Make spaCy and Thinc up to 8 × faster on macOS by calling into Apple's native libraries.

Make sure you have Xcode installed and then install with pip:

pip install thinc-apple-ops

Matrix multiplication is one of the primary operations in machine learning. Since matrix multiplication is computationally expensive, using a fast matrix multiplication implementation can speed up training and prediction significantly.

Most linear algebra libraries provide matrix multiplication in the form of the standardized BLAS gemm functions. The work behind scences is done by a set of matrix multiplication kernels that are meticulously tuned for specific architectures. Matrix multiplication kernels use architecture-specific SIMD instructions for data-level parallism and can take factors such as cache sizes and intstruction latency into account. Thinc uses the BLIS linear algebra library, which provides optimized matrix multiplication kernels for most x86_64 and some ARM CPUs.

Recent Apple Silicon CPUs, such as the M-series used in Macs, differ from traditional x86_64 and ARM CPUs in that they have a separate matrix co-processor(s) called AMX. Since AMX is not well-documented, it is unclear how many AMX units Apple M CPUs have. It is certain that the (single) performance cluster of the M1 has an AMX unit and there is empirical evidence that both performance clusters of the M1 Pro/Max have an AMX unit.

Even though AMX units use a set of undocumented instructions, the units can be used through Apple's Accelerate linear algebra library. Since Accelerate implements the BLAS interface, it can be used as a replacement of the BLIS library that is used by Thinc. This is where the thinc-apple-ops package comes in. thinc-apple-ops extends the default Thinc ops, so that gemm matrix multiplication from Accelerate is used in place of the BLIS implementation of gemm. As a result, matrix multiplication in Thinc is performed on the fast AMX unit(s).

Using thinc-apple-ops leads to large speedups in prediction and training on Apple Silicon Macs, as shown by the benchmarks below.

This first benchmark compares prediction speed of the de_core_news_lg spaCy model between the M1 with and without thinc-apple-ops. Results for an Intel Mac Mini and AMD Ryzen 5900X are also provided for comparison. Results are in words per second. In this prediction benchmark, using thinc-apple-ops improves performance by 4.3 times.

In the second benchmark, we compare the training speed of the de_core_news_lg spaCy model (without NER). The results are in training iterations per second. Using thinc-apple-ops improves training time by 3.0 times.