Benchmarks Performance benchmarks for SpeedyWeather.jl across architectures. This page is auto-generated at doc-build time from SpeedyWeather/benchmark/assets/benchmark_results.json, which itself is updated by SpeedyWeather/benchmark/manual_benchmarking.jl. Running the benchmark script on a new architecture adds a column to the overview table, a series to each comparison figure, and a per-architecture section to this page.
All simulations are benchmarked over several seconds (wallclock time) without output. Benchmarking excludes initialization. By default each configuration is timed up to 5 times (as long as all its timed runs together take at most 30 s), each architecture's section states its settings. SYPD is the median and Spread shows how much the runs differ. Timings still vary between processes and machines, so treat the numbers as rough rather than precise.
Overview: PrimitiveWet resolution across architectures Simulated years per wallclock day (SYPD) for the PrimitiveWetModel resolution sweep, one column per architecture. Each (T, L) configuration is reported for both the standard Legendre transform and fast Fourier transform (LT+FFT) and the single matrix transform (MT). Empty cells (—) mean the architecture has either not been benchmarked yet or did not run that specific configuration. The labels in the figures give the average grid spacing of each resolution in degrees.
T L Transform cpu-arm cpu-x86 gpu-nvidia gpu-amd reactant-cpu reactant-nvidia-gpu 32 8 LT+FFT 1308 796 8349 3763 1210 33951 32 8 MT 772 841 8249 3696 1203 21090 43 8 LT+FFT 559 343 6088 2149 584 20532 43 8 MT 253 306 6025 2830 464 17556 64 8 LT+FFT 152 100 3354 828 195 11068 64 8 MT 52 65 3355 826 120 8023 86 8 LT+FFT 58 40 1269 246 87 6583 86 8 MT 14 21 1283 245 39 2975 86 16 LT+FFT 29 21 799 202 52 3415 86 16 MT 8.8 12 796 150 25 1882 86 24 LT+FFT 20 14 700 196 27 2440 86 24 MT 6.1 9.0 699 108 16 1326 128 8 LT+FFT 14 11 423 86 29 2383 128 8 MT 1.2 — 239 39 — 692 128 16 LT+FFT 7.6 5.5 363 80 15 1339 128 16 MT 1.2 — 145 23 — 577 128 24 LT+FFT 5.5 3.6 321 76 7.7 885 128 24 MT 1.0 — 128 17 — 493 171 8 LT+FFT 5.7 4.1 215 44 10 1178 171 8 MT — — 52 — — 168 171 16 LT+FFT 3.1 2.1 179 41 5.5 645 171 16 MT — — 36 — — 97 171 24 LT+FFT 1.9 1.4 155 38 2.7 372 171 24 MT — — 31 — — 78 256 8 LT+FFT 1.4 1.0 78 16 2.6 374 256 16 LT+FFT 0.8 0.5 61 13 1.5 185 256 24 LT+FFT 0.5 0.3 50 11 0.8 121 342 8 LT+FFT — — 36 — — 130 342 16 LT+FFT — — 26 — — 71 342 24 LT+FFT — — 21 — — 42 512 8 LT+FFT — — 11 — — 32 512 16 LT+FFT — — 7.6 — — 17 512 24 LT+FFT — — 5.7 — — 12
Architecture: cpu-arm Created for SpeedyWeather.jl v0.23.0 on Tue, 06 Oct 2026 12:25:03.
Machine details — cpu-arm julia julia > versioninfo () Julia Version 1.12 . 6 Commit 15346901f00 ( 2026 - 04 - 09 19 : 20 UTC) Build Info : Official https :// julialang . org release Platform Info : OS : macOS (arm64 - apple - darwin24. 0.0 ) CPU : 8 × Apple M3 WORD_SIZE : 64 LLVM : libLLVM - 18.1 . 7 (ORCJIT, apple - m3) GC : Built with stock GC Threads : 1 default, 1 interactive, 1 GC (on 4 virtual cores) Models, default setups — cpu-arm Model truncation L Physics Δt SYPD Memory BarotropicModel 32 1 false 1800 33737 725.53 KB ShallowWaterModel 32 1 false 2400 30728 907.89 KB PrimitiveDryModel 32 8 true 2400 2063 4.78 MB PrimitiveWetModel 32 8 true 2400 1275 5.82 MB
Shallow water model, resolution — cpu-arm Model truncation L Rings Δt SYPD Memory ShallowWaterModel 32 1 48 2400 35956 907.89 KB ShallowWaterModel 43 1 64 1800 14935 1.59 MB ShallowWaterModel 64 1 96 1200 4135 3.57 MB ShallowWaterModel 86 1 128 900 1595 6.49 MB ShallowWaterModel 128 1 192 600 376 15.36 MB ShallowWaterModel 171 1 256 450 127 29.00 MB ShallowWaterModel 256 1 384 300 27 73.09 MB
Primitive wet model, resolution — cpu-arm Model truncation L Rings Transform Δt SYPD Memory PrimitiveWetModel 32 8 48 default 2400 1308 5.82 MB PrimitiveWetModel 43 8 64 default 1800 559 9.82 MB PrimitiveWetModel 64 8 96 default 1200 152 20.87 MB PrimitiveWetModel 86 8 128 default 900 58 36.31 MB PrimitiveWetModel 128 8 192 default 600 14 79.95 MB PrimitiveWetModel 171 8 256 default 450 5.7 141.86 MB PrimitiveWetModel 256 8 384 default 300 1.4 322.22 MB PrimitiveWetModel 86 16 128 default 900 29 62.14 MB PrimitiveWetModel 128 16 192 default 600 7.6 135.95 MB PrimitiveWetModel 171 16 256 default 450 3.1 239.79 MB PrimitiveWetModel 256 16 384 default 300 0.8 538.55 MB PrimitiveWetModel 86 24 128 default 900 20 88.01 MB PrimitiveWetModel 128 24 192 default 600 5.5 192.01 MB PrimitiveWetModel 171 24 256 default 450 1.9 337.81 MB PrimitiveWetModel 256 24 384 default 300 0.5 755.00 MB PrimitiveWetModel 32 8 48 matrix 2400 772 34.26 MB PrimitiveWetModel 43 8 64 matrix 1800 253 92.95 MB PrimitiveWetModel 64 8 96 matrix 1200 52 396.37 MB PrimitiveWetModel 86 8 128 matrix 900 14 1.18 GB PrimitiveWetModel 128 8 192 matrix 600 1.2 5.49 GB PrimitiveWetModel 86 16 128 matrix 900 8.8 1.21 GB PrimitiveWetModel 128 16 192 matrix 600 1.2 5.55 GB PrimitiveWetModel 86 24 128 matrix 900 6.1 1.23 GB PrimitiveWetModel 128 24 192 matrix 600 1.0 5.60 GB
Primitive Equation, Float32 vs Float64 — cpu-arm Model NF truncation L Δt SYPD Memory PrimitiveWetModel Float32 32 8 2400 1306 5.82 MB PrimitiveWetModel Float64 32 8 2400 1206 11.04 MB
Grids — cpu-arm Model truncation L Grid Rings Δt SYPD Memory PrimitiveWetModel 64 8 FullGaussianGrid 96 1200 86 30.59 MB PrimitiveWetModel 64 8 FullClenshawGrid 127 1200 61 51.09 MB PrimitiveWetModel 64 8 OctahedralGaussianGrid 96 1200 154 20.87 MB PrimitiveWetModel 64 8 OctahedralClenshawGrid 127 1200 92 32.77 MB PrimitiveWetModel 64 8 HEALPixGrid 127 1200 95 24.19 MB PrimitiveWetModel 64 8 OctaHEALPixGrid 127 1200 72 30.06 MB
Number of vertical layers — cpu-arm Model truncation L Δt SYPD Memory PrimitiveWetModel 32 4 2400 2226 3.72 MB PrimitiveWetModel 32 8 2400 1388 5.82 MB PrimitiveWetModel 32 12 2400 972 7.93 MB PrimitiveWetModel 32 16 2400 754 10.04 MB
PrimitiveDryModel: Physics or dynamics only — cpu-arm Model truncation L Dynamics Physics Δt SYPD Memory PrimitiveDryModel 32 8 true true 2400 2363 4.78 MB PrimitiveDryModel 32 8 true false 2400 3957 4.78 MB PrimitiveDryModel 32 8 false true 2400 2631 4.78 MB
PrimitiveWetModel: Physics or dynamics only — cpu-arm Model truncation L Dynamics Physics Δt SYPD Memory PrimitiveWetModel 32 8 true true 2400 1371 5.82 MB PrimitiveWetModel 32 8 true false 2400 3048 5.82 MB PrimitiveWetModel 32 8 false true 2400 1531 5.82 MB
Individual dynamics functions — cpu-arm PrimitiveWetModel | Float32 | T31 L8 | OctahedralGaussianGrid | 48 Rings — cpu-arm Function Time Memory Allocations pressure_gradient_flux! 41.458 μs 31.98 KiB 200 linear_virtual_temperature! 2.069 μs 0 bytes 0 geopotential! 7.521 μs 384 bytes 6 vertical_integration! 14.583 μs 0 bytes 0 surface_pressure_tendency! 11.625 μs 15.66 KiB 96 vertical_velocity! 23.333 μs 0 bytes 0 linear_pressure_gradient! 2.079 μs 0 bytes 0 vertical_advection! 110.167 μs 2.44 KiB 32 vordiv_tendencies! 224.584 μs 231.83 KiB 284 temperature_tendency! 291.333 μs 344.77 KiB 401 humidity_tendency! 280.000 μs 344.09 KiB 396 bernoulli_potential! 91.833 μs 114.28 KiB 129
Architecture: cpu-x86 Created for SpeedyWeather.jl v0.24.0-DEV on Mon, 09 Nov 2026 12:50:51.
Machine details — cpu-x86 julia julia > versioninfo () Julia Version 1.12 . 2 Commit ca9b6662be4 ( 2025 - 11 - 20 16 : 25 UTC) Build Info : Official https :// julialang . org release Platform Info : OS : Linux (x86_64 - linux - gnu) CPU : 128 × AMD EPYC 9554 64 - Core Processor WORD_SIZE : 64 LLVM : libLLVM - 18.1 . 7 (ORCJIT, znver4) GC : Built with stock GC Threads : 1 default, 1 interactive, 1 GC (on 128 virtual cores) Environment : LD_LIBRARY_PATH = / usr /local/ lib :/ usr /local/ lib : Settings: 10 timed runs per configuration (--repeats 10), the PrimitiveWet resolution sweep 2× longer (--length 2).
Models, default setups — cpu-x86 Model truncation L Physics Δt SYPD Runs Spread Memory BarotropicModel 32 1 false 1800 25968 10 1% 725.97 KB ShallowWaterModel 32 1 false 2400 18232 10 6% 908.33 KB PrimitiveDryModel 32 8 true 2400 1245 10 1% 4.97 MB PrimitiveWetModel 32 8 true 2400 803 10 1% 6.08 MB
Shallow water model, resolution — cpu-x86 Model truncation L Rings Δt SYPD Runs Spread Memory ShallowWaterModel 32 1 48 2400 18634 10 1% 908.33 KB ShallowWaterModel 43 1 64 1800 7847 10 4% 1.59 MB ShallowWaterModel 64 1 96 1200 2095 10 2% 3.57 MB ShallowWaterModel 86 1 128 900 866 10 0% 6.49 MB ShallowWaterModel 128 1 192 600 225 10 1% 15.36 MB ShallowWaterModel 171 1 256 450 83 10 1% 29.00 MB ShallowWaterModel 256 1 384 300 17 10 2% 73.09 MB
Primitive wet model, resolution — cpu-x86 Model truncation L Rings Transform Δt SYPD Runs Spread Memory PrimitiveWetModel 32 8 48 default 2400 796 10 0% 6.08 MB PrimitiveWetModel 43 8 64 default 1800 343 10 0% 10.24 MB PrimitiveWetModel 64 8 96 default 1200 100 10 0% 21.74 MB PrimitiveWetModel 86 8 128 default 900 40 10 1% 37.77 MB PrimitiveWetModel 128 8 192 default 600 11 10 0% 83.07 MB PrimitiveWetModel 171 8 256 default 450 4.1 10 0% 147.25 MB PrimitiveWetModel 256 8 384 default 300 1.0 10 1% 334.04 MB PrimitiveWetModel 86 16 128 default 900 21 10 1% 63.11 MB PrimitiveWetModel 128 16 192 default 600 5.5 10 0% 138.00 MB PrimitiveWetModel 171 16 256 default 450 2.1 10 1% 243.28 MB PrimitiveWetModel 256 16 384 default 300 0.5 10 3% 546.12 MB PrimitiveWetModel 86 24 128 default 900 14 10 1% 88.49 MB PrimitiveWetModel 128 24 192 default 600 3.6 10 2% 192.99 MB PrimitiveWetModel 171 24 256 default 450 1.4 10 1% 339.40 MB PrimitiveWetModel 256 24 384 default 300 0.3 10 1% 758.33 MB PrimitiveWetModel 32 8 48 matrix 2400 841 10 2% 34.52 MB PrimitiveWetModel 43 8 64 matrix 1800 306 10 2% 93.38 MB PrimitiveWetModel 64 8 96 matrix 1200 65 10 4% 397.23 MB PrimitiveWetModel 86 8 128 matrix 900 21 10 4% 1.18 GB PrimitiveWetModel 86 16 128 matrix 900 12 10 4% 1.21 GB PrimitiveWetModel 86 24 128 matrix 900 9.0 10 5% 1.23 GB
Primitive Equation, Float32 vs Float64 — cpu-x86 Model NF truncation L Δt SYPD Runs Spread Memory PrimitiveWetModel Float32 32 8 2400 802 10 1% 6.08 MB PrimitiveWetModel Float64 32 8 2400 668 10 1% 11.62 MB
Grids — cpu-x86 Model truncation L Grid Rings Δt SYPD Runs Spread Memory PrimitiveWetModel 64 8 FullGaussianGrid 96 1200 67 10 0% 32.24 MB PrimitiveWetModel 64 8 FullClenshawGrid 127 1200 40 10 1% 54.20 MB PrimitiveWetModel 64 8 OctahedralGaussianGrid 96 1200 102 10 0% 21.74 MB PrimitiveWetModel 64 8 OctahedralClenshawGrid 127 1200 56 10 2% 34.42 MB PrimitiveWetModel 64 8 HEALPixGrid 127 1200 450 10 1% 25.20 MB PrimitiveWetModel 64 8 OctaHEALPixGrid 127 1200 65 10 1% 31.49 MB
Number of vertical layers — cpu-x86 Model truncation L Δt SYPD Runs Spread Memory PrimitiveWetModel 32 4 2400 1162 10 2% 4.01 MB PrimitiveWetModel 32 8 2400 801 10 0% 6.08 MB PrimitiveWetModel 32 12 2400 580 10 1% 8.15 MB PrimitiveWetModel 32 16 2400 457 10 1% 10.23 MB
PrimitiveDryModel: Physics or dynamics only — cpu-x86 Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory PrimitiveDryModel 32 8 true true 2400 1232 10 1% 4.97 MB PrimitiveDryModel 32 8 true false 2400 2236 10 2% 4.97 MB PrimitiveDryModel 32 8 false true 2400 1410 10 2% 4.97 MB
PrimitiveWetModel: Physics or dynamics only — cpu-x86 Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory PrimitiveWetModel 32 8 true true 2400 801 10 0% 6.08 MB PrimitiveWetModel 32 8 true false 2400 1719 10 1% 6.08 MB PrimitiveWetModel 32 8 false true 2400 957 10 1% 6.08 MB
Individual dynamics functions — cpu-x86 PrimitiveWetModel | Float32 | T32 L8 | OctahedralGaussianGrid | 48 Rings — cpu-x86 Function Time Memory Allocations pressure_gradient_flux! 0.000 ns 0 bytes 0 linear_virtual_temperature! 0.000 ns 0 bytes 0 geopotential! 0.000 ns 0 bytes 0 vertical_integration! 0.000 ns 0 bytes 0 surface_pressure_tendency! 0.000 ns 0 bytes 0 vertical_velocity! 0.000 ns 0 bytes 0 linear_pressure_gradient! 0.000 ns 0 bytes 0 vertical_advection! 0.000 ns 0 bytes 0 vordiv_tendencies! 0.000 ns 0 bytes 0 temperature_tendency! 0.000 ns 0 bytes 0 humidity_tendency! 0.000 ns 0 bytes 0 bernoulli_potential! 0.000 ns 0 bytes 0
Architecture: gpu-nvidia Created for SpeedyWeather.jl v0.24.0-DEV on Mon, 09 Nov 2026 09:59:05.
