800GbE AI NIC comparison
Broadcom Thor Ultra vs NVIDIA ConnectX-9: 800G AI NICs
Broadcom Thor Ultra and NVIDIA ConnectX-9 target the same difficult problem: moving data between accelerators at up to 800GbE without turning the NIC into the bottleneck. Both use PCIe Gen6 x16-class host connectivity, but their ecosystem strategies differ. Thor Ultra emphasizes UEC-compliant open Ethernet, packet-level multipathing and advanced RoCE, while ConnectX-9 is the SuperNIC endpoint for NVIDIA Spectrum-X Ethernet and Quantum-X InfiniBand.
Interactive calculator
Thor Ultra vs ConnectX-9 Deployment Comparator
Enter your own topology and bandwidth assumptions. These planning calculators estimate raw links, capacity and ratios; they do not certify fabric goodput, rail mapping, cable reach, firmware interoperability, electrical design, cooling or OEM compatibility.
Live Amazon supporting hardware
Current supporting components and Price Options
Compare current Amazon listings relevant to this guide. Product availability and prices can change.
Quick answer
Which 800G NIC is the better fit?
Choose based on fabric architecture, not the shared 800G headline. Thor Ultra is Broadcom’s UEC-oriented Ethernet NIC with packet-level multipathing, out-of-order placement, selective retransmission and programmable congestion control. ConnectX-9 supports up to 800GbE and XDR InfiniBand, PCIe Gen6 x16 and tight integration with Spectrum-X, Quantum-X and Vera Rubin.
The surrounding switch and software stack should drive the decision
An operator building an open UEC Ethernet environment may value Thor Ultra’s interoperability and advanced RoCE design. A Vera Rubin or NVIDIA end-to-end fabric has a stronger ConnectX-9 integration path. Compare supported protocols, firmware maturity, host form factor, congestion behavior, management tooling and qualification with the exact XPU and switch.
Thor Ultra vs ConnectX-9
The comparison separates confirmed public capabilities from ecosystem positioning.
| Feature | Broadcom Thor Ultra | NVIDIA ConnectX-9 | Why it matters |
|---|---|---|---|
| Peak Ethernet | 800GbE | Up to 800GbE per port | Equal headline link class |
| Host interface | PCIe Gen6 x16 | PCIe Gen6 x16 | Required for full host bandwidth |
| Ethernet transport | Advanced RoCE / UEC compliant | Spectrum-X Ethernet / RoCE acceleration | Fabric software differs |
| Multipathing | Packet-level multipathing | Spectrum-X plane-aware load balancing in supported design | Path selection architecture matters |
| InfiniBand | No InfiniBand positioning | Up to XDR InfiniBand | Important for dual-protocol environments |
| Ecosystem emphasis | Open UEC switch/XPU interoperability | NVIDIA Spectrum-X / Quantum-X / Vera Rubin | Qualification can outweigh raw specs |
Before you use the result for procurement
Draw the physical topology
Map every server-facing port, leaf uplink, spine link and network plane. Aggregate bandwidth alone can hide impossible port or lane assumptions.
Verify exact endpoints
Confirm NIC form factor, PCIe generation, host lane budget, port speed, connector and firmware support on the exact server platform.
Qualify optics and cables
Match OSFP/QSFP form factor, lane rate, breakout, reach, fiber type and both endpoint qualification lists. Do not treat equal headline speed as automatic compatibility.
Test failure and congestion behavior
Validate oversubscription, ECMP or multiplane path behavior, switch failure domains and recovery under the traffic patterns the AI workload will actually generate.
Both NICs target 800G AI fabrics
Thor Ultra and ConnectX-9 are designed for AI and HPC environments where a 100G or 200G host link can become the limiting factor. Each reaches the 800GbE class and uses a PCIe Gen6 x16 host interface, so the basic bandwidth conversation begins at a similar point.
That similarity makes ecosystem features more important. Buyers should compare how each NIC handles congestion, path diversity, memory placement, telemetry and switch integration rather than stopping at the line rate printed on the product page.
