A leading operator in Singapore has deployed Hermesys HydraX™ 800G, 400G and 100G optical transceivers in its AI computing data centre. It is the first time the operator has used third-party compatible optics in a core data centre project. Before deployment, our 800G OSFP module went through a full compatibility and stress test on NVIDIA InfiniBand equipment. This article explains how it was tested and what the results showed.
The Project: Optics for a Singapore AI Data Centre
AI training clusters move huge volumes of data between GPU servers, so every link in the fabric needs high bandwidth, low error rates and long-term stability. The operator needed 800G, 400G and 100G optics for its AI computing data centre. The modules had to work with the NVIDIA switches and network adapters already chosen for the cluster.
Original-brand optics are a large share of an AI cluster's cost, and they can have long lead times. Third-party compatible transceivers are a practical alternative, but only when they are proven to behave exactly like the original parts on the target hardware. That is why the operator asked for a structured test on NVIDIA equipment before approving HydraX™ optical transceivers for its core project.
Figure 1. HydraX™ 800G OSFP transceiver in flat top and finned top versions
The Module Under Test: 800G OSFP 2xSR4
The module tested is a twin-port 800Gb/s OSFP transceiver built to NVIDIA's specification for InfiniBand and Ethernet. It is a hot-swappable multimode module that uses 100G PAM4 signalling and VCSEL lasers, supporting 800G interconnects over distances of up to 100 metres. In this test, it was paired with a 400G OSFP SR4 module on the adapter side over a 50-metre link.
| Item | Switch side | Adapter side |
|---|---|---|
| Test equipment | NVIDIA QM9700 switch | NVIDIA MCX75310AAS-NEAT adapter (ConnectX-7) |
| Module model | OSFP-800G-2xSR4H | OSFP-400G-SR4H |
| Module type | OSFP (800G 2xSR4) | IB NDR, 400G-SR4 |
| Wavelength | Not applicable | 854 nm |
| Test link distance | 50 m | 50 m |
| Port and link rate | IB1/17/1, NDR 400 Gbps | mlx5_0, 400 Gbps |
How We Tested Compatibility on NVIDIA Equipment
The test plan covers three areas: hardware compatibility, software compatibility and performance. Each item has a clear pass criterion, so the results can be checked against the same standard the operator uses for original-brand optics.
Figure 2. Physical connection of the test setup
| Category | Test | Pass criterion |
|---|---|---|
| Hardware | Plug and unplug | After more than five insertions, the module still fits securely in the switch and adapter ports, without becoming loose or stuck. |
| Hardware | Port restart | After restarting the switch or adapter by command, the link is restored within 3 minutes. |
| Hardware | Power off and restart | After a power cycle, the link is restored within 5 minutes. |
| Software | Connectivity | The module links up with devices at both ends and the port status shows "up". |
| Software | Shutdown / no shutdown | After a port is manually shut down and reopened, the link reconnects normally. |
| Software | Parameter check | Part number, version and serial number read by the device match the module label. |
| Performance | DDM monitoring | None of the five Digital Diagnostic Monitoring (DDM) parameters exceed their thresholds. |
| Performance | Bit error rate | Real-time bit error rate is measured before and after Forward Error Correction (FEC). |
| Performance | Stability | 7 days of continuous transmission with no link interruption and no speed reduction. |
Test Results
Hardware compatibility
The module was inserted into and removed from both the switch port and the adapter port five times. In every cycle it seated firmly, showed no looseness, and released smoothly using the module latch.
| Check | 1 | 2 | 3 | 4 | 5 |
|---|---|---|---|---|---|
| Insertion into switch port | Pass | Pass | Pass | Pass | Pass |
| Stability in switch port | Pass | Pass | Pass | Pass | Pass |
| Removal from switch | Pass | Pass | Pass | Pass | Pass |
| Insertion into adapter port | Pass | Pass | Pass | Pass | Pass |
| Stability in adapter port | Pass | Pass | Pass | Pass | Pass |
| Removal from adapter | Pass | Pass | Pass | Pass | Pass |
After hot and cold reboots of the switch and the adapter, and after a full system restart, the link recovered within 1 minute, well inside the 3- and 5-minute criteria.
