Meta has introduced its self-developed Vistara custom chip solution, which addresses the challenges of high global hardware prices and memory shortages by repurposing DDR4 memory modules from retired servers and integrating them into new servers that rely on DDR5. This solution utilizes a self-developed CXL ASIC chip to bridge DDR4 memory to the host processor, expanding memory capacity and reducing latency. Vistara has already been deployed in production within Meta's hyperscale infrastructure, covering millions of servers for scenarios such as disaggregated machine learning inference and big data processing, reducing the number of servers required for inference AI by up to 25%.