Machine details — gpu-nvidia julia julia > versioninfo () Julia Version 1.12 . 2 Commit ca9b6662be4 ( 2025 - 11 - 20 16 : 25 UTC) Build Info : Official https :// julialang . org release Platform Info : OS : Linux (x86_64 - linux - gnu) CPU : 128 × AMD EPYC 9554 64 - Core Processor WORD_SIZE : 64 LLVM : libLLVM - 18.1 . 7 (ORCJIT, znver4) GC : Built with stock GC Threads : 1 default, 1 interactive, 1 GC (on 128 virtual cores) Environment : LD_LIBRARY_PATH = / usr /local/ lib :/ usr /local/ lib : julia julia > CUDA . versioninfo () CUDA toolchain : - runtime 13.4 . 0 , artifact installation - driver 580.126 . 9 for CUDA 13.5 , forward - compatible - compiler 13.5 . 37 , artifact installation CUDA libraries : - cuBLAS : 13.8 . 0 - cuSPARSE : 12.8 . 6 - cuSOLVER : 12.3 . 5 - cuFFT : 12.4 . 0 - cuRAND : 10.4 . 4 - CUPTI : 2026.3 . 1 (API 13.4 . 1 ) - NVML : 13.0 . 0 + 580.126 . 9 Julia packages : - CUDACore : 6.4 . 1 - GPUArrays : 11.5 . 16 - GPUCompiler : 2.10 . 0 - KernelAbstractions : 0.9 . 45 - CUDA_Driver_jll : 13.5 . 0 + 0 - CUDA_Compiler_jll : 0.6 . 4 + 0 - CUDA_Runtime_jll : 0.25 . 0 + 3 - NVPTX_LLVM_Backend_jll : 23.1 . 2 + 2 Toolchain : - Julia : 1.12 . 2 - LLVM : 18.1 . 7 1 device : 0 : NVIDIA H100 80 GB HBM3 (sm_90, 75.628 GiB / 79.647 GiB available) compiles to sm_90a / PTX 9.4 (LLVM : sm_90a / PTX 9.2 ) Settings: 10 timed runs per configuration (--repeats 10), the PrimitiveWet resolution sweep 2× longer (--length 2).
Models, default setups — gpu-nvidia Model truncation L Physics Δt SYPD Runs Spread Memory BarotropicModel 32 1 false 1800 21657 10 1% 217.07 KB ShallowWaterModel 32 1 false 2400 16298 10 4% 221.44 KB PrimitiveDryModel 32 8 true 2400 10705 10 6% 375.98 KB PrimitiveWetModel 32 8 true 2400 7860 10 3% 385.97 KB
Shallow water model, resolution — gpu-nvidia Model truncation L Rings Δt SYPD Runs Spread Memory ShallowWaterModel 32 1 48 2400 15995 10 1% 221.44 KB ShallowWaterModel 43 1 64 1800 11880 10 2% 365.32 KB ShallowWaterModel 64 1 96 1200 4591 10 0% 775.94 KB ShallowWaterModel 86 1 128 900 1640 10 0% 1.35 MB ShallowWaterModel 128 1 192 600 298 10 0% 3.44 MB ShallowWaterModel 171 1 256 450 156 10 0% 5.89 MB ShallowWaterModel 256 1 384 300 63 10 0% 12.75 MB
Primitive wet model, resolution — gpu-nvidia Model truncation L Rings Transform Δt SYPD Runs Spread Memory PrimitiveWetModel 32 8 48 default 2400 8349 10 2% 385.97 KB PrimitiveWetModel 43 8 64 default 1800 6088 10 2% 630.19 KB PrimitiveWetModel 64 8 96 default 1200 3354 10 0% 1.33 MB PrimitiveWetModel 86 8 128 default 900 1269 10 0% 2.30 MB PrimitiveWetModel 128 8 192 default 600 423 10 0% 5.61 MB PrimitiveWetModel 171 8 256 default 450 215 10 0% 9.68 MB PrimitiveWetModel 256 8 384 default 300 78 10 0% 21.17 MB PrimitiveWetModel 342 8 512 default 225 36 10 0% 37.11 MB PrimitiveWetModel 512 8 768 default 150 11 10 0% 82.37 MB PrimitiveWetModel 86 16 128 default 900 799 10 1% 3.35 MB PrimitiveWetModel 128 16 192 default 600 363 10 0% 7.97 MB PrimitiveWetModel 171 16 256 default 450 179 10 0% 13.87 MB PrimitiveWetModel 256 16 384 default 300 61 10 0% 30.60 MB PrimitiveWetModel 342 16 512 default 225 26 10 0% 53.89 MB PrimitiveWetModel 512 16 768 default 150 7.6 10 0% 120.12 MB PrimitiveWetModel 86 24 128 default 900 700 10 0% 4.40 MB PrimitiveWetModel 128 24 192 default 600 321 10 0% 10.33 MB PrimitiveWetModel 171 24 256 default 450 155 10 0% 18.07 MB PrimitiveWetModel 256 24 384 default 300 50 10 0% 40.04 MB PrimitiveWetModel 342 24 512 default 225 21 10 0% 70.67 MB PrimitiveWetModel 512 24 768 default 150 5.7 10 0% 157.87 MB PrimitiveWetModel 32 8 48 matrix 2400 8249 10 1% 385.97 KB PrimitiveWetModel 43 8 64 matrix 1800 6025 10 1% 630.19 KB PrimitiveWetModel 64 8 96 matrix 1200 3355 10 0% 1.33 MB PrimitiveWetModel 86 8 128 matrix 900 1283 10 1% 2.30 MB PrimitiveWetModel 128 8 192 matrix 600 239 10 0% 5.09 MB PrimitiveWetModel 171 8 256 matrix 450 52 10 0% 8.99 MB PrimitiveWetModel 86 16 128 matrix 900 796 10 0% 3.35 MB PrimitiveWetModel 128 16 192 matrix 600 145 10 0% 7.45 MB PrimitiveWetModel 171 16 256 matrix 450 36 10 0% 13.19 MB PrimitiveWetModel 86 24 128 matrix 900 699 10 0% 4.40 MB PrimitiveWetModel 128 24 192 matrix 600 128 10 0% 9.81 MB PrimitiveWetModel 171 24 256 matrix 450 31 10 0% 17.38 MB
Primitive Equation, Float32 vs Float64 — gpu-nvidia Model NF truncation L Δt SYPD Runs Spread Memory PrimitiveWetModel Float32 32 8 2400 8588 10 5% 385.97 KB PrimitiveWetModel Float64 32 8 2400 8207 10 1% 387.15 KB
Grids — gpu-nvidia Model truncation L Grid Rings Δt SYPD Runs Spread Memory PrimitiveWetModel 64 8 FullGaussianGrid 96 1200 2508 10 1% 1.33 MB PrimitiveWetModel 64 8 FullClenshawGrid 127 1200 1703 10 1% 2.29 MB PrimitiveWetModel 64 8 OctahedralGaussianGrid 96 1200 3315 10 1% 1.33 MB PrimitiveWetModel 64 8 OctahedralClenshawGrid 127 1200 2528 10 1% 2.29 MB PrimitiveWetModel 64 8 HEALPixGrid 127 1200 13283 10 16% 2.29 MB PrimitiveWetModel 64 8 OctaHEALPixGrid 127 1200 2752 10 0% 2.29 MB
