Vitis™ BLAS Library is a performance-optimized implementation of the standard Basic Linear Algebra Subroutines (BLAS), designed to bring you out-of-the-box acceleration on Xilinx platforms for a range of different applications like Multi-layer Perceptron (MLP) based Machine Learning, Computer Vision, Quantitative Finance among others.
Vitis BLAS library APIs like General Matrix Multiply (GEMM) and General Matrix-Vector Multiply (GEMV) are available as pre-compiled accelerators with C, C++, and Python function interfaces. Call them directly call in your application, without any additional hardware configuration. Drop-in and replace your CPU and GPU-based BLAS operations with the Vitis BLAS Library APIs for rapid prototyping and evaluation of performance benefits of Xilinx platforms.
Vitis BLAS Library Primitives and Kernels provide greater flexibility and control while designing your own unique accelerated algorithms for deployment across edge, on-premise, or cloud.
|Matrix Size||Vitis BLAS GEMM API
|Intel® MKL (16 threads, no caching) (Tops/Sec)||Speedup|
CPU: 2 Intel(R) Xeon(R) CPU E5-2640 v3 @2.60GHz, 8 cores per processor and 2 threads per core.
Xilinx: Vitis BLAS library v1.0 running 1 Alveo U200
FPGA Execution time Includes data transfer between host and device