Thor Ultra is built around UEC-compliant Ethernet
Broadcom introduced Thor Ultra as an 800G AI Ethernet NIC fully feature compliant with the Ultra Ethernet Consortium specification. The company emphasizes open interoperability across XPUs, optics and switches and positions the device as an alternative to vertically integrated fabrics.
UEC compliance is valuable when an operator wants a standards-led multi-vendor environment, but a label should still be backed by interoperability testing. Validate the exact switch OS, optic, XPU peer-memory path and congestion configuration intended for production.
ConnectX-9 spans Ethernet and InfiniBand
NVIDIA ConnectX-9 supports 800/400/200/100 Gb/s Ethernet on the high-end OSFP model and also supports XDR/NDR-class InfiniBand. That dual-protocol capability makes it relevant to organizations that operate both Spectrum-X Ethernet and Quantum-X InfiniBand fabrics.
Thor Ultra is positioned as Ethernet. If InfiniBand is a real requirement, ConnectX-9 has a structural advantage before performance testing begins. If the deployment is definitively Ethernet-only, the comparison can focus more narrowly on UEC/RoCE behavior and software integration.
Packet-level multipathing is central to Thor Ultra
Broadcom lists packet-level multipathing as a core advanced-RDMA feature. The goal is to distribute traffic efficiently across available paths instead of pinning a large transfer to one route and waiting for congestion to clear.
Multipathing is particularly valuable in AI because synchronized flows can create hotspots that are much larger than typical enterprise traffic. The important benchmark is not whether packets use several paths, but whether the technique improves job completion time without creating excessive reordering or control overhead.
Out-of-order placement changes the receiver model
Thor Ultra supports out-of-order packet delivery directly to XPU memory. Traditional reliable transport often waits for missing packets before placing later data, which can waste available network capacity when paths have different latency.
Out-of-order placement must be understood together with selective retransmission and the receiving software stack. The promise is better utilization under multipath conditions, but production value depends on the exact XPU memory and communication framework integration.
Selective retransmission reduces recovery work
Broadcom also highlights hardware-based selective retransmission, allowing the NIC to resend missing data rather than replaying a much larger sequence after loss. In large fabrics, avoiding unnecessary retransmitted bytes can protect congestion headroom during fault or packet-loss events.
The best validation combines controlled packet loss with application-level measurements. Record useful throughput, tail latency and job progress while introducing loss or trimming. A transport feature is successful when the XPU spends less time waiting, not merely when the NIC counters look efficient.
ConnectX-9 is tightly tied to Spectrum-X Multiplane
NVIDIA’s latest Multiplane architecture places the plane-aware load-balancing function in the ConnectX SuperNIC. Each plane keeps independent congestion state, and the NIC filters congested or failed planes before selecting an egress path.
This integration is a meaningful advantage inside an NVIDIA reference architecture because switch and endpoint behavior are designed together. It can be less decisive in a multi-vendor UEC fabric where the operator prefers independent components and standards-based behavior.
PCIe Gen6 x16 is necessary but not sufficient
An 800G network link can exceed the practical headroom of older host interfaces once protocol overhead and bidirectional traffic are considered. Both families therefore move to PCIe Gen6 x16-class connectivity.
System qualification still matters. ConnectX-9 documentation explicitly warns about data-center power and airflow requirements, and any 800G NIC needs an appropriate slot, cooling path and BIOS configuration. A server with the connector but inadequate thermal design is not a valid host.
Form factor can decide the procurement before protocol does
Broadcom lists standard PCIe CEM and OCP 3.0 form factors for Thor Ultra. NVIDIA offers ConnectX-9 in PCIe and OCP variants with different port configurations. The selected server chassis may only accept one of those mechanical and cooling designs.
Create a host compatibility matrix covering slot type, lane width, bracket, airflow, power, cable exit and firmware. High-speed NIC comparisons are often lost on mechanical details that would have been obvious if the physical server had been checked first.
Encryption and platform security belong in the comparison
Thor Ultra includes line-rate encryption/decryption with PSP offload, secure boot, signed firmware and device attestation. ConnectX-9 also includes secure-boot and crypto-enabled variants within NVIDIA’s data-center platform.
