A domestic thousand-GPU cluster has been live and stable for months, continuously carrying real traffic and serving real users.
Project information
Environment, workload, and validation approach.
Background and challenges
The project targeted a thousand-GPU production cluster at a domestic accelerator company. The goal was not only to complete test validation, but also to keep the network carrying real workloads long term and serving users continuously.
Results
This case presents project value with production facts and does not package operating results that lack a unified comparison baseline as performance percentages.
The domestic compute cluster went live.
Continuously carried workloads and ran stably for months.
Continuously served real users.
Formed a production loop from cluster delivery to user feedback.
Value summary
ZCube moved from test validation to production validation, forming a loop of sustainable operation, real service, and customer feedback on a domestic thousand-GPU cluster.
Customer case studies
See more ZCube validation and production practice across compute environments, workloads, and project stages.
Using a coding inference service and comparing against traditional ROFT, this case validates that ZCube unlocks more effective inference compute through a flat fabric and more balanced network communication.
15%
Average GPU inference throughput increase
40.6%
Time to first token (TTFT) P99 reduction
33%
Network hardware cost reduction
In a domestic AI computing ecosystem, compare ZCube with traditional Clos across collective communication, training, and inference.
14.5%
All-to-All throughput increase
22%–30%
TTFT P99 reduction
7%–10%
Inference throughput increase
Compare ZCube with Clos and validate performance gains in collective communication, model training, and multi-tenant end-to-end inference.
119%
Collective communication performance increase
20.1%
Model-training throughput increase
11.5%
Inference latency reduction
Whether you are building a new cluster, upgrading an existing network, tuning performance, or deploying intelligent operations, we can assess how the network can unlock more effective compute.