From 5b0acf168f19df3020ea29f1c6f8019856bc208b Mon Sep 17 00:00:00 2001 From: Alessandro Masat <232780120+amasat01@users.noreply.github.com> Date: Fri, 6 Nov 2026 13:04:44 +0000 Subject: [PATCH 1/4] CI: cache pip installs in the pages jobs --- .github/workflows/pages.yml | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/.github/workflows/pages.yml b/.github/workflows/pages.yml index b9d18f5..bdae0b4 100644 --- a/.github/workflows/pages.yml +++ b/.github/workflows/pages.yml @@ -47,6 +47,8 @@ jobs: - uses: actions/setup-python@v7 with: python-version: "3.12" + cache: pip + cache-dependency-path: requirements.txt # the chart layout measures text with Pillow and DejaVu Sans (tools/visuals.py); without them it falls back to # an estimate and lays the charts out differently from the committed pages - name: text measuring for the chart layout @@ -86,6 +88,8 @@ jobs: - uses: actions/setup-python@v7 with: python-version: "3.12" + cache: pip + cache-dependency-path: requirements.txt # the gallery animations are built from the paired hawk and eagle checkouts (same-named branch if it exists, else main) - name: hawk and eagle checkouts for the gallery animations env: From 91d0a9692001b69e7941654693d0280c05ad4c93 Mon Sep 17 00:00:00 2001 From: Alessandro Masat <232780120+amasat01@users.noreply.github.com> Date: Sat, 7 Nov 2026 14:01:39 +0000 Subject: [PATCH 2/4] Link each package's Best practices section from Start here --- start_here.md | 6 ++++++ 1 file changed, 6 insertions(+) diff --git a/start_here.md b/start_here.md index 3f24a1a..be350c2 100644 --- a/start_here.md +++ b/start_here.md @@ -290,4 +290,10 @@ Every site has the same sections (Start here, Tutorials, How-to guides, Explanation, Reference, Contributing), and the words the family shares are in one [glossary](https://amasat01.github.io/raptor/content/glossary.html). +**Best practices**, one short checklist per package: +[hawk](https://amasat01.github.io/hawk/content/best_practices.html), +[eagle](https://amasat01.github.io/eagle/content/userguide/best_practices.html), +[aether](https://amasat01.github.io/aether/content/installation.html#best-practices), +[raptor](https://amasat01.github.io/raptor/content/installation.html#best-practices). + *deeper: [where RAPTOR fits](where_raptor_fits.md) has the full four-part story.* From 4ba7a3879af6426ed1d9ede1451d01cbc2216de2 Mon Sep 17 00:00:00 2001 From: Alessandro Masat <232780120+amasat01@users.noreply.github.com> Date: Tue, 10 Nov 2026 13:28:33 +0000 Subject: [PATCH 3/4] Cards: the site's CPU numbers, radars and sources follow the 0.7.5 cards --- _generated/SOURCES.json | 97 +++++++++++++++++++++++++++------- _generated/compare.md | 2 +- _generated/compatibility.md | 5 +- _generated/fastest_arm_grid.md | 6 +-- _generated/perf_glance.md | 25 +++++---- _generated/perf_switch.md | 6 +-- _generated/radar_cpu.md | 17 ++++-- _generated/radar_detailed.md | 32 +++++++++-- 8 files changed, 145 insertions(+), 45 deletions(-) diff --git a/_generated/SOURCES.json b/_generated/SOURCES.json index b325198..9f49472 