
Max domain72; 1,152 on roadmap1,024 (spec)72
In-network computeSHARP, shippingCommon Spec 2.0Not stated
Switch suppliersNVIDIAAstera Labs, Upscale, Marvell custom (announced)Broadcom Tomahawk 6
StatusProduction shipments from fall 2026No native switch confirmed shippingIn production; deployments from Q4 2026
Figure 01
Native UALink and Helios UALoE
Native UALinkOpen specification
  • UALink 1.0 defines 200G-per-lane links for pods of up to 1,024 accelerators; that scale is an addressing ceiling.
  • No vendor has publicly confirmed shipping a native UALink switch. Vendors with dates point to 2027.
  • Interoperability between vendors remains unproven until compliance programs exist.
Helios UALoEEthernet-based scale-up
  • Helios runs UALink over Ethernet on Broadcom Tomahawk 6 switches, rather than native UALink silicon.
  • Helios and NVLink racks both offer 72-GPU all-to-all domains for delivery in 2026.
  • HPE calls UALoE an open standard, but the consortium's published specifications contain no UALoE document.

Neither AMD nor HPE has said that Helios UALoE trays can be swapped for native UALink switches. Assume they cannot unless it is in the contract.

Availability reality check

Statuses use each vendor's own wording.

ProductStatus as of 2026-10-28
NVIDIA Vera Rubin NVL72 (NVLink 6)Production: "ramping into full production", shipments "starting this fall" [13]
AMD Helios (UALoE), OEM reference designProduction: "now in production"; OpenAI expects it online "beginning in the fourth quarter of 2026" [17]; "volume deployments expected in 2H 2026" [15]
HPE Juniper QFX5252 (Tomahawk 6 UALoE tray)Ordered: first Helios order, $1.2B from Vultr, 2026-09-30; no delivery date given [16]
Astera Labs native UALink switchAnnounced: "poised to expand further into 2027 with the introduction of UALink-based products" [18]
Astera Labs Scorpio X 320-laneShipping family, ramping part: Scorpio is "Now Shipping to Leading Hyperscalers" with "Open and Platform-Specific Protocols"; the X 320-lane has a "production ramp in 2H 2026" [19]. Not described as UALink.
Upscale SkyFabriX (115.2 Tbps)Early access: supports ESUN, UALoE, SUE-T and Ethernet/IP, plus "evolving protocols including UALink"; GA "planned for early 2027" [20]
Marvell custom UALink (SerDes, controller, switch-core IP)Announced 2025-06-11, no date given [21]; bought XConn to "expand our UALink scale-up switch team" [22]
Synopsys UALink IPLicensable IP, status not stated [23]; UALinkSec_200 runs on the "silicon-proven" 224G PHY [24]
Cadence UALink IPIP in silicon: "UALink-200G controller and PHY running in 3nm" [25]

The consortium-hosted white paper puts switching silicon at "Multiple vendors in development, evaluation hardware expected in late 2026" [10]. That matches what vendors say: controller and PHY IP runs in silicon, UALoE ships in Helios, and native switches arrive in 2027.

Procurement implications

  • Lock-in sits in the switch. NVIDIA supplies the switch for NVLink, including under Fusion. Helios splits suppliers across AMD, Broadcom and the OEM. However, UALoE is not a published consortium standard, so check the documents behind any "multi-vendor" claim.
  • Native UALink buys optionality from 2027, not in 2026. Interoperability between vendors stays unproven until compliance programs exist.
  • Upgrade paths. Neither AMD nor HPE has said that Helios UALoE trays can be swapped for native UALink switches. Assume they cannot unless it is in the contract.

Questions to ask any scale-up vendor:

  1. What is the per-GPU scale-up bandwidth for this exact SKU, in each direction, and where is it published?
  2. Which document defines the protocol on the wire?
  3. What all-reduce and all-to-all bandwidth have you measured on a full 72-GPU domain, and with which collective library version?
  4. Is in-network reduction available now, and in which software release?
  5. For native UALink: is the switch silicon announced, sampling or in production, and does it have compliance results and tested accelerator partners?
  6. Can the domain be partitioned for multi-tenancy, and is UALinkSec or an equivalent encryption feature enabled?
Figure 02
Questions for scale-up vendors
  1. 01Confirm exact-SKU bandwidth

    Ask for per-GPU scale-up bandwidth in each direction for the exact SKU, and where it is published.

