The RecAccel, can perform 500,000 inferences per second, the company said and when running the PyTorch deep learning recommendation model (DLRM) software outperforms server class CPU and inference GPU by factors of 28x and 65x, respectively.
The engine includes a memory subsystem for embedding table lookup and a parallel compute FPGA area for neural network inferencing. When mounted on PCIe Gen3 host interface the RecAccel is suitable for data center deployment.
"Fast and accurate recommendation inference is the key to e-commerce business success," said Youn-Long Lin, CEO of Neuchips, in a statement.
Related links and articles: