MetaX C600

MetaX C600 (曦云C600) is MetaX's second-generation training and inference GPU and the first built end to end on a domestic Chinese supply chain, covering design, manufacture, packaging and test. It is an OAM 2.0 module with a maximum board power of 1000W, supporting both air and liquid cooling, and MetaX describes its memory only qualitatively, as an ultra-large-capacity high-bandwidth VRAM with full end-to-end ECC protection, alongside built-in ECC/RAS safety blocks aimed at finance and government workloads. A capacity of 144 GB of HBM3e is widely reported and was previously recorded here, but MetaX's product page states no capacity figure and no bandwidth figure at all, so neither is published in the specification fields. The C600 is built on MetaX's XCORE1.5 instruction set, which doubles the FP16 and BF16 tensor matrix rate over the previous generation and adds FP8 and BF8 tensor instructions plus tensor transpose and FP4 conversion instructions for mixed-precision inference. FP8 compute support is the headline architectural change over the C500. Scale-up runs over MetaXLink, with a flexible MetaXLink-E Ethernet interface reaching supernodes of up to 128 GPUs, and the MXMACA software stack provides compatibility with mainstream GPU ecosystems and frameworks such as PyTorch and TensorFlow. MetaX has deliberately not published throughput figures at any precision, stating only that single-card compute leads the NVIDIA H20, so this entry carries no FLOPS values rather than repeat the unverified numbers that circulate online. MetaX has also disclosed that the move to a domestic process cost it slightly on card-to-card interconnect bandwidth and cost it power versus the C500 and C550. The chip itself is named MXC600 and also underpins the N300 and X300 product families; it passed China's national security and reliability certification in May 2026. MetaX benchmarks the following C700 generation, not the C600, against the NVIDIA H100.