Part 6/12:
Unlike NumPy, which uses a less scalable pseudo-random number generator, JAX employs counter-based PRNG algorithms that support parallelization, making it better suited for large-scale, distributed training processes.
Core Data Structures: The DeviceArray (or Dissory)
Moving beyond traditional NumPy arrays, JAX introduces its fundamental data structure called DeviceArray—akin to a "ND array" but optimized for accelerators. Creating and manipulating these arrays is straightforward:
import jax.numpy as jnp
# Create a JAX array
array_jax = jnp.array([1, 2, 3])