Number of vertical layers — gpu-nvidia Model truncation L Δt SYPD Runs Spread Memory PrimitiveWetModel 32 4 2400 8527 10 2% 312.24 KB PrimitiveWetModel 32 8 2400 8601 10 1% 385.97 KB PrimitiveWetModel 32 12 2400 8371 10 6% 459.70 KB PrimitiveWetModel 32 16 2400 8556 10 1% 533.42 KB
PrimitiveDryModel: Physics or dynamics only — gpu-nvidia Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory PrimitiveDryModel 32 8 true true 2400 10838 10 1% 375.98 KB PrimitiveDryModel 32 8 true false 2400 13811 10 0% 375.98 KB PrimitiveDryModel 32 8 false true 2400 15410 10 0% 375.98 KB
PrimitiveWetModel: Physics or dynamics only — gpu-nvidia Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory PrimitiveWetModel 32 8 true true 2400 8611 10 1% 385.97 KB PrimitiveWetModel 32 8 true false 2400 12417 10 1% 385.97 KB PrimitiveWetModel 32 8 false true 2400 11534 10 1% 385.97 KB
Individual dynamics functions — gpu-nvidia PrimitiveWetModel | Float32 | T32 L8 | OctahedralGaussianGrid | 48 Rings — gpu-nvidia Function Time Memory Allocations pressure_gradient_flux! 0.000 ns 0 bytes 0 linear_virtual_temperature! 0.000 ns 0 bytes 0 geopotential! 0.000 ns 0 bytes 0 vertical_integration! 0.000 ns 0 bytes 0 surface_pressure_tendency! 0.000 ns 0 bytes 0 vertical_velocity! 0.000 ns 0 bytes 0 linear_pressure_gradient! 0.000 ns 0 bytes 0 vertical_advection! 0.000 ns 0 bytes 0 vordiv_tendencies! 0.000 ns 0 bytes 0 temperature_tendency! 0.000 ns 0 bytes 0 humidity_tendency! 0.000 ns 0 bytes 0 bernoulli_potential! 0.000 ns 0 bytes 0
Architecture: gpu-amd Created for SpeedyWeather.jl v0.22.1 on Tue, 29 Sep 2026 16:46:03.
Machine details — gpu-amd julia julia > versioninfo () Julia Version 1.12 . 7 Commit 6 d172b025e4 ( 2026 - 08 - 15 08 : 05 UTC) Build Info : Official https :// julialang . org release Platform Info : OS : Linux (x86_64 - linux - gnu) CPU : 128 × AMD EPYC 7 A53 64 - Core Processor WORD_SIZE : 64 LLVM : libLLVM - 18.1 . 7 (ORCJIT, znver3) GC : Built with stock GC Threads : 1 default, 1 interactive, 1 GC (on 128 virtual cores) Environment : LD_LIBRARY_PATH = / opt / cray / pe / papi / 7.2 . 0.1 / lib64 :/ opt / cray / libfabric / 1.22 . 0 / lib64 JULIA_DEPOT_PATH = / projappl / project_462000008 / decristoforo /. julia : julia julia > AMDGPU . versioninfo () AMDGPU versioninfo (AMDGPU . versioninfo () failed : ErrorException ( "could not load symbol \" hiptensorGetVersion \" : \n /opt/rocm-6.3.4/lib/libhiptensor.so: undefined symbol: hiptensorGetVersion" )) Models, default setups — gpu-amd Model truncation L Physics Δt SYPD Memory BarotropicModel 32 1 false 1800 9483 436.34 KB ShallowWaterModel 32 1 false 2400 7389 440.48 KB PrimitiveDryModel 32 8 true 2400 4148 590.11 KB PrimitiveWetModel 32 8 true 2400 3790 596.78 KB
Shallow water model, resolution — gpu-amd Model truncation L Rings Δt SYPD Memory ShallowWaterModel 32 1 48 2400 8214 440.48 KB ShallowWaterModel 43 1 64 1800 4029 753.06 KB ShallowWaterModel 64 1 96 1200 1038 1.64 MB ShallowWaterModel 86 1 128 900 84 3.26 MB ShallowWaterModel 128 1 192 600 37 6.98 MB ShallowWaterModel 171 1 256 450 20 12.12 MB ShallowWaterModel 256 1 384 300 9.0 26.62 MB
Primitive wet model, resolution — gpu-amd Model truncation L Rings Transform Δt SYPD Memory PrimitiveWetModel 32 8 48 default 2400 3763 596.78 KB PrimitiveWetModel 43 8 64 default 1800 2149 1.01 MB PrimitiveWetModel 64 8 96 default 1200 828 2.19 MB PrimitiveWetModel 86 8 128 default 900 246 4.35 MB PrimitiveWetModel 128 8 192 default 600 86 9.30 MB PrimitiveWetModel 171 8 256 default 450 44 16.12 MB PrimitiveWetModel 256 8 384 default 300 16 35.35 MB PrimitiveWetModel 86 16 128 default 900 202 5.40 MB PrimitiveWetModel 128 16 192 default 600 80 11.66 MB PrimitiveWetModel 171 16 256 default 450 41 20.31 MB PrimitiveWetModel 256 16 384 default 300 13 44.79 MB PrimitiveWetModel 86 24 128 default 900 196 6.45 MB PrimitiveWetModel 128 24 192 default 600 76 14.02 MB PrimitiveWetModel 171 24 256 default 450 38 24.50 MB PrimitiveWetModel 256 24 384 default 300 11 54.22 MB PrimitiveWetModel 32 8 48 matrix 2400 3696 596.78 KB PrimitiveWetModel 43 8 64 matrix 1800 2830 1.01 MB PrimitiveWetModel 64 8 96 matrix 1200 826 2.19 MB PrimitiveWetModel 86 8 128 matrix 900 245 3.83 MB PrimitiveWetModel 128 8 192 matrix 600 39 8.52 MB PrimitiveWetModel 86 16 128 matrix 900 150 4.88 MB PrimitiveWetModel 128 16 192 matrix 600 23 10.87 MB PrimitiveWetModel 86 24 128 matrix 900 108 5.93 MB PrimitiveWetModel 128 24 192 matrix 600 17 13.23 MB
Primitive Equation, Float32 vs Float64 — gpu-amd Model NF truncation L Δt SYPD Memory PrimitiveWetModel Float32 32 8 2400 3701 596.78 KB PrimitiveWetModel Float64 32 8 2400 3101 597.37 KB
Grids — gpu-amd Model truncation L Grid Rings Δt SYPD Memory PrimitiveWetModel 64 8 FullGaussianGrid 96 1200 515 2.28 MB PrimitiveWetModel 64 8 FullClenshawGrid 127 1200 325 3.97 MB PrimitiveWetModel 64 8 OctahedralGaussianGrid 96 1200 826 2.19 MB PrimitiveWetModel 64 8 OctahedralClenshawGrid 127 1200 509 3.80 MB PrimitiveWetModel 64 8 HEALPixGrid 127 1200 742 3.72 MB PrimitiveWetModel 64 8 OctaHEALPixGrid 127 1200 566 3.77 MB
