Memory subsystem solution for next-generation AI training chip

“Artificial intelligence training requires far greater memory bandwidth than standard compute applications. Rambus’ proven HBM2 memory subsystem IP delivers the ultra-high bandwidth performance our AI chips need,” explained Arthur Zhang, COO of Enflame Technology. “Using Rambus interface IP cores, we are revolutionising what’s possible in AI technology.”

Optimised for low-latency and high-bandwidth memory applications, the Rambus HBM2 interface solution delivers maximum performance and flexibility in minimal form factor and power envelope. The comprehensive HBM2 subsystem solution of PHY and memory controller provides 2Tb/s of bandwidth, making it the perfect fit for Enflame Technology’s cloud AI training needs.

Complementing this interface IP, Rambus provides silicon interposer and package reference designs, and supports signal and power integrity (SI/PI) analysis.

“Enflame Technology’s repeat choice of Rambus showcases our HBM2 PHY and memory controller IP as the ideal solution for complex neural network-based AI and machine learning chips,” added Hemant Dhulla, Vice President and General Manager of IP Cores at Rambus. “Our HBM portfolio of solutions already in volume production deliver the memory performance needed at the frontier of AI computing.”