Overview

CuPy 13.0, released on January 15, 2024, prepares NumPy 2.0 compatibility and improves CUDA kernel performance.

Main Features

NumPy 2.0 compatibility

CuPy 13 aligns its API with NumPy 2.0, supporting new dtypes and deprecating old APIs.

python 
import cupy as cp
import numpy as np

# API aligned with NumPy 2.0
x = cp.array([1, 2, 3, 4, 5])
print(cp.mean(x))        # 3.0 (GPU)
print(cp.asnumpy(x))     # convert to NumPy
print(type(x.get()))     # numpy.ndarray

Optimized CUDA kernels

Reduction kernels and sort operations are optimized for recent GPUs (Ampere, Hopper).

python 
import cupy as cp

# Optimized GPU operations
x = cp.random.randn(1_000_000)
y = cp.sort(x)           # optimized GPU sort
z = cp.sum(x ** 2)       # GPU reduction

cp.cuda.Stream.null.synchronize()
print(f'Sum of squares: {float(z):.2f}')

Sources