Number of vertical layers — gpu-amd Model truncation L Δt SYPD Memory PrimitiveWetModel 32 4 2400 3890 523.05 KB PrimitiveWetModel 32 8 2400 3935 596.78 KB PrimitiveWetModel 32 12 2400 3801 670.51 KB PrimitiveWetModel 32 16 2400 3653 744.24 KB
PrimitiveDryModel: Physics or dynamics only — gpu-amd Model truncation L Dynamics Physics Δt SYPD Memory PrimitiveDryModel 32 8 true true 2400 4635 590.11 KB PrimitiveDryModel 32 8 true false 2400 5238 590.11 KB PrimitiveDryModel 32 8 false true 2400 8782 590.11 KB
PrimitiveWetModel: Physics or dynamics only — gpu-amd Model truncation L Dynamics Physics Δt SYPD Memory PrimitiveWetModel 32 8 true true 2400 3830 596.78 KB PrimitiveWetModel 32 8 true false 2400 4740 596.78 KB PrimitiveWetModel 32 8 false true 2400 7114 596.78 KB
Individual dynamics functions — gpu-amd PrimitiveWetModel | Float32 | T31 L8 | OctahedralGaussianGrid | 48 Rings — gpu-amd Function Time Memory Allocations pressure_gradient_flux! 127.728 μs 19.23 KiB 421 linear_virtual_temperature! 26.051 μs 3.64 KiB 66 geopotential! 38.305 μs 5.06 KiB 129 vertical_integration! 55.427 μs 9.19 KiB 188 surface_pressure_tendency! N/A N/A N/A vertical_velocity! 48.564 μs 10.22 KiB 211 linear_pressure_gradient! 25.540 μs 3.27 KiB 66 vertical_advection! 52.752 μs 13.06 KiB 224 vordiv_tendencies! N/A N/A N/A temperature_tendency! N/A N/A N/A humidity_tendency! N/A N/A N/A bernoulli_potential! N/A N/A N/A
Architecture: reactant-cpu Created for SpeedyWeather.jl v0.24.0-DEV on Mon, 09 Nov 2026 16:38:37.
Machine details — reactant-cpu julia julia > versioninfo () Julia Version 1.12 . 2 Commit ca9b6662be4 ( 2025 - 11 - 20 16 : 25 UTC) Build Info : Official https :// julialang . org release Platform Info : OS : Linux (x86_64 - linux - gnu) CPU : 128 × AMD EPYC 9554 64 - Core Processor WORD_SIZE : 64 LLVM : libLLVM - 18.1 . 7 (ORCJIT, znver4) GC : Built with stock GC Threads : 1 default, 1 interactive, 1 GC (on 128 virtual cores) Environment : JULIA_DEPOT_PATH = / p / tmp / maxgelbr / julia - depot - cpu :/ home / maxgelbr /. julia : LD_LIBRARY_PATH = / usr /local/ lib :/ usr /local/ lib : Reactant backend: cpu (via Reactant.set_default_backend("cpu")), 1 device(s): cpu (one used), Reactant v0.2.307, Reactant_jll v0.0.442+0.
Settings: 10 timed runs per configuration (--repeats 10), the PrimitiveWet resolution sweep 2× longer (--length 2).
Models, default setups — reactant-cpu Model truncation L Physics Δt SYPD Runs Spread Memory Compile BarotropicModel 32 1 false 1800 12873 10 1% 1.81 MB 28.0 s ShallowWaterModel 32 1 false 2400 11636 10 5% 2.00 MB 38.1 s PrimitiveDryModel 32 8 true 2400 1776 10 5% 5.77 MB 176.1 s PrimitiveWetModel 32 8 true 2400 1201 10 4% 6.88 MB 213.1 s
Shallow water model, resolution — reactant-cpu Model truncation L Rings Δt SYPD Runs Spread Memory Compile ShallowWaterModel 32 1 48 2400 11771 10 6% 2.00 MB 1.8 s ShallowWaterModel 43 1 64 1786 4413 10 3% 4.28 MB 8.3 s ShallowWaterModel 64 1 96 1200 1173 10 6% 12.85 MB 6.9 s ShallowWaterModel 86 1 128 893 529 10 5% 29.04 MB 7.3 s ShallowWaterModel 128 1 192 600 105 10 1% 91.85 MB 9.2 s ShallowWaterModel 171 1 256 449 60 10 2% 211.92 MB 9.9 s ShallowWaterModel 256 1 384 300 9.7 10 1% 693.70 MB 11.0 s
Primitive wet model, resolution — reactant-cpu Model truncation L Rings Transform Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 8 48 default 2400 1210 10 3% 6.88 MB 10.5 s PrimitiveWetModel 43 8 64 default 1786 584 10 2% 12.43 MB 33.5 s PrimitiveWetModel 64 8 96 default 1200 195 10 7% 29.94 MB 27.5 s PrimitiveWetModel 86 8 128 default 893 87 10 2% 58.40 MB 28.2 s PrimitiveWetModel 128 8 192 default 600 29 10 1% 155.32 MB 27.6 s PrimitiveWetModel 171 8 256 default 449 10 10 3% 322.67 MB 28.7 s PrimitiveWetModel 256 8 384 default 300 2.6 10 6% 937.89 MB 30.6 s PrimitiveWetModel 86 16 128 default 893 52 10 4% 81.57 MB 187.2 s PrimitiveWetModel 128 16 192 default 600 15 10 6% 205.42 MB 32.2 s PrimitiveWetModel 171 16 256 default 449 5.5 10 2% 410.16 MB 33.1 s PrimitiveWetModel 256 16 384 default 300 1.5 10 11% 1.13 GB 35.7 s PrimitiveWetModel 86 24 128 default 893 27 10 4% 104.78 MB 255.8 s PrimitiveWetModel 128 24 192 default 600 7.7 10 5% 255.58 MB 61.3 s PrimitiveWetModel 171 24 256 default 449 2.7 10 3% 497.74 MB 62.8 s PrimitiveWetModel 256 24 384 default 300 0.8 10 3% 1.32 GB 66.0 s PrimitiveWetModel 32 8 48 matrix 2400 1203 10 2% 34.08 MB 195.1 s PrimitiveWetModel 43 8 64 matrix 1786 464 10 1% 92.59 MB 12.4 s PrimitiveWetModel 64 8 96 matrix 1200 120 10 1% 395.47 MB 12.9 s PrimitiveWetModel 86 8 128 matrix 893 39 10 2% 1.18 GB 13.9 s PrimitiveWetModel 86 16 128 matrix 893 25 10 10% 1.20 GB 176.0 s PrimitiveWetModel 86 24 128 matrix 893 16 10 2% 1.23 GB 204.1 s
Primitive Equation, Float32 vs Float64 — reactant-cpu Model NF truncation L Δt SYPD Runs Spread Memory Compile PrimitiveWetModel Float32 32 8 2400 1188 10 3% 6.88 MB 8.6 s PrimitiveWetModel Float64 32 8 2400 823 10 2% 13.61 MB 279.1 s