  2. 02Identify the protocol document

    Ask which document defines the protocol on the wire.

  3. 03Request collective measurements

    Ask for measured all-reduce and all-to-all bandwidth on a full 72-GPU domain, including the collective library version.

  4. 04Check native switch readiness

    For native UALink, ask whether silicon is announced, sampling or in production, and request compliance results and tested accelerator partners.

Decision guide

This is our assessment, not a benchmark.

  • MoE at scale, delivered in 2026: NVLink racks or Helios, both 72-GPU all-to-all domains. Compare measured all-to-all bandwidth and software maturity, not headline TB/s.
  • Memory-bound long-context inference: HBM per domain matters more than the fabric. AMD lists 31 TB of HBM4 per Helios rack [15]; NVIDIA lists 20.7 TB per Vera Rubin NVL72 [26].
  • Your own XPU, 2027-2028: carry both NVLink Fusion and native UALink through design review. Fusion offers a fabric that already ships. UALink offers a choice of switch vendors, if Astera Labs, Upscale or Marvell hit their 2027 dates.
  • A few racks for enterprise or sovereign use: the fabric comes with the GPU vendor. Do not pay extra for "UALink-ready" until a native switch ships.

Recheck after AMD's and Astera Labs' next earnings reports and any announcement of a UALink compliance program.

GPU Smith helps teams compare scale-up fabrics and turn vendor claims into acceptance tests. Contact us.

