Huawei Ascend 310

The Huawei Ascend 310, codenamed Ascend-Mini, was the first Ascend processor to ship and the chip behind the entire Atlas 200 and Atlas 500 edge line. Announced by Eric Xu at Huawei Connect 2018, it pairs two Da Vinci AI cores with eight Arm Cortex-A55 host cores on a 12nm die of 104.4 square millimetres, and delivers 8 TFLOPS of FP16 or 16 TOPS of INT8 inside an 8 watt envelope, which is the efficiency point the whole design was built around. Alongside the AI cores it carries a substantial fixed-function media block, 16 channels of H.264 and H.265 decode plus one channel of encode, and an 8 MB on-chip buffer feeding the matrix units. Memory is 128-bit LPDDR4X, but capacity and bandwidth are deliberately left blank here because they are a board-level choice rather than a chip specification: Huawei's own Atlas 200 module ships the same silicon with either 4 GB or 8 GB. The Da Vinci Cube in this generation implements 4,096 FP16 MACs and 8,192 INT8 MACs and has no FP32 or BF16 matrix path at all, which is why this part carries only FP16 and INT8 figures.