Architecture upgrade
-
GPU support
As early as version 1.x of Milvus, we used to support GPU, but in version 2.x due to switching to a distributed architecture, and due to cost considerations, GPU support has not been added for the time being. More than a year after the release of Milvus 2.0, the Milvus community has become more and more vocal about GPU, coupled with the strong cooperation of NVIDIA engineers - adding the latest RAFT algorithm support to Knowhere (Milvus indexing engine), making Milvus Not only has GPU support been added back, but also the latest algorithm in the industry has been supported at the fastest speed. After testing, compared with the CPU HNSW index, the GPU version has a QPS improvement of more than 3 times, and some data sets have a nearly 10-fold improvement.
The following table is the QPS data of GPU-IVF-FLAT and HNSW on Milvus E2E, the host size is 8c32g, NVIDIA A100 GPU. NQ is 100:
-
Arm64 support
With the popularity of Arm64 CPU