Sources

  1. UALink Consortium Releases the Ultra Accelerator Link 200G 1.0 Specification (news release), UALink Consortium, 2025-04-08. https://ualinkconsortium.org/wp-content/uploads/2025/04/UALink-1.0-Specification-PR_FINAL.pdf
  2. Introducing UALink 200G 1.0 Specification (white paper), UALink Consortium, 2025-04. https://ualinkconsortium.org/wp-content/uploads/2025/04/UALink-1.0-White_Paper_FINAL.pdf
  3. Specifications, UALink Consortium, accessed 2026-10-28. https://ualinkconsortium.org/specification/
  4. UALink Consortium Publishes Four Specifications Defining In-Network Compute, Chiplets, Manageability and 200G Performance (news release), UALink Consortium, 2026-04-07. https://ualinkconsortium.org/wp-content/uploads/2026/04/UALink-2.0-Specification-PR_FINAL.pdf
  5. UALink Roadmap Insights: Accelerating Open, Scalable AI Networking, UALink Consortium blog (Kurtis Bowman), 2026-03-06. https://ualinkconsortium.org/blog/ualink-roadmap-insights-accelerating-open-scalable-ai-networking-1296/
  6. UALink: An Open, High-Efficiency Scale-Up Interconnect for AI (white paper), TASK Consultancy, hosted by UALink Consortium, 2026. https://ualinkconsortium.org/wp-content/uploads/2026/01/UALink_White_Paper_Publication_FINAL_UPDATED.pdf
  7. NVIDIA NVLink, NVIDIA, accessed 2026-10-28. https://www.nvidia.com/en-us/data-center/nvlink/
  8. NVIDIA Vera Rubin NVL72, NVIDIA, accessed 2026-10-28. https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/
  9. NVIDIA NVLink Fusion, NVIDIA, accessed 2026-10-28. https://www.nvidia.com/en-us/data-center/nvlink-fusion/
  10. NVIDIA GB200 NVL72, NVIDIA, accessed 2026-10-28. https://www.nvidia.com/en-us/data-center/gb200-nvl72/
  11. NVIDIA Vera Rubin Ramps Into Full Production (press release), NVIDIA, 2026-05-31. https://nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory
  12. AMD Helios Rackscale Solution, AMD, accessed 2026-10-28. https://www.amd.com/en/products/rackscale-solutions/helios.html
  13. AAI 2026: AMD Launches AMD Helios Rackscale Solution for Frontier AI, AMD Newsroom, 2026-07-23. https://newsroom.amd.com/news/aai-2026-helios-update/
  14. AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era (press release), AMD, 2026-07-23. https://ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era
  15. HPE accelerates AI deployments with first AMD "Helios" AI rack-scale architecture with open, scale-up networking built with Broadcom (press release), HPE, 2025-12-02. https://www.hpe.com/us/en/newsroom/press-release/2025/12/hpe-accelerates-ai-deployments-with-first-amd-helios-ai-rack-scale-architecture-with-open-scale-up-networking-built-with-broadcom.html
  16. HPE secures its first AMD Helios order in $1.2 billion deal with Vultr (press release), HPE, 2026-09-30. https://www.hpe.com/us/en/newsroom/press-release/2026/09/hpe-secures-its-first-amd-helios-order-in-12-billion-deal-with-vultr.html
  17. Astera Labs Extends Leadership in Open, AI Scale-Up Networking with New 320 Lane Scorpio X-Series Smart Fabric Switch (press release), Astera Labs, 2026-05-05. https://asteralabs.gcs-web.com/news-releases/news-release-details/astera-labs-extends-leadership-open-ai-scale-networking-new-320
  18. Astera Labs (ALAB) Q1 2026 Earnings Call Transcript, The Motley Fool, 2026-05-05. https://www.fool.com/earnings/call-transcripts/2026/05/05/astera-labs-alab-q1-2026-earnings-transcript/
  19. Astera Labs Reports Second Quarter 2026 Financial Results (Form 8-K Exhibit 99.1), Astera Labs via SEC EDGAR, 2026-08-04. https://www.sec.gov/Archives/edgar/data/0001736297/000173629726000033/q226exhibit991.htm
  20. Upscale Introduces Token Fabric (press release), Upscale, 2026-10-08. https://upscaleai.com/blogs/upscale-introduces-token-fabric-the-industrys-most-comprehensive-standards-based-networking-portfolio-for-ai-factories
  21. Marvell Expands Custom Compute Platform with UALink Scale-up Solution for AI Accelerated Infrastructure (press release), Marvell, 2025-06-11. https://investor.marvell.com/news-events/press-releases/detail/94/marvell-expands-custom-compute-platform-with-ualink-scale-up-solution-for-ai-accelerated-infrastructure
  22. Marvell to Acquire XConn Technologies, Expanding Leadership in AI Data Center Connectivity (press release), Marvell, 2026-01-06. https://investor.marvell.com/news-events/press-releases/detail/1004/marvell-to-acquire-xconn-technologies-expanding-leadership-in-ai-data-center-connectivity
  23. UALink IP, Synopsys, accessed 2026-10-28. https://www.synopsys.com/designware-ip/interface-ip/ualink.html
  24. Securing UALink: Introducing Synopsys UALinkSec_200 Security Module, Synopsys, 2026-01-15. https://www.synopsys.com/articles/securing-ualink-security-module.html
  25. Industry's First UALink-200G Controller and PHY Running in 3nm, Cadence Community blog, 2026-04-13. https://community.cadence.com/cadence_blogs_8/b/data-center/posts/industry-first-ualink-200g-controller-and-phy-running-in-3nm
  26. NVIDIA Kicks Off the Next Generation of AI With Rubin — Six New Chips, One Incredible AI Supercomputer (press release), NVIDIA, 2026-01-05. https://nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
  27. NVIDIA Vera Rubin datasheet, NVIDIA, 2026-10-14. https://dam-cdn.nvd.orangelogic.com/AssetLink/56p68o47y6f1yucpubl0ump1c0aif422.pdf
External Sources (26)
[1]ualinkconsortium.org/wp-content/uploads/2025/04/UALink-1.0-Specification-PR_FINAL.pdf
[2]hpe.com/us/en/newsroom/press-release/2025/12/hpe-accelerates-ai-deployments-with-first-amd-helios-ai-rack-scale-architecture-with-open-scale-up-networking-built-with-broadcom.html
[3]nvidia.com/en-us/data-center/nvlink
[4]ualinkconsortium.org/wp-content/uploads/2025/04/UALink-1.0-White_Paper_FINAL.pdf
[5]newsroom.amd.com/news/aai-2026-helios-update
[6]ualinkconsortium.org/wp-content/uploads/2026/04/UALink-2.0-Specification-PR_FINAL.pdf
[7]ualinkconsortium.org/specification
[8]ualinkconsortium.org/blog/ualink-roadmap-insights-accelerating-open-scalable-ai-networking-1296
[9]sec.gov/Archives/edgar/data/0001736297/000173629726000033/q226exhibit991.htm
[10]ualinkconsortium.org/wp-content/uploads/2026/01/UALink_White_Paper_Publication_FINAL_UPDATED.pdf
[11]nvidianews.nvidia.com/news/rubin-platform-ai-supercomputer
[12]nvidia.com/en-us/data-center/gb200-nvl72
[13]nvidianews.nvidia.com/news/vera-rubin-full-production-agentic-ai-factory
[14]nvidia.com/en-us/data-center/nvlink-fusion
[15]amd.com/en/products/rackscale-solutions/helios.html
[16]hpe.com/us/en/newsroom/press-release/2026/09/hpe-secures-its-first-amd-helios-order-in-12-billion-deal-with-vultr.html
[17]ir.amd.com/news-events/press-releases/detail/1294/aai-2026-amd-delivers-full-stack-compute-for-the-agentic-ai-era
[18]fool.com/earnings/call-transcripts/2026/05/05/astera-labs-alab-q1-2026-earnings-transcript
[19]asteralabs.gcs-web.com/news-releases/news-release-details/astera-labs-extends-leadership-open-ai-scale-networking-new-320
[20]upscaleai.com/blogs/upscale-introduces-token-fabric-the-industrys-most-comprehensive-standards-based-networking-portfolio-for-ai-factories
[21]investor.marvell.com/news-events/press-releases/detail/94/marvell-expands-custom-compute-platform-with-ualink-scale-up-solution-for-ai-accelerated-infrastructure
[22]investor.marvell.com/news-events/press-releases/detail/1004/marvell-to-acquire-xconn-technologies-expanding-leadership-in-ai-data-center-connectivity
[23]synopsys.com/designware-ip/interface-ip/ualink.html
[24]synopsys.com/articles/securing-ualink-security-module.html
[25]community.cadence.com/cadence_blogs_8/b/data-center/posts/industry-first-ualink-200g-controller-and-phy-running-in-3nm
[26]nvidia.com/en-us/data-center/vera-rubin-nvl72
ualink vs nvlinkualink 200g 1.0 specificationualink 1024 acceleratorsnvlink fusion vs ualinkscale up vs scale out ai fabricualink switch availability