Grids — reactant-cpu Model truncation L Grid Rings Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 64 8 FullGaussianGrid 96 1200 110 10 2% 39.66 MB 218.3 s PrimitiveWetModel 64 8 FullClenshawGrid 127 1200 92 10 3% 67.05 MB 238.2 s PrimitiveWetModel 64 8 OctahedralGaussianGrid 96 1200 193 10 2% 29.94 MB 9.4 s PrimitiveWetModel 64 8 OctahedralClenshawGrid 127 1200 127 10 1% 48.31 MB 217.5 s PrimitiveWetModel 64 8 HEALPixGrid 127 1200 126 10 10% 30.27 MB 215.4 s PrimitiveWetModel 64 8 OctaHEALPixGrid 127 1200 68 10 3% 44.41 MB 236.2 s
Number of vertical layers — reactant-cpu Model truncation L Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 4 2400 2440 10 6% 4.97 MB 171.8 s PrimitiveWetModel 32 8 2400 1312 10 10% 6.88 MB 8.4 s PrimitiveWetModel 32 12 2400 1015 10 7% 8.79 MB 177.6 s PrimitiveWetModel 32 16 2400 700 10 4% 10.71 MB 36.7 s
PrimitiveDryModel: Physics or dynamics only — reactant-cpu Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory Compile PrimitiveDryModel 32 8 true true 2400 1851 10 6% 5.77 MB 6.1 s PrimitiveDryModel 32 8 true false 2400 2142 10 3% 5.77 MB 5.5 s PrimitiveDryModel 32 8 false true 2400 2480 10 7% 5.77 MB 7.9 s
PrimitiveWetModel: Physics or dynamics only — reactant-cpu Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 8 true true 2400 1218 10 2% 6.88 MB 11.0 s PrimitiveWetModel 32 8 true false 2400 1734 10 4% 6.88 MB 6.1 s PrimitiveWetModel 32 8 false true 2400 1553 10 3% 6.88 MB 10.6 s
Individual dynamics functions — reactant-cpu PrimitiveWetModel | Float32 | T32 L8 | OctahedralGaussianGrid | 48 Rings — reactant-cpu Function Time Memory Allocations pressure_gradient_flux! 0.000 ns 0 bytes 0 linear_virtual_temperature! 0.000 ns 0 bytes 0 geopotential! 0.000 ns 0 bytes 0 vertical_integration! 0.000 ns 0 bytes 0 surface_pressure_tendency! 0.000 ns 0 bytes 0 vertical_velocity! 0.000 ns 0 bytes 0 linear_pressure_gradient! 0.000 ns 0 bytes 0 vertical_advection! 0.000 ns 0 bytes 0 vordiv_tendencies! 0.000 ns 0 bytes 0 temperature_tendency! 0.000 ns 0 bytes 0 humidity_tendency! 0.000 ns 0 bytes 0 bernoulli_potential! 0.000 ns 0 bytes 0
Primitive wet model, reverse mode (AD) with Reactant — reactant-cpu Wallclock time of the time steps forward and of the reverse-mode gradient with respect to the initial state over the same time steps, minimum over up to 3 runs, compilation excluded. The loss is the squared norm of the vorticity after the time steps.
truncation L Rings Δt Steps Checkpointing Forward Gradient Gradient/Forward Compile forward Compile gradient Status 32 8 48 2400 50 Binomial(32) 0.278 s 1.24 s 4.5 34.1 s 297.1 s ok 32 8 48 2400 100 Binomial(32) 0.557 s 2.69 s 4.8 11.8 s 46.1 s ok 32 8 48 2400 200 Binomial(32) 1.04 s 5.62 s 5.4 9.5 s 33.8 s ok 32 8 48 2400 2000 Binomial(32) 11.1 s 68.2 s 6.1 9.6 s 33.7 s non-finite or zero gradient 64 8 96 1200 50 Binomial(32) 0.878 s 3.81 s 4.3 13.6 s 46.8 s ok 64 8 96 1200 100 Binomial(32) 1.74 s 8.28 s 4.8 14.0 s 48.2 s ok 64 8 96 1200 200 Binomial(32) 3.48 s 17 s 4.9 13.8 s 48.8 s ok 64 8 96 1200 2000 Binomial(32) 34.8 s 206 s 5.9 10.7 s 35.8 s non-finite or zero gradient 128 8 192 600 50 Binomial(32) 2.94 s 13.8 s 4.7 11.6 s 35.3 s ok 128 8 192 600 100 Binomial(32) 5.83 s 29.5 s 5.1 15.1 s 47.5 s ok 128 8 192 600 200 Binomial(32) 11.6 s 61.8 s 5.3 15.2 s 47.9 s ok 128 8 192 600 2000 Binomial(32) 117 s 744 s 6.4 15.2 s 49.0 s non-finite or zero gradient
SPEEDY.f90-like configuration — reactant-cpu The PrimitiveWet model in a configuration closer to SPEEDY.f90, at the resolutions of the PrimitiveWet resolution sweep with the default transform: seasonal prescribed climatologies of sea surface temperature (SeasonalOceanClimatology), land surface temperature (SeasonalLandTemperature) and soil moisture (SeasonalSoilMoisture) instead of the slab ocean and the land bucket model, constant surface roughness lengths over land and ocean (ConstantSurfaceRoughness) instead of the learned ones, and SPEEDY.f90's near-surface wind of 0.95 times the lowermost-layer wind (ConstantWindSlowdown) with 5 m/s gusts (ConstantWindGust) instead of the learned wind slowdown and Beljaars gusts; sea ice stays thermodynamic, driven by the prescribed sea surface temperature.
Model truncation L Rings Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 8 48 2400 1228 10 4% 7.40 MB 217.0 s PrimitiveWetModel 43 8 64 1786 591 10 3% 13.29 MB 17.7 s PrimitiveWetModel 64 8 96 1200 205 10 4% 31.73 MB 14.8 s PrimitiveWetModel 86 8 128 893 86 10 5% 61.47 MB 15.9 s PrimitiveWetModel 128 8 192 600 28 10 5% 161.93 MB 16.4 s PrimitiveWetModel 171 8 256 449 9.7 10 4% 334.17 MB 20.6 s PrimitiveWetModel 256 8 384 300 2.5 10 4% 963.21 MB 19.4 s PrimitiveWetModel 86 16 128 893 49 10 6% 84.63 MB 165.4 s PrimitiveWetModel 128 16 192 600 15 10 7% 212.03 MB 22.7 s PrimitiveWetModel 171 16 256 449 5.4 10 3% 421.66 MB 23.6 s PrimitiveWetModel 256 16 384 300 1.4 10 4% 1.16 GB 26.6 s PrimitiveWetModel 86 24 128 893 26 10 4% 107.85 MB 231.9 s PrimitiveWetModel 128 24 192 600 7.6 10 3% 262.19 MB 70.1 s PrimitiveWetModel 171 24 256 449 2.6 10 2% 509.25 MB 67.9 s PrimitiveWetModel 256 24 384 300 0.8 10 11% 1.35 GB 69.4 s
Architecture: reactant-nvidia-gpu Created for SpeedyWeather.jl v0.24.0-DEV on Mon, 09 Nov 2026 14:45:29.
