A naukri.com initiative
Amazon
2d
15
Image Credit: Amazon
Save big on OpenSearch: Unleashing Intel AVX-512 for binary vector performance
- OpenSearch version 2.19 now supports hardware-accelerated enhanced latency and throughput for binary vectors using Intel AVX-512 acceleration.
- With AVX-512, there is up to a 48% throughput improvement compared to previous-generation R5 instances and a 10% improvement compared to OpenSearch 2.17 and below.
- Binary vectors store embeddings more efficiently than full precision vectors, saving on storage and memory costs.
- The AVX-512 accelerator enhances Hamming distance calculations for binary vectors by utilizing the popcount operation efficiently.
- OpenSearch 2.19 integrates advanced Intel AVX-512 instructions in the FAISS engine to improve performance on Intel Xeon processors without requiring additional configurations.
- Benchmarking reveals up to 20% gains on indexing throughput and over 40% gains on search throughput with R7i instances compared to R5 instances.
- The performance improvements come from utilizing the avx512_spr optimization mode, which utilizes advanced Intel instructions for faster computation.
- Optimizing vector search with newer AVX-512 instructions can lead to a 10% throughput improvement on indexing and search operations.
- The article details testing datasets of 10 million binary vectors to analyze performance gains, showcasing the benefits of upgrading to newer Intel instances.
- Authors from Intel and AWS provide insights into the optimizations for OpenSearch workloads and suggest utilizing R7i instances for improved price-performance and TCO gains.
Read Full Article
Like
For uninterrupted reading, download the app