About

GPUSmith

Plan and build private AI compute with GPU Smith. We connect workload requirements with GPU hardware, networking, deployment and operating decisions so your infrastructure fits the work it must perform.

GPU Smith provides independent engineering and research for private AI infrastructure. We help technical buyers and operators reason about GPU workload sizing, hardware selection, procurement, deployment and inference operations, from the compute system to the networking, power and cooling requirements around it.

Start with the workload

Useful infrastructure decisions begin with the models, data, throughput, latency and operating constraints a team actually has. GPU Smith helps connect those requirements with system design and deployment choices. The goal is a privately controlled AI environment whose capacity and operational demands are understood before a purchasing decision.

Hardware and supplier research

Our public reference library covers GPUs, complete systems and networking components. The server-vendor directory and manufacturing research help teams investigate suppliers, compare options and follow sources behind technical claims. We distinguish manufacturer specifications from measured benchmarks and advertised capabilities from tested configurations. Reference pages are research resources, not a live inventory listing or a binding equipment quote.

Deployment and operations

GPU Smith's engineering scope includes infrastructure integration, networking, power, cooling and the practical operation of inference workloads. Our published articles explain the tradeoffs behind deployment, procurement and ongoing operations so teams can ask better questions and document decisions.

Work with GPU Smith

Explore the hardware library, GPU references, networking references and vendor research. Contact GPU Smith with your workloads and deployment constraints to discuss a project and confirm current scope, pricing and availability.

Inclusion of a manufacturer or product in our research does not imply a partnership, certification or endorsement.

Disclaimer

This document is provided for informational purposes only. No representations or warranties are made regarding the accuracy, completeness, or reliability of its contents. Any use of this information is at your own risk. GPUSmith shall not be liable for any damages arising from the use of this document. This content was generated with assistance from artificial intelligence tools, which may contain errors or inaccuracies. Readers should verify critical information independently. All product names, trademarks, and registered trademarks mentioned are property of their respective owners and are used for identification purposes only. Use of these names does not imply endorsement. This document does not constitute professional or legal advice. For specific guidance related to your needs, please consult qualified professionals.