Machine details — reactant-nvidia-gpu julia julia > versioninfo () Julia Version 1.12 . 2 Commit ca9b6662be4 ( 2025 - 11 - 20 16 : 25 UTC) Build Info : Official https :// julialang . org release Platform Info : OS : Linux (x86_64 - linux - gnu) CPU : 128 × AMD EPYC 9554 64 - Core Processor WORD_SIZE : 64 LLVM : libLLVM - 18.1 . 7 (ORCJIT, znver4) GC : Built with stock GC Threads : 1 default, 1 interactive, 1 GC (on 128 virtual cores) Environment : LD_LIBRARY_PATH = / usr /local/ lib :/ usr /local/ lib : julia julia > CUDA . versioninfo () CUDA toolchain : - runtime 13.4 . 0 , artifact installation - driver 580.126 . 9 for CUDA 13.5 , forward - compatible - compiler 13.5 . 37 , artifact installation CUDA libraries : - cuBLAS : 13.8 . 0 - cuSPARSE : 12.8 . 6 - cuSOLVER : 12.3 . 5 - cuFFT : 12.4 . 0 - cuRAND : 10.4 . 4 - CUPTI : 2026.3 . 1 (API 13.4 . 1 ) - NVML : 13.0 . 0 + 580.126 . 9 Julia packages : - CUDACore : 6.4 . 1 - GPUArrays : 11.5 . 16 - GPUCompiler : 2.10 . 0 - KernelAbstractions : 0.9 . 45 - CUDA_Driver_jll : 13.5 . 0 + 0 - CUDA_Compiler_jll : 0.6 . 4 + 0 - CUDA_Runtime_jll : 0.25 . 0 + 3 - NVPTX_LLVM_Backend_jll : 23.1 . 2 + 2 Toolchain : - Julia : 1.12 . 2 - LLVM : 18.1 . 7 1 device : 0 : NVIDIA H100 80 GB HBM3 (sm_90, 18.934 GiB / 79.647 GiB available) compiles to sm_90a / PTX 9.4 (LLVM : sm_90a / PTX 9.2 ) Reactant backend: gpu (via Reactant.set_default_backend("gpu")), 1 device(s): NVIDIA H100 80GB HBM3 (one used), Reactant v0.2.307, Reactant_jll v0.0.442+0.
Settings: 10 timed runs per configuration (--repeats 10), the PrimitiveWet resolution sweep 2× longer (--length 2).
Models, default setups — reactant-nvidia-gpu Model truncation L Physics Δt SYPD Runs Spread Memory Compile BarotropicModel 32 1 false 1800 155202 10 13% 1.81 MB 28.5 s ShallowWaterModel 32 1 false 2400 148129 10 1% 2.00 MB 33.4 s PrimitiveDryModel 32 8 true 2400 35862 10 2% 5.77 MB 184.8 s PrimitiveWetModel 32 8 true 2400 33880 10 2% 6.88 MB 222.2 s
Shallow water model, resolution — reactant-nvidia-gpu Model truncation L Rings Δt SYPD Runs Spread Memory Compile ShallowWaterModel 32 1 48 2400 147160 10 14% 2.00 MB 1.6 s ShallowWaterModel 43 1 64 1786 105532 10 1% 4.28 MB 7.6 s ShallowWaterModel 64 1 96 1200 62888 10 2% 12.85 MB 7.9 s ShallowWaterModel 86 1 128 893 38207 10 4% 29.04 MB 7.8 s ShallowWaterModel 128 1 192 600 14612 10 1% 91.85 MB 8.1 s ShallowWaterModel 171 1 256 449 6884 10 0% 211.92 MB 8.9 s ShallowWaterModel 256 1 384 300 2050 10 1% 693.70 MB 10.6 s
Primitive wet model, resolution — reactant-nvidia-gpu Model truncation L Rings Transform Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 8 48 default 2400 33951 10 3% 6.88 MB 15.0 s PrimitiveWetModel 43 8 64 default 1786 20532 10 2% 12.43 MB 33.8 s PrimitiveWetModel 64 8 96 default 1200 11068 10 3% 29.94 MB 32.0 s PrimitiveWetModel 86 8 128 default 893 6583 10 3% 58.40 MB 32.7 s PrimitiveWetModel 128 8 192 default 600 2383 10 3% 155.32 MB 33.9 s PrimitiveWetModel 171 8 256 default 449 1178 10 2% 322.67 MB 33.6 s PrimitiveWetModel 256 8 384 default 300 374 10 1% 937.89 MB 34.6 s PrimitiveWetModel 342 8 512 default 225 130 10 1% 2.05 GB 39.2 s PrimitiveWetModel 512 8 768 default 150 32 10 0% 6.35 GB 50.0 s PrimitiveWetModel 86 16 128 default 893 3415 10 3% 81.57 MB 198.8 s PrimitiveWetModel 128 16 192 default 600 1339 10 3% 205.42 MB 41.0 s PrimitiveWetModel 171 16 256 default 449 645 10 2% 410.16 MB 41.7 s PrimitiveWetModel 256 16 384 default 300 185 10 1% 1.13 GB 45.8 s PrimitiveWetModel 342 16 512 default 225 71 10 0% 2.39 GB 49.0 s PrimitiveWetModel 512 16 768 default 150 17 10 0% 7.10 GB 59.3 s PrimitiveWetModel 86 24 128 default 893 2440 10 9% 104.78 MB 229.2 s PrimitiveWetModel 128 24 192 default 600 885 10 3% 255.58 MB 72.5 s PrimitiveWetModel 171 24 256 default 449 372 10 2% 497.74 MB 71.4 s PrimitiveWetModel 256 24 384 default 300 121 10 1% 1.32 GB 75.1 s PrimitiveWetModel 342 24 512 default 225 42 10 1% 2.73 GB 82.3 s PrimitiveWetModel 512 24 768 default 150 12 10 1% 7.86 GB 94.9 s PrimitiveWetModel 32 8 48 matrix 2400 21090 10 1% 34.08 MB 199.5 s PrimitiveWetModel 43 8 64 matrix 1786 17556 10 1% 92.59 MB 17.7 s PrimitiveWetModel 64 8 96 matrix 1200 8023 10 4% 395.47 MB 17.7 s PrimitiveWetModel 86 8 128 matrix 893 2975 10 1% 1.18 GB 18.2 s PrimitiveWetModel 128 8 192 matrix 600 692 10 5% 5.48 GB 17.7 s PrimitiveWetModel 171 8 256 matrix 449 168 10 3% 16.83 GB 19.6 s PrimitiveWetModel 86 16 128 matrix 893 1882 10 4% 1.20 GB 190.3 s PrimitiveWetModel 128 16 192 matrix 600 577 10 2% 5.54 GB 29.8 s PrimitiveWetModel 171 16 256 matrix 449 97 10 1% 16.92 GB 31.7 s PrimitiveWetModel 86 24 128 matrix 893 1326 10 3% 1.23 GB 218.9 s PrimitiveWetModel 128 24 192 matrix 600 493 10 3% 5.59 GB 60.0 s PrimitiveWetModel 171 24 256 matrix 449 78 10 1% 17.02 GB 60.0 s
Primitive Equation, Float32 vs Float64 — reactant-nvidia-gpu Model NF truncation L Δt SYPD Runs Spread Memory Compile PrimitiveWetModel Float32 32 8 2400 31808 10 15% 6.88 MB 16.3 s PrimitiveWetModel Float64 32 8 2400 15820 10 2% 13.61 MB 259.9 s
Grids — reactant-nvidia-gpu Model truncation L Grid Rings Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 64 8 FullGaussianGrid 96 1200 9596 10 3% 39.66 MB 221.0 s PrimitiveWetModel 64 8 FullClenshawGrid 127 1200 8176 10 8% 67.05 MB 219.6 s PrimitiveWetModel 64 8 OctahedralGaussianGrid 96 1200 10708 10 3% 29.94 MB 13.4 s PrimitiveWetModel 64 8 OctahedralClenshawGrid 127 1200 9377 10 4% 48.31 MB 224.2 s PrimitiveWetModel 64 8 HEALPixGrid 127 1200 10960 10 9% 30.27 MB 222.8 s PrimitiveWetModel 64 8 OctaHEALPixGrid 127 1200 9890 10 6% 44.41 MB 221.1 s
