Meta has unveiled a new generation AI processing chip, following in the footsteps of the MTIA v1 (Meta Training and Inference Accelerator) introduced last year. This next-generation chip, known as MTIA Next Gen, focuses on customization for AI model deployment on the platform.
MTIA Next Gen boasts enhanced specs, featuring a custom 5nm TSMC manufacturing technology chip, a clock signal of 1.35GHz, maximum performance of 708 TOPS (sparse INT8) or 354 TOPS (sparse FP16/BF16), with an energy consumption rate increased to a TDP of 90W. The design also includes 64 PE sub-processing units arranged in an 8×8 configuration, with an added 384KB of SRAM per unit and a memory bandwidth increased to 1TB/s per unit.
For installation, Meta has developed a rack consisting of processing accelerator modules, with 3 chassis, each containing 12 circuit boards installed with 2 chip modules each, totaling 72 processing accelerators. The upgraded connection is PCIe Gen5, with a maximum power consumption of 90W to support complex or larger-than-normal models.
Source: Meta
TLDR: Meta introduces MTIA Next Gen, a high-performance AI processing chip with advanced customization and enhanced specifications for model deployment.
Leave a Comment