Software compatibility
The module linked up normally with both the NVIDIA QM9700 switch and the MCX75310AAS-NEAT adapter. The port LEDs on both devices stayed solid green, and both ports reported an active 400 Gbps link.
| Item | NVIDIA QM9700 | NVIDIA MCX75310AAS-NEAT |
|---|---|---|
| Port number | IB1/17/1, IB1/17/2 | mlx5_0 |
| Port status | Active | Active |
| Port link rate | NDR 400 Gbps | 400 Gbps |
In the shutdown / no shutdown test, the link recovered within 1 minute after the switch and adapter ports were disabled and re-enabled. The model name, serial number and other information read by both devices matched the product label. Transmission distance, wavelength and module type also matched the datasheet.
Performance
DDM. Across three readings on both devices, no DDM alarm was raised and all values stayed normal. Readings fluctuated by less than 10% between tests.
| DDM parameter | NVIDIA QM9700 | NVIDIA MCX75310AAS-NEAT |
|---|---|---|
| DDM alert | No alarm | No alarm |
| Temperature | Normal | Normal |
| Voltage | Normal | Normal |
| Bias current | Normal | Normal |
| Tx optical power | Normal | Normal |
| Rx optical power | Normal | Normal |
Bit error rate. The pre-FEC bit error rate stayed below 1E-5 on both devices. After FEC, no errors were measured and no error packets were detected on the ports.
| Item | NVIDIA QM9700 | NVIDIA MCX75310AAS-NEAT |
|---|---|---|
| Port number | IB1/17/1 | mlx5_0 |
| BER before FEC | 5E-9 | 1E-7 |
| BER after FEC | No errors | No errors |
Stability. The module ran at full speed for 7 days, 12 hours and 43 minutes on both the switch and the adapter, with no link interruptions and no speed reductions.
Results at a Glance
| Test | Criterion | Result |
|---|---|---|
| Plug and unplug | 5+ cycles, no looseness or jamming | Pass (5/5) |
| Port restart | Link back within 3 min | Pass (< 1 min) |
| Power off and restart | Link back within 5 min | Pass (< 1 min) |
| Connectivity | Port status "up" at both ends | Pass (400 Gbps) |
| Shutdown / no shutdown | Link reconnects normally | Pass (< 1 min) |
| Parameter check | Information matches label | Pass |
| DDM monitoring | No parameter over threshold | Pass (< 10% fluctuation) |
| Bit error rate | No errors after FEC | Pass |
| Stability | 7 days without interruption | Pass (7 d 12 h 43 min) |
The HydraX™ 800G OSFP module passed every hardware compatibility, software compatibility and performance test on NVIDIA equipment, demonstrating full compatibility with the operator's AI network.
In Production at a Singapore AI Data Centre
Following the tests, Hermesys HydraX™ 800G, 400G and 100G optical transceivers were supplied to the operator's AI computing data centre and passed all on-site stress tests. They are now in production use alongside the latest NVIDIA networking equipment, including the NVIDIA SN5600 switch.
Figure 3. Hermesys 800G OSFP transceivers working with an NVIDIA SN5600 switch at one of Singapore's leading data centres
Figure 4. HydraX™ high-speed optical transceivers deployed in a local data centre in Singapore
For the operator, the project shows that properly tested compatible optics can meet the reliability standard of a core AI data centre. Hermesys will keep investing in testing and product development so that customers can scale their AI infrastructure with confidence.
Frequently Asked Questions
Are third-party 800G OSFP transceivers compatible with NVIDIA switches?
They can be, provided they are built to NVIDIA's specification and tested on the target hardware. The HydraX™ 800G OSFP module in this article passed hardware, software and performance tests on an NVIDIA QM9700 switch and a ConnectX-7 adapter.
What is the difference between 800G OSFP 2xSR4 and SR8?
Both carry 800G over multimode fibre using eight 100G PAM4 lanes. A 2xSR4 module presents the eight lanes as two independent 400G ports, so one switch port can connect to two 400G adapters. This is the twin-port design used in NVIDIA InfiniBand NDR networks.
How long should a transceiver be stress-tested before deployment?
For core data centre links, a multi-day continuous run at full speed is a sensible baseline. In this project, the pass criterion was 7 days without interruption, and the module ran for 7 days, 12 hours and 43 minutes without a single link drop.
Need Tested Optics for Your AI Data Centre?
Hermesys supplies HydraX™ 800G, 400G and 100G optical transceivers, tested for compatibility with mainstream switches and network adapters, along with the high-speed cables and fibre cabling to connect them.
Contact Our Team View 800G OSFP Transceiver