100644 --- a/_generated/SOURCES.json +++ b/_generated/SOURCES.json @@ -1,5 +1,5 @@ { - "generated_utc": "2026-11-05T14:09:38Z", + "generated_utc": "2026-11-10T13:24:52Z", "sources": [ { "path": "benchmarks/device_names.json", @@ -13,9 +13,21 @@ "path": "benchmarks/perf_card/card_nvidia-l4.json", "md5": "5e7a57b5e623c649cc32c141d12f57d7" }, + { + "path": "benchmarks/perf_card/cpu_card_aarch64.json", + "md5": "6d62a9a0db5de2563dcceff60877ade8" + }, { "path": "benchmarks/perf_card/cpu_card_amd-epyc-9b45.json", - "md5": "db5950d1b58ec0b0d13a1e1d00f7a825" + "md5": "c64d1a5e4eee4923e73e309efce50709" + }, + { + "path": "benchmarks/perf_card/cpu_card_apple-m4-pro.json", + "md5": "46725ca065571883f279792941cf15c9" + }, + { + "path": "benchmarks/perf_card/cpu_card_intel-r-xeon-r-w-2125-cpu-4-00ghz.json", + "md5": "941ee1d88a8944dede03065df0b3aece" }, { "path": "benchmarks/perf_card/feature_card_cpu-amd-epyc-9b45.json", @@ -55,6 +67,7 @@ "numbers": [ "-0", "-0.125", + "-00", "-04", "-04,", "-04.", @@ -64,9 +77,14 @@ "-1", "-1.0", "-10", + "-10,", "-11", "-13", "-14", + "-2100.4", + "-2125", + "-2125,", + "-4", "-7", "-9", "0", @@ -87,15 +105,15 @@ "0.028", "0.038", "0.039", + "0.041", "0.043", "0.053", "0.060", - "0.061", "0.062", "0.063", "0.065", "0.069", - "0.072", + "0.071", "0.091", "0.092", "0.094", @@ -107,18 +125,18 @@ "0.13", "0.15", "0.16", - "0.17", "0.18", "0.2", "0.20", "0.21", - "0.23", + "0.22", "0.25", "0.26", "0.28", "0.29", "0.30", "0.31", + "0.32", "0.33", "0.34", "0.35", @@ -144,6 +162,7 @@ "0.60", "0.63", "0.64", + "0.65", "0.66", "0.68", "0.7", @@ -185,8 +204,8 @@ "1.0", "1.00", "1.03", - "1.04", "1.05", + "1.07", "1.08", "1.1", "1.10", @@ -199,7 +218,6 @@ "1.24", "1.29", "1.3", - "1.33", "1.37", "1.39", "1.4", @@ -209,6 +227,7 @@ "1.5", "1.51e+03", "1.53", + "1.6", "1.64", "1.65", "1.66", @@ -243,6 +262,7 @@ "113", "115", "117", + "118", "12", "12.3", "12.6", @@ -253,12 +273,15 @@ "13", "13.0", "13.3", + "13.4", + "13.5", "130", "131,072", "131,072,", "132", "1340", "135", + "136", "14", "14.2", "14.4", @@ -268,26 +291,32 @@ "141", "149", "15", + "15.0", "15.1", "15.6", "15.9", + "150", "1506", "158", + "159", "16", "16,", "16,777,216", "16,777,216,", + "16.0", "16.1", - "16.5", + "16.4", + "16.7", + "16.9", "160", "162", "167", - "168", "17.0", + "17.1", "17.5", "17.6", "17.7", - "17.9", + "17.8", "170", "18.2", "18.9", @@ -305,7 +334,7 @@ "2,", "2,000,000", "2.0", - "2.08", + "2.06", "2.1", "2.11", "2.12", @@ -316,7 +345,9 @@ "2.3", "2.32", "2.4", + "2.40", "2.48", + "2.49", "2.5", "2.58", "2.6", @@ -327,21 +358,28 @@ "2.8", "2.9", "20", - "20.9", "200", "2000", "201", + "20250110", "2026", + "209", "21", + "21.0", "21.1", "21.5", "21.8", "210", "22", + "22.5", + "22.6", "22.7", "23", + "23.5", "23.9", "239", + "24", + "24.04", "24.3", "246", "25", @@ -360,7 +398,8 @@ "288", "29.0", "29.1", - "29.4", + "29.9", + "297", "299", "3", "3,", @@ -373,7 +412,6 @@ "3.36", "3.5", "3.63", - "3.69", "3.7", "3.71", "3.76", @@ -388,6 +426,7 @@ "32.6", "33.0", "33.1", + "33.9", "333", "34", "34.9", @@ -414,6 +453,7 @@ "4,194,304,", "4.", "4.0", + "4.00", "4.03", "4.1", "4.14", @@ -431,17 +471,23 @@ "4.68", "4.7", "4.79", + "4.8", "4.87", "4.91", + "4.92", "40.6", "40.8", "404", + "407", + "41", "41.2", + "41.6", "41.9", "411", "42", "420", "43.6", + "438", "44", "44.7", "441", @@ -452,6 +498,7 @@ "46", "46.8", "467", + "47.2", "47.5", "470", "475", @@ -462,6 +509,7 @@ "491", "5", "5.0", + "5.06", "5.14", "5.3", "5.39", @@ -472,6 +520,7 @@ "51.3", "52.8", "52.9", + "520", "529", "53.5", "541", @@ -479,6 +528,7 @@ "55.1", "55.3", "55.9", + "56.4", "560.35", "565", "566", @@ -492,22 +542,27 @@ "593", "6", "6.01", + "6.02", "6.1", + "6.11", "6.2", + "6.20", + "6.3", "6.44", "6.47", "6.5", "6.54", + "6.6", "6.70", "6.8", "6.84", - "60.4", "60.5", "609", "616", "619", "62.0", "62.3", + "62.9", "629", "630", "633", @@ -528,7 +583,6 @@ "693", "694", "69639", - "6e00", "7", "7.17", "7.2", @@ -538,7 +592,8 @@ "7.27", "7.5", "7.6", - "7.73", + "7.7", + "7.70", "7.8", "7.9", "70.6", @@ -559,15 +614,18 @@ "8.06", "8.5", "8.59", + "8.6", "8.66", "8.81", "8.9", "80", "80.0", "802", + "803", "807", "811", "823", + "83.7", "84", "86", "87.5", @@ -585,6 +643,8 @@ "9.4", "9.43", "9.49", + "9.5", + "9.51", "9.80", "9.9", "9.92", @@ -596,6 +656,7 @@ "94", "979", "982", + "988", "989" ] } diff --git a/_generated/compare.md b/_generated/compare.md index 4c4dde6..f28ecc8 100644 --- a/_generated/compare.md +++ b/_generated/compare.md @@ -46,6 +46,6 @@ Every hawk kernel carries its reverse-mode (`hawk.diff.vjp`) and forward-mode (` Every kernel in this family runs unchanged on CPU threads (OpenMP), no GPU, no CUDA toolchain: -- Adaptive RKF7(8) Kepler orbits, N = 1,000,000: eagle CPU, 8 threads, 3.69 s. +- Adaptive RKF7(8) Kepler orbits, N = 1,000,000: eagle CPU, 8 threads, 2.49 s. Every number on this page traces to a committed card: [eagle's performance page](https://amasat01.github.io/eagle/content/performance.html). diff --git a/_generated/compatibility.md b/_generated/compatibility.md index f6c9f7c..f9fd285 100644 --- a/_generated/compatibility.md +++ b/_generated/compatibility.md @@ -7,4 +7,7 @@ | CPU measured | cores (logical) | platform | host compiler | |---|---|---|---| -| AMD EPYC 9B45 | None (16) | – | x86_64-conda-linux-gnu-g++ (conda-forge gcc 14.4.0-7) 14.4.0 | +| Google Axion (Arm Neoverse-V2) | None (8) | – | g++-14 (Ubuntu 14.2.0-4ubuntu2~24.04.1) 14.2.0 | +| AMD EPYC 9B45 | None (16) | – | g++-14 (Ubuntu 14.2.0-4ubuntu2~24.04.1) 14.2.0 | +| Apple M4 Pro | None (14) | – | Apple clang version 21.0.0 (clang-2100.4.1.1) | +| Intel(R) Xeon(R) W-2125 CPU @ 4.00GHz | None (8) | – | g++ (GCC) 14.2.1 20250110 (Red Hat 14.2.1-7) | diff --git a/_generated/fastest_arm_grid.md b/_generated/fastest_arm_grid.md index 346f6cf..f85587b 100644 --- a/_generated/fastest_arm_grid.md +++ b/_generated/fastest_arm_grid.md @@ -13,8 +13,8 @@ | configuration | N = 1,000,000 | |---|---| -| spread, each sample stops at its own time | eagle CPU, 8 threads (3.69 s) | -| uniform, every sample the same orbit | eagle CPU, 8 threads (2.08 s) | -| spread-short | eagle CPU, 8 threads (168 ms) | +| spread, each sample stops at its own time | eagle CPU, 8 threads (2.49 s) | +| uniform, every sample the same orbit | eagle CPU, 8 threads (2.06 s) | +| spread-short | eagle CPU, 8 threads (136 ms) | Every arm, every column, every N: [eagle's performance page](https://amasat01.github.io/eagle/content/performance.html). diff --git a/_generated/perf_glance.md b/_generated/perf_glance.md index 8a7add9..36ece80 100644 --- a/_generated/perf_glance.md +++ b/_generated/perf_glance.md @@ -29,23 +29,26 @@ | workload | arm | wall | × hawk + eagle | memory | |---|---|---:|---:|---:| -| spread | eagle CPU, 8 threads | 3.69 s | 1.0× | – | -| spread | Numba prange | 20.9 s | 5.6× | – | -| spread | eagle CPU, 1 thread | 29.4 s | 7.9× | – | -| spread | JAX jit(vmap(while_loop)) | 60.4 s | 16× | – | -| uniform | eagle CPU, 8 threads | 2.08 s | 1.0× | – | -| uniform | eagle CPU, 1 thread | 16.5 s | 7.9× | – | -| uniform | Numba prange | 17.9 s | 8.5× | – | +| spread | eagle CPU, 8 threads | 2.49 s | 1.0× | – | +| spread | Numba prange | 19.6 s | 7.8× | – | +| spread | eagle CPU, 1 thread | 19.9 s | 7.9× | – | +| spread | JAX jit(vmap(while_loop)) | 60.5 s | 24× | – | +| uniform | eagle CPU, 8 threads | 2.06 s | 1.0× | – | +| uniform | eagle CPU, 1 thread | 16.4 s | 7.9× | – | +| uniform | Numba prange | 17.8 s | 8.6× | – | | uniform | JAX jit(vmap(while_loop)) | 28.8 s | 13× | – | -| spread-short | eagle CPU, 8 threads | 168 ms | 1.0× | – | -| spread-short | Numba prange | 1.04 s | 6.1× | – | -| spread-short | eagle CPU, 1 thread | 1.33 s | 7.8× | – | -| spread-short | JAX jit(vmap(while_loop)) | 2.67 s | 15× | – | +| spread-short | eagle CPU, 8 threads | 136 ms | 1.0× | – | +| spread-short | Numba prange | 1.05 s | 7.7× | – | +| spread-short | eagle CPU, 1 thread | 1.07 s | 7.8× | – | +| spread-short | JAX jit(vmap(while_loop)) | 2.67 s | 19× | – | **Measured on:** - Kepler orbits, GPU: [NVIDIA H100 NVL](https://amasat01.github.io/eagle/content/performance.html#perf-card-gpu-nvidia-h100-nvl) - Kepler orbits, GPU: [NVIDIA L4](https://amasat01.github.io/eagle/content/performance.html#perf-card-gpu-nvidia-l4) +- Kepler orbits, CPU: [Google Axion (Arm Neoverse-V2)](https://amasat01.github.io/eagle/content/performance.html#perf-card-cpu-aarch64) - Kepler orbits, CPU: [AMD EPYC 9B45](https://amasat01.github.io/eagle/content/performance.html#perf-card-cpu-amd-epyc-9b45) +- Kepler orbits, CPU: [Apple M4 Pro](https://amasat01.github.io/eagle/content/performance.html#perf-card-cpu-apple-m4-pro) +- Kepler orbits, CPU: [Intel(R) Xeon(R) W-2125 CPU @ 4.00GHz](https://amasat01.github.io/eagle/content/performance.html#perf-card-cpu-intel-r-xeon-r-w-2125-cpu-4-00ghz) Full cards, every workload and column: [eagle's performance page](https://amasat01.github.io/eagle/content/performance.html). diff --git a/_generated/perf_switch.md b/_generated/perf_switch.md index 3cc6bfa..da9c16a 100644 --- a/_generated/perf_switch.md +++ b/_generated/perf_switch.md @@ -70,8 +70,8 @@
