GPU-Accelerated Gravitational Lensing and Dynamical (GLaD) modeling for cosmology and galaxies

Home

All issues

Volume 701 (September 2025)

A&A, 701 (2025) A280

Full HTML

Open Access

Table 2.

Time comparison for one-step sampling in the joint modeling.

	Process	Type of implementation	Runtime (CPU)	Runtime (GPU)
Previous work	Extended Image	Suyu et al. (2006) in `C`	2 s	–
	MGE 1D fit (1 profile; total)	Shajib (2019) in `NumPy`	∼2 × 10⁻⁴ s; 1 × 10⁻³ s	–
	MGE 1D fit (1 profile; total)	`mge.fit_1D(linear = True)` in `NumPy`¹	∼3 × 10⁻³ s; 0.02 s	–
	$\bar{v_{r}^{2}}$ $\overline{v_r^2}$ , $\bar{v_{θ}^{2}}$ $\overline{v_\theta^2}$ , $\bar{v_{ϕ}^{2}}$ $\overline{v_\phi^2}$ calculations	Integral solver (adaptive) in `NumPy` ²	13 s	–
	$v_{rms}^{pre}$ ${\boldsymbol{v}}_{\mathbf{rms}}^{\mathbf{pre}}$ calculation	`jam.axi.proj` in `NumPy`²	14 s	–

This paper	Extended image	Follows Suyu et al. (2006) in `JAX`	10 s	0.21 s
	MGE 1D fit (1 profile; total)	Follows Shajib (2019) in `JAX`	∼0.13 s; 0.52 s	∼2 × 10⁻⁴ s; 6 × 10⁻⁴ s
	$\bar{v_{r}^{2}}$ $\overline{v_r^2}$ , $\bar{v_{θ}^{2}}$ $\overline{v_\theta^2}$ , $\bar{v_{ϕ}^{2}}$ $\overline{v_\phi^2}$ calculations	Integral solver (non-adaptive) in `JAX`	118 s	0.32 s
	$v_{rms}^{pre}$ ${\boldsymbol{v}}_{\mathbf{rms}}^{\mathbf{pre}}$ calculation	`jam.axi.proj` in `JAX`	119 s	0.33 s

Notes.

https://pypi.org/project/mgefit/.

https://pypi.org/project/jampy/.

The joint modeling is based on a composite mass model with a 64 × 64 source grid (see Sect. 3.3 for the adopted profiles). The computations were performed on a 2.10 GHz CPU and an NVIDIA A100 GPU. This table presents the runtime for the MGE 1D fit for a single mass or light profile (1 profile) and for the decomposition of all mass and light profiles in the modeling (total). The comparison of all MGE 1D fits was conducted using the same number of 21 Gaussians and the same number of log radii. The running time of the integral solver shows the calculation time for the second velocity moment on the diagonal of the tensor, which is the most time-consuming part for deriving $v_{rms}^{pre}$ ${\boldsymbol{v}}_{\mathbf{rms}}^{\mathbf{pre}}$ . We adopted the same number of Gaussians for the testing, i.e., 42 Gaussians for the lens light and 95 Gaussians for the composite mass model. JAX is primarily designed to maximize the parallelization performance on GPUs. We present the running time of the JAX code on a CPU to isolate the impact of the GPU acceleration. In practice, the code is intended to run on GPUs.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.