
Shangxuan Xinzhi proposes MRAM-based architecture for AI inference
The AMW Read
The proposed architecture adds an early-stage inference chip design, but its impact depends on fabricated hardware results.
Shangxuan Xinzhi proposes MRAM-based architecture for AI inference
Chinese chip designer Shangxuan Xinzhi has introduced MITA, an MRAM in-memory tensor architecture, and defined an MRAM-based processing unit called MPU. The company says its first planned chip will target feed-forward network calculations in cloud-based large language model inference. Founder Lei Kun told DoNews that the design is intended to handle selected matrix operations alongside existing GPUs or AI accelerators.
The proposal addresses a central inference constraint: moving model weights between memory and compute units can add latency, energy use, and system cost. Shangxuan Xinzhi proposes storing weights in nonvolatile MRAM and performing matrix calculations either in the memory array or nearby. The company says MRAM arrays can offer greater density than SRAM, but these advantages remain design claims for its proposed system. The architecture adds another approach to the competition among inference hardware designs that seek to reduce data movement.
For builders and investors, the key test is whether a fabricated chip can deliver useful power, performance, and area results within an existing accelerator system. Lei said prospective customers need those results before buying, and that chip iterations take time. The company plans its first tape-out for March 2027 and expects the chip back in June. It has not reported results from that chip or disclosed an amount for its nearly completed third funding round.