-
SimulationAICPUCPU and GPUGPU
-
SimulationAICPUCPU andGPUGPU
+
SimulationAICPUCPU and GPUGPU
+
SimulationAICPUCPU andGPUGPU

first-class: a glowing icon with its name not first-class: a faint icon on the middle band, no name number: the grade, 0 to 5; blank when no source confirms one

@@ -88,7 +88,7 @@
Compare all tools
Where each tool has a documented, first-class way to do each capability: CPU, GPU or both. An empty cell is not first-class. Every filled cell links to its source. The small number is the grade, 0 to 5; a blank means no source confirms one.
CapabilityraptorNVIDIA WarpJAXPyTorchCuPyNumbaNumPy
Hand-tuned kernels0both 5GPU 42GPU 4both CPU GPU 50
Fixed-step ODEsboth 5both 3both 4both 3GPU 3both CPU GPU 3CPU 3
Differentiable sensitivitiesboth 5both 4both 4both 4000
Zero-copy hand-over (DLPack)both 5both 5both 4both 5GPU 5both CPU GPU 4CPU 4
Per-sample models and RHSboth 4both 4both 5both 42both CPU GPU 51
Half precision0both 4both 4both 5GPU 3GPU 3CPU 3
Dense linear algebra, large solvers2both 3both 5both 5GPU 5CPU 3CPU 5
Ensembles and Monte Carloboth 5both 4both 5both 4GPU 4both CPU GPU 4CPU 4
Divergent workloadsboth 512
Events and early exitboth 5CPU11
Adaptive ODEs, embedded pairsboth 5both 311
Deployment outside Pythonboth 5both 3both 3both 420
Host and GPU steps in one graphGPU 41GPU 31GPU 300
Differentiable simulation into learningboth 5both 4both 4both 4000
Ensemble data for trainingboth 4both 5both 4both 5GPU 5both CPU GPU 4CPU 4
Hybrid, physics-informed modelsboth 5both 4both 4both 4000
Dense tensor training0both 5both 5000
Low-precision inference0both 3both 3000
Array-style programs0both 4both 5GPU 52CPU 5

What each tool has a documented, first-class way to do, on CPU and on GPU; speed is in the bars and the radar.

An icon lights up only when the tool's own documentation shows a designed feature or API for it; a workaround (masking every lane, raw CUDA strings, running the slowest sample to the end) does not count. Every covered cell links to its source. The tools are complementary; overlap is not a ranking.

- +
diff --git a/_generated/radar_cpu.md b/_generated/radar_cpu.md index 178282b..6e567fd 100644 --- a/_generated/radar_cpu.md +++ b/_generated/radar_cpu.md @@ -11,17 +11,26 @@
-Ensemble speedfaster is betterFootprint andturnaroundlower is betterCapabilitiesfaster is betterTensors andprecision(qualitative)One-sampleauthoring(qualitative)Low-levelcontrol(qualitative)Reach(qualitative) +Ensemble speedfaster is betterFootprint andturnaroundlower is betterCapabilitiesfaster is betterTensors andprecision(qualitative)One-sampleauthoring(qualitative)Low-levelcontrol(qualitative)Reach(qualitative)
  1. 1 Ensemble speed faster is better
  2. 2 Footprint and turnaround lower is better
  3. 3 Capabilities faster is better
  4. 4 Tensors and precision (qualitative)
  5. 5 One-sample authoring (qualitative)
  6. 6 Low-level control (qualitative)
  7. 7 Reach (qualitative)
Ensemble speed
How fast a batch of independent samples runs: uneven, regular, small and adaptive workloads.
