To meet its goal of AI leadership, China must break its dependence on Nvidia. The BAT companies—Baidu, Alibaba, and Tencent—play an important part in this plan. Rather than follow the others in developing its own AI chips, Tencent is backing a domestic startup. Based in Shanghai, Enflame (燧原) Technology closed a 1.8-billion-yuan Series C round in January 2021, bringing its total funding to a sizable 3.1 billion yuan ($462 million). It now offers a series of AI accelerators for both training and inference.

Enflame sampled its first chip, the CloudBlazer (云燧) T10, in November 2019 and achieved production the following quarter. It sells the T10 on a standard dual-slot PCIe card, and it offers a version called the T11 in an OCP-compliant OAM module. Both provide about 20Tflop/s for FP32 data and target AI training. The company recently announced the CloudBlazer I10, a PCIe card for inference that’s slated for 1Q21 production.

The CloudBlazer T10 features 32 cores that Enflame calls scalable intelligent processors (SIPs). The cores implement a proprietary instruction set and support a range of data types including FP32, FP16, Bfloat16, INT32, INT16, and INT8. GlobalFoundries manufactures the chip in its 12nm process and handles the 2.5D packaging that combines the accelerator with HBM2 memory.