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.

Meta has traditionally relied on using CPU-based servers for running inference workloads, specifically Deep Learning Recommendation Models (DLRM), but the increasing compute and memory requirements of these models have pushed the company towards using specialized solutions such as GPUs or other hardware accelerators. This paper describes the company's effort in constructing its first silicon specifically designed for recommendation systems; it describes the accelerator architecture and platform design, the software stack for enabling and optimizing PyTorch-based models and provides an initial performance evaluation. With our emerging software stack, we have made significant progress towards reaching the same or higher efficiency as the GPU: We averaged 0.9x perf/W across various DLRMs, and benchmarks show operators such as GEMMs reaching 2x perf/W. Finally, the paper describes the lessons we learned during this journey which can improve the performance and programmability of future generations of architecture.

Rabin fingerprints are short tags for large objects that can be used in a wide range of applications, such as data deduplication, web querying, packet routing, and caching. We present a pipelined hardware architecture for computing Rabin fingerprints on data being transferred on a high throughput bus. The design conducts real-time fingerprinting with short latencies, and can be tuned for optimized clock rate with "split fresh" technique. A pipelined sampling logic selects fingerprints based on the Minwise theory and adds only a few clock cycles of latency before returning the final results. The design can be replicated to work in parallel for higher throughput data traffic. This architecture is implemented on a Xilinx Virtex-6 FPGA, and is tested on a storage prototyping platform. The implementation shows that the design can achieve clock rates above 300 MHz with an order of magnitude improvement in latency over prior software implementations, while consuming little hardware resource. The scheme is extensible to other types of fingerprints and CRC computations, and is readily applicable to primary storages and caches in hybrid storage systems.

Data deduplication has proven important in backup storage systems as large amount of identical or similar data chunks exist. Recent studies have shown the great potential of data deduplication in primary storage and storage caches. Deduplications in these environments require high speed processing not to drag down production performance. This paper presents a hardware accelerator for similarity based data deduplication. It implements three compute-intensive kernel modules to improve throughput and latency in dedupe systems: sketch computation for data blocks, index searching for reference block, and delta encoding over similar blocks. Adopting pipelined computation and parallel data lookup across multiple hardware modules, our HW design is capable of processing high throughput data traffic by working on multiple data units concurrently, thus enabling wire speed dedupe for data stream where similar blocks present. Using a PC host system connected to the FPGA-based accelerator through a PCIe Gen 2×4 interface, our experiments show that the similarity based data dedupe performs 30% better in data reduction ratio than conventional dedupe techniques that look at identical blocks only. By comparing the hardware implementation with its software counterpart, the experimental results show that our preliminary FPGA implementation with maximum clock speed of 250MHz achieves at least 6 times improvement in latency over the software implementation running on state-of-art servers.