Moffett S40 Compute Card is built for AI inference applications in the data center. As a general-purpose deep learning inference gas pedal, the form factor is in dual-slot, full-height, full-length PCIe x16 gold finger physical form factor. Moffett S40 Compute Card supports 80GB of LPDDR4x memory, with a theoretical peak memory bandwidth of 256GB/s, and a maximum power consumption of 300W.The passively cooled board design keeps the compute card operating within its thermal limitations and through system airflow.

Moffett S40 Compute Card is built on Moffett AI Antoum^®️ architecture. The architecture is designed to emphasize balanced structured sparsity through tight integration of hardware and software, supporting high sparsity rates of up to 32x. Based on Antoum^®️ architecture, Moffett S40 compute card supports BF16 and INT8 computation. At the same time, Moffett S40 Compute Card supports end-to-end software solutions including software toolchain, compiler and runtime with integrated model sparser, ensuring that mainstream AI inference jobs can be quickly realized.

The tightly integrated hardware and software design makes Antoum^®️ an efficient AI system-on-chip processor. In addition, Moffett S40 Compute Card supports hardware video codecs and JPEG decoders, enabling it to handle a variety of video and image application scenarios. Also, Moffett S40 Compute Card ships with the device with ECC enabled for the system DDR to prevent detectable errors in memory.