Cambricon MLU290-M5

The Cambricon (寒武纪) MLU290-M5 is the accelerator card built around 思元290 (SiYuan 290), Cambricon's first training chip. Cambricon states it is fabricated on TSMC 7 nm with 46 billion transistors on the MLUv02 extended architecture, and that the card uses the open OAM module form factor with 64 MLU Cores, 1.23 TB/s of memory bandwidth and the MLU-Link multi-chip interconnect, delivering up to 512 TOPS of 8-bit fixed-point AI performance at a 350 W maximum thermal power. Cambricon compares the chip with its predecessor 思元270 at four times the peak compute, twelve times the memory bandwidth and nineteen times the chip-to-chip bandwidth. The card is the compute element of the 玄思1000 appliance, which holds four of them. Per-card memory capacity is not listed here because Cambricon publishes only a 128 GB total for the four-card chassis, and no floating-point throughput is listed because Cambricon publishes none, though the platform does support FP32 and FP16.