Datacom & AI Compute
Goilips 800G/1.6T Ultra-Low-Latency Interconnect for AI Compute Clusters
Large-scale AI model training clusters with tens of thousands of GPUs, high-concurrency networks for GPU servers (e.g. H100, B200).
Large-scale AI model training clusters with tens of thousands of GPUs, high-concurrency networks for GPU servers (e.g. H100, B200).
With the rapid advancement of trillion-parameter large language model (LLM) training and generative AI, GPU cluster scale has moved into the tens-of-thousands (10K+ GPUs) and even hundreds-of-thousands range. In AI distributed training, frequent inter-node parameter synchronization and gradient reduction (All-Reduce/All-To-All) mean that network communication efficiency directly determines the utilization (MFU) of extremely expensive compute resources.
The Goilips 800G/1.6T Ultra-Low-Latency Interconnect for AI Compute Clusters is purpose-built for next-generation, hyperscale AI compute centers. It uses Goilips' innovative LPO (Linear-drive Pluggable Optics) and Silicon Photonics architecture to reinvent data transmission at the physical layer, overcoming the core bottlenecks of high power consumption, high latency, and high heat output in traditional optical modules — helping customers build a compute foundation that is “high-throughput, non-blocking, and low-power.”
For AI training infrastructure at 10K+ GPU scale that requires weeks of uninterrupted, high-intensity parallel computation.
Perfectly suited for high-concurrency RoCE v2 or InfiniBand networks built on NVIDIA H100 / H200 / B200 / GB200 and various domestic high-performance AI chips.
For modern data centers with extremely high rack power density and stringent power/thermal requirements for optical modules.
In a typical non-blocking Fat-Tree or Spine-Leaf network topology, Goilips provides precisely matched optical modules for each layer:
Core Deployment：Goilips 800G QSFP-DD / OSFP LPO modules are deployed end-to-end.
Advantage：Eliminates the DSP (Digital Signal Processing) chip inside traditional modules entirely, using the powerful equalization capability of the switch and NIC ASICs to drive the optical signal directly — reducing latency to the picosecond/nanosecond range and removing nearly half of the heat sources.
Core Deployment：Goilips 800G/1.6T DR8 / 2xFR4 Silicon Photonics modules are deployed.
Advantage：Uses a highly integrated silicon photonics chip process in place of traditional discrete EML lasers, delivering excellent signal integrity and single-lane 200G scalability to ensure lossless, ultra-high-throughput forwarding at the backbone layer.
In traditional DSP-based solutions, signal error correction and reconstruction (FEC) introduces additional microsecond-level latency. Goilips LPO modules use a purely linear analog signal transmission architecture that eliminates the signal-processing latency introduced by a DSP, bringing physical-layer latency down to the picosecond range. Over long parallel training runs, this meaningfully reduces GPU Wait Time and significantly improves cluster MFU (Model Flops Utilization).
The power draw of 800G/1.6T pluggable optical modules has always been the “thermal demon” of the data hall. By removing the power-hungry DSP, Goilips 800G LPO modules cut per-module power consumption to below 12W-14W (roughly a 40% reduction versus the 22W-30W of traditional DSP-based solutions). Across a 10K-GPU facility deploying tens of thousands of modules, this saves tens of kilowatts of power and greatly eases hot-aisle congestion at the rack.
Nothing derails large model training more than a mid-run network failure forcing a restart. Goilips' 1.6T series introduces proprietary wafer-scale silicon photonic integration technology, reducing discrete optical components by more than 80% and significantly improving the modules' shock and heat resistance. Module MTBF (Mean Time Between Failures) is improved by over 30%, safeguarding training runs that can last for months.
Goilips' full range of 800G/1.6T modules strictly follows IEEE 802.3 and OIF industry standards, and has been thoroughly compatibility-tested with switches from mainstream vendors (Arista, Cisco, NVIDIA Quantum/Spectrum, H3C, Huawei, and others) and high-spec smart NICs. Plug-and-play with no need to re-architect existing networks, significantly lowering the total cost of ownership (TCO) for building and operating a compute center.
Simplify your network deployment and reduce complexity. Connect with our engineers for a solution design tailored to your project.
Expert customer service on call, 5x24 phone support.
For general questions, email us anytime.
Our live agent is ready to chat and help in real time.