General Comments

The authors present a hardware-accelerated implementation of the EPICC air quality model’s horizontal advection scheme, demonstrating impressive computational performance improvements achieved through clever and careful source code translation and refactoring, although these improvements are offset by data transfer overhead. The authors couple the accelerated advection module back into the full model and demonstrate that this heterogeneous model implementation retains code integrity to produce reasonable simulation results. Moreover, the authors present a thorough performance analysis. Taken together, this is a comprehensive and impressive effort that may offer an impetus for the air quality modeling community to collectively modernize model codebases towards enabling the kind of ultra-high-resolution simulations that are necessary to characterize urban- and neighborhood-scale pollution, the scales at which communities experience harmful exposures. My comments and brief and minor.

Specific comments

Line 130: Please include, along with brief term definitions, the species continuity equation explicitly.

Line 217, 599-600: Considering the substantial CPU-GPU data transfer overhead, could a cluster deploying EPICC with GPU-HADVPPM4HIP V1.0 on a cluster with PCIe 5.0 buses, offering double the bandwidth of PCIe 4.0, expect to see additional efficiency improvements? Please include a brief discussion of hardware outlook. Regardless, I commend the authors’ focus on software optimization rather than relying purely on hardware capabilities, so that the model is maximally performant across diverse hardware and not only the latest generation.

Technical corrections

Line 45: “Air pollution, a source of…” may be better worded as “Air pollution, including…”

Line 83: Please define “HPL score”.

Line 92: Reword “enabling automated generates…” to “enabling automated generation of…”.

Line 108: “Air quality models-the advection module…” may be scan better with a comma instead of a hyphen.

Line 127: “Air pollution complex in China” maybe is meant to be “complex air pollution in China”?

Line 133: Change “developed using Fortran program language…” to include “the: between “using” and “Fortran”.

Figure 1: Wet scavenging typo for RADM; aqueous-phase chemistry seems to reference the vertical diffusion schemes.

Line 216: Change “two China’s domestic CPU…” to “two of China’s domestic CPU…”

The authors have addressed all of my major comments from the original submission. The revised manuscript shows significant improvements. I believe that GPU-accelerated computing will be important in the future if the data transfer bottleneck between the host and device can be resolved. Future generations of accelerated devices and CPUs are expected to be more tightly integrated.

Figures 6 and 7 have not been referenced or discussed in the main text.

Figure 6, the caption should be expanded to describe the left, middle, and right panels. Specify the units of color bar (ppbV, ugm-3, %). Reduce the color bar scale to see the difference between the two models. Include O3 in your evaluation because it is an important air pollutant.

Tables 5 and 6, include the normalized mean bias, as it is an important metric. Since each species has different concentration range. For example, the mean concentration of O3 is about 30 ppbV, while HONO is around 1 ppbV.

Provide the definition of GPU-like accelerator.