The rapid growth of AI workloads at Meta has motivated our inhouse development of AI chips, aiming to significantly reduce the total cost of ownership and mitigate risks posed by unpredictable GPU supplies.At ISCA'23, we presented Meta's first-generation AI chip, MTIA 1.This paper describes its successor, MTIA 2i, now deployed at scale and serving billions of users.MTIA 2i significantly improves upon MTIA 1, reducing total cost of ownership by 44% compared to GPUs while delivering competitive performance per watt.A key differentiator is its memory hierarchy: instead of costly HBM, it uses large SRAM alongside LPDDR.Although there has been a proliferation of publications on AI chips, they often focus on architectural design and overlook three critical aspects:(1) co-designing and optimizing ML models to work effectively with the AI chip; (2) demonstrating sufficient flexibility to support a wide range of models; and (3) during the productionization process, addressing challenges unanticipated or decisions deferred at design time, such as dealing with memory errors, safe overclocking, reducing provisioned power, and implementing real-time firmware updates to mitigate silicon design defects.A key contribution of this paper is sharing our experience with these aspects, based on our journey of productionizing MTIA 2i at scale.

The growing complexity in manufacturing and operating the hardware in AI clusters leads to significant challenges in reliability. Hyperscalars have reported various AI hardware failures during high-stake jobs such as GenAI model training, where one GPU failure could bring down the entire training job. To tackle this issue, we present CP-Bench, an open-source, Configurable and Parameterizable, PyTorch-level test suite designed to test AI hardware failure, performance degradation, and silent data corruption (SDC). Built upon open-source projects, CP-Bench contains 30+ AI workloads (e.g., Llama), and implements various checks (e.g., SDC check) within these workloads. We have deployed CP-Bench throughout Meta’s AI hardware lifecycle, spanning manufacturing, in-production diagnostics, and device RMA; CP-Bench identified various hardware issues, some of which were not caught by vendor’s tooling. Notably, vendor has acknowledged to establish CP-Bench as a valid RMA criteria and plan to integrate CP-Bench into its tooling. CP-Bench is open-sourced at https://github.com/facebookincubator/CP-Bench.

Deep learning recommendation models (DLRMs) have been used across many business-critical services at Meta and are the single largest AI application in terms of infrastructure demand in its data-centers. In this paper, we present Neo, a software-hardware co-designed system for high-performance distributed training of large-scale DLRMs. Neo employs a novel 4D parallelism strategy that combines table-wise, row-wise, column-wise, and data parallelism for training massive embedding operators in DLRMs. In addition, Neo enables extremely high-performance and memory-efficient embedding computations using a variety of critical systems optimizations, including hybrid kernel fusion, software-managed caching, and quality-preserving compression. Finally, Neo is paired with ZionEX, a new hardware platform co-designed with Neo's 4D parallelism for optimizing communications for large-scale DLRM training. Our evaluation on 128 GPUs using 16 ZionEX nodes shows that Neo outperforms existing systems by up to 40× for training 12-trillion-parameter DLRM models deployed in production.