Enflame Cloud Blazer POD P2010

System Details

Interconnect

Two 200Gb/s links between compute nodes, as filed. The prospectus describes the cabinet as four to eight AI servers and several network switches joined by high-speed cabling; it gives no scale-up fabric for the accelerators.

Notes

Enflame's single-cabinet training appliance, branded Cloud Blazer POD and described in the company's STAR Market prospectus as a single-cabinet, single-compute-node device that forms the smallest unit of a compute cluster. It integrates 32 Yunsui T20 accelerators across four to eight AI servers, each server carrying eight cards. The filing lists it for training, air-cooled, with a PUE of 1.30 or better for the inference model and 1.25 or better for this one. Its 1TB of accelerator memory matches 32 cards at 32 GB exactly. Enflame publishes no throughput figure for this product or for the cards inside it.