Security features should be evaluated against the organization’s threat model and key-management workflow. Check which functions are available in the exact SKU and firmware rather than comparing broad family claims.
Switch choice determines how much ecosystem value is unlocked
A Thor Ultra NIC paired with a UEC-compliant Broadcom or third-party switch can benefit from the open Ethernet design Broadcom promotes. ConnectX-9 paired with Spectrum-X can use NVIDIA’s coordinated congestion, routing and Multiplane behavior.
Cross-vendor combinations may work at the Ethernet protocol level, but feature depth can change. If the procurement strategy intentionally mixes vendors, require a proof of concept that covers the features being relied on rather than assuming full parity from standards compliance.
Benchmark the same workload and failure conditions
Vendor feature lists are not a substitute for application measurements. Test the same GPU/XPU count, model communication pattern, message sizes, transport settings and switch topology on both NIC options where possible.
Include congestion and failure cases. The winning adapter is the one that keeps accelerators productive, recovers predictably and integrates cleanly with operations at an acceptable total cost, not the one with the longest list of transport acronyms.
Methodology and sources
Cloudzat compares only capabilities explicitly documented by Broadcom and NVIDIA, then separates ecosystem fit from raw specifications. The interactive score is a deployment heuristic and does not claim benchmark superiority for either adapter. Availability and firmware feature status must be checked at purchase time.
- Broadcom Thor Ultra 800G AI Ethernet NIC announcement
- Broadcom P1800GO 800GbE PCIe NIC
- Broadcom BCM57708 800Gb/s NIC product brief
- NVIDIA ConnectX-9 SuperNIC specifications
- NVIDIA ConnectX-9 SuperNIC introduction
As an Amazon Associate, Cloudzat may earn from qualifying purchases. Marketplace listings on these pages are supporting networking hardware such as NICs, switches, optics and high-speed cables. A marketplace row is not represented as a Spectrum-6 switch, ConnectX-9 SuperNIC, Thor Ultra NIC or qualified NVIDIA fabric unless the exact listing evidence supports that identity. Verify model, speed, connector, firmware, warranty and OEM qualification before purchase.
Frequently asked questions
Are Thor Ultra and ConnectX-9 both 800GbE NICs?
Yes. Broadcom positions Thor Ultra at 800GbE, and NVIDIA ConnectX-9 supports up to 800GbE on its high-end models.
Do both use PCIe Gen6 x16?
Yes. Broadcom specifies PCIe Gen6 x16 for Thor Ultra, and NVIDIA documents PCIe Gen6 x16 on ConnectX-9.
Which NIC supports InfiniBand?
ConnectX-9 supports Ethernet and InfiniBand, including XDR-class InfiniBand. Thor Ultra is an Ethernet/UEC product.
What is special about Thor Ultra RDMA?
Broadcom highlights packet-level multipathing, out-of-order data placement, selective retransmission and programmable sender/receiver congestion control.
What is special about ConnectX-9 in Spectrum-X?
ConnectX-9 is the SuperNIC endpoint for the Vera Rubin Spectrum-X generation and participates in plane-aware routing and congestion behavior.
Is Thor Ultra already broadly available?
Availability can change. Broadcom originally announced sampling, and current product pages should be checked for the exact adapter and distributor status before procurement.
Can Thor Ultra work with non-Broadcom switches?
Broadcom emphasizes UEC-compliant open interoperability, but exact optics, firmware and feature behavior should still be tested with the selected switch.
Can ConnectX-9 work with ordinary Ethernet switches?
NVIDIA documents Ethernet interoperability across supported rates, but the full Spectrum-X feature set depends on compatible NVIDIA platform components.
Which is better for Vera Rubin?
ConnectX-9 has the direct NVIDIA Vera Rubin and Spectrum-X integration path.
Which is better for an open UEC fabric?
Thor Ultra is explicitly positioned for UEC-compliant open Ethernet. Benchmark it against other qualified NICs in the exact multi-vendor environment.