Enflame L600

The L600 is Enflame's fourth-generation accelerator and its first part aimed at training rather than inference. It is an OAM module carrying one 邃思 400 chip, rated at 700 W and built on the fourth generation of Enflame's own GCU-CARA compute unit, which the company describes as a deliberate refusal of the GPGPU model: its own instruction set, its own accelerated compute unit answering the Tensor Core, and its own GCU-LARE fabric answering NVLink. The generation's distinguishing feature is native FP8, which Enflame says makes it one of the few Chinese vendors supporting the datatype in hardware rather than by conversion. Up to 128 modules connect in a single-layer all-to-all topology, and Enflame builds two systems on it: OGX400, an eight-module node released in July 2025, and ESL32/64, a 32 or 64 module rack-scale superpod still waiting on volume production. Enflame publishes no throughput figure for the part, in any document. Its IPO prospectus prints power, FP16 throughput, memory type, capacity, bandwidth and interconnect bandwidth for six NVIDIA and AMD parts on one page, and gives the L600 a single merged cell reading only that it leads the domestic field for training and inference cards. As of the prospectus signing the chip had come back from fab but was not yet in volume delivery.