jax: FFT on CPU noticeably slower than SciPy's FFT

The FFT implementation in JAX seems to be noticeably slower than the one in SciPy even though both use some flavor of the PocketFFT FFT implementation.

from jax.config import config

config.update("jax_enable_x64", True)

import timeit
import numpy as np
from scipy import fft as sp_fft
from jax import numpy as jnp
from jax import jit, random

key = random.PRNGKey(42)
jr = random.normal(key, (2**26, ))
r = np.array(jr)

N_IT = 7

timing = timeit.timeit(lambda: np.fft.fft(r), number=N_IT) / N_IT
print(f"NumPy: {timing} s")
timing = timeit.timeit(lambda: sp_fft.fft(r), number=N_IT) / N_IT
print(f"SciPy: {timing} s")

jax_fft = jnp.fft.fft
timing = timeit.timeit(
    lambda: jax_fft(jr).block_until_ready(), number=N_IT
) / N_IT
print(f"JAX (unjitted): {timing} s")

jax_fft_jit = jit(jax_fft)
jax_fft_jit(jr)  # Warm-up
timing = timeit.timeit(
    lambda: jax_fft_jit(jr).block_until_ready(), number=N_IT
) / N_IT
print(f"JAX (jitted): {timing} s")

On an AMD Ryzen 7 4800H (with Radeon Graphics) with JAX 0.2.18 and jaxlib 0.1.69 installed from PyPI, I get the following timings:

WARNING:absl:No GPU/TPU found, falling back to CPU. (Set TF_CPP_MIN_LOG_LEVEL=0 and rerun for more info.)
NumPy: 3.5420487681424544 s
SciPy: 1.9076589474290293 s
JAX (unjitted): 4.5806920641433475 s
JAX (jitted): 4.63735637514245 s

The timings improve if I compile JAX(lib) myself though it still compares unfavorably to SciPy (compiled JAX (unjitted): 3.506302507857202 s).

My hypothesis is that JAX is distributing binaries that are suboptimal for recent AMD CPUs and much more importantly JAX is probably using some outdated version of PocketFFT.