Number of vertical layers — reactant-nvidia-gpu Model truncation L Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 4 2400 41984 10 5% 4.97 MB 174.7 s PrimitiveWetModel 32 8 2400 32811 10 14% 6.88 MB 15.5 s PrimitiveWetModel 32 12 2400 26430 10 2% 8.79 MB 197.3 s PrimitiveWetModel 32 16 2400 12063 10 19% 10.71 MB 42.3 s
PrimitiveDryModel: Physics or dynamics only — reactant-nvidia-gpu Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory Compile PrimitiveDryModel 32 8 true true 2400 35864 10 12% 5.77 MB 15.8 s PrimitiveDryModel 32 8 true false 2400 43463 10 15% 5.77 MB 13.6 s PrimitiveDryModel 32 8 false true 2400 61428 10 19% 5.77 MB 10.3 s
PrimitiveWetModel: Physics or dynamics only — reactant-nvidia-gpu Model truncation L Dynamics Physics Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 8 true true 2400 32525 10 16% 6.88 MB 15.9 s PrimitiveWetModel 32 8 true false 2400 44148 10 22% 6.88 MB 10.3 s PrimitiveWetModel 32 8 false true 2400 47853 10 23% 6.88 MB 13.9 s
Individual dynamics functions — reactant-nvidia-gpu PrimitiveWetModel | Float32 | T32 L8 | OctahedralGaussianGrid | 48 Rings — reactant-nvidia-gpu Function Time Memory Allocations pressure_gradient_flux! 0.000 ns 0 bytes 0 linear_virtual_temperature! 0.000 ns 0 bytes 0 geopotential! 0.000 ns 0 bytes 0 vertical_integration! 0.000 ns 0 bytes 0 surface_pressure_tendency! 0.000 ns 0 bytes 0 vertical_velocity! 0.000 ns 0 bytes 0 linear_pressure_gradient! 0.000 ns 0 bytes 0 vertical_advection! 0.000 ns 0 bytes 0 vordiv_tendencies! 0.000 ns 0 bytes 0 temperature_tendency! 0.000 ns 0 bytes 0 humidity_tendency! 0.000 ns 0 bytes 0 bernoulli_potential! 0.000 ns 0 bytes 0
Primitive wet model, reverse mode (AD) with Reactant — reactant-nvidia-gpu Wallclock time of the time steps forward and of the reverse-mode gradient with respect to the initial state over the same time steps, minimum over up to 3 runs, compilation excluded. The loss is the squared norm of the vorticity after the time steps.
truncation L Rings Δt Steps Checkpointing Forward Gradient Gradient/Forward Compile forward Compile gradient Status 32 8 48 2400 50 Binomial(32) 0.0111 s 0.0765 s 6.9 40.4 s 318.6 s ok 32 8 48 2400 100 Binomial(32) 0.0206 s 0.14 s 6.8 15.7 s 55.0 s ok 32 8 48 2400 200 Binomial(32) 0.0378 s 0.269 s 7.1 15.8 s 55.5 s ok 32 8 48 2400 2000 Binomial(32) 0.362 s 2.93 s 8.1 15.8 s 55.4 s non-finite or zero gradient 64 8 96 1200 50 Binomial(32) 0.0168 s 0.0948 s 5.6 15.5 s 53.8 s ok 64 8 96 1200 100 Binomial(32) 0.0308 s 0.182 s 5.9 15.5 s 52.6 s ok 64 8 96 1200 200 Binomial(32) 0.0591 s 0.353 s 6.0 15.5 s 53.1 s ok 64 8 96 1200 2000 Binomial(32) 0.574 s 3.94 s 6.9 15.5 s 53.3 s non-finite or zero gradient 128 8 192 600 50 Binomial(32) 0.0363 s 0.172 s 4.7 17.2 s 59.4 s ok 128 8 192 600 100 Binomial(32) 0.0697 s 0.348 s 5.0 17.1 s 57.2 s ok 128 8 192 600 200 Binomial(32) 0.137 s 0.696 s 5.1 17.0 s 57.0 s ok 128 8 192 600 2000 Binomial(32) 1.35 s 8.03 s 6.0 17.0 s 57.1 s non-finite or zero gradient 256 8 384 300 50 Binomial(32) 0.112 s 0.509 s 4.5 20.0 s 61.0 s ok 256 8 384 300 100 Binomial(32) 0.221 s 1.07 s 4.8 19.8 s 59.2 s ok 256 8 384 300 200 Binomial(32) 0.439 s 2.2 s 5.0 19.9 s 59.1 s ok 256 8 384 300 2000 Binomial(32) 4.37 s 26.2 s 6.0 19.9 s 59.2 s non-finite or zero gradient
SPEEDY.f90-like configuration — reactant-nvidia-gpu The PrimitiveWet model in a configuration closer to SPEEDY.f90, at the resolutions of the PrimitiveWet resolution sweep with the default transform: seasonal prescribed climatologies of sea surface temperature (SeasonalOceanClimatology), land surface temperature (SeasonalLandTemperature) and soil moisture (SeasonalSoilMoisture) instead of the slab ocean and the land bucket model, constant surface roughness lengths over land and ocean (ConstantSurfaceRoughness) instead of the learned ones, and SPEEDY.f90's near-surface wind of 0.95 times the lowermost-layer wind (ConstantWindSlowdown) with 5 m/s gusts (ConstantWindGust) instead of the learned wind slowdown and Beljaars gusts; sea ice stays thermodynamic, driven by the prescribed sea surface temperature.
Model truncation L Rings Δt SYPD Runs Spread Memory Compile PrimitiveWetModel 32 8 48 2400 32344 10 6% 7.40 MB 206.8 s PrimitiveWetModel 43 8 64 1786 20314 10 27% 13.29 MB 21.5 s PrimitiveWetModel 64 8 96 1200 10895 10 67% 31.73 MB 20.1 s PrimitiveWetModel 86 8 128 893 6306 10 5% 61.47 MB 20.2 s PrimitiveWetModel 128 8 192 600 2172 10 9% 161.93 MB 22.0 s PrimitiveWetModel 171 8 256 449 1112 10 3% 334.17 MB 21.7 s PrimitiveWetModel 256 8 384 300 362 10 4% 963.21 MB 23.6 s PrimitiveWetModel 342 8 512 225 126 10 2% 2.10 GB 26.4 s PrimitiveWetModel 512 8 768 150 31 10 1% 6.45 GB 35.6 s PrimitiveWetModel 86 16 128 893 3187 10 22% 84.63 MB 175.4 s PrimitiveWetModel 128 16 192 600 1255 10 5% 212.03 MB 30.6 s PrimitiveWetModel 171 16 256 449 616 10 8% 421.66 MB 30.0 s PrimitiveWetModel 256 16 384 300 177 10 1% 1.16 GB 33.8 s PrimitiveWetModel 342 16 512 225 70 10 1% 2.44 GB 36.9 s PrimitiveWetModel 512 16 768 150 17 10 1% 7.20 GB 46.8 s PrimitiveWetModel 86 24 128 893 2224 10 18% 107.85 MB 208.3 s PrimitiveWetModel 128 24 192 600 835 10 11% 262.19 MB 62.7 s PrimitiveWetModel 171 24 256 449 358 10 3% 509.25 MB 61.4 s PrimitiveWetModel 256 24 384 300 118 10 2% 1.35 GB 64.1 s PrimitiveWetModel 342 24 512 225 40 10 1% 2.78 GB 70.6 s PrimitiveWetModel 512 24 768 150 12 10 0% 7.96 GB 82.1 s