Footprint and turnaround
What the work costs around the run: the memory it holds, the time to compile and the code to write.
Capabilities
What the libraries support beyond the ensemble loop: wide values, fused steps, control flow, small linear algebra, contractions, FFT, scatter and reductions, each against the libraries the feature card records.
Tensors and precision (qualitative)
Dense tensor maths, array-style code and half precision: where PyTorch and JAX lead.
One-sample authoring (qualitative)
Writing the code for one sample and getting derivatives of it.
Low-level control (qualitative)
How much of the execution you control from the tool's own language.
Reach (qualitative)
Where the tool fits in: your arrays, deployment outside Python and extension to new targets.
-

Radar for AMD EPYC 9B45, card of 2026-11-04, script 103c6e00. Each axis is defined in the list below the chart; hover or tap one for the numbers behind it.

-
Show the data behind the radar
AxisraptorNumbaJAXPyTorchNumPy
Ensemble speed1.00 of the radius (mean of 4)0.56 of the radius (mean of 4)0.36 of the radius (mean of 4)0.20 of the radius (mean of 1 of 4)0.35 of the radius (mean of 1 of 4)
Uneven ensembles3.69 s (1.0× the best)20.9 s (5.6× the best)60.4 s (16× the best)––
Regular ensembles2.08 s (1.0× the best)17.9 s (8.5× the best)28.8 s (13× the best)––
Small jobs167 µs (1.0× the best)368 µs (2.1× the best)491 µs (2.9× the best)44.7 ms (260× the best)15.6 ms (92× the best)
Adaptive ODEs3.76 s (1.0× the best)19.9 s (5.3× the best)60.5 s (16× the best)––
Footprint and turnaround0.60 of the radius (mean of 2 of 3)0.36 of the radius (mean of 2 of 3)0.96 of the radius (mean of 2 of 3)–0.97 of the radius (mean of 1 of 3)
Memory–––––
Compile time565 ms (7.8× the best)449 ms (6.2× the best)71.5 ms (1.0× the best)––
Code size22 lines (1.0× the best)58 lines (2.6× the best)25 lines (1.1× the best)–23 lines (1.0× the best)
Capabilities0.68 of the radius (mean of 7 of 9)––0.73 of the radius (mean of 3 of 9)0.60 of the radius (mean of 2 of 9)
Wide values90.5 ms (1.9× the best)––47.5 ms (1.0× the best)105 ms (2.2× the best)
Fused steps259 ms (2.5× the best)––101 ms (1.0× the best)–
Control flow694 ms (1.0× the best)––––
Small linear algebra, per lane41.2 ms (1.0× the best)––299 ms (7.2× the best)–
Small linear algebra, per team–––––
Contractions78.8 µs (1.0× the best)––––
FFT–––––
Scatter9.80 ms (2.2× the best)–––4.36 ms (1.0× the best)
Reductions3.63 ms (1.0× the best)––––
Tensors and precision (qualitative)1.7 of 52.7 of 54.7 of 55.0 of 54.0 of 5
Tensor workloads (qualitative)2 of 53 of 55 of 55 of 54 of 5
Array-style programming (qualitative)2 of 52 of 54 of 55 of 55 of 5
Half precision and below (qualitative)1 of 53 of 55 of 55 of 53 of 5
One-sample authoring (qualitative)5.0 of 52.0 of 54.5 of 53.5 of 51.0 of 5
Single-thread abstraction (qualitative)5 of 53 of 54 of 53 of 51 of 5
Differentiable kernels (qualitative)5 of 51 of 55 of 54 of 51 of 5
Low-level control (qualitative)3.0 of 55.0 of 52.0 of 52.0 of 52.0 of 5
Reach (qualitative)3.7 of 52.7 of 54.0 of 54.0 of 52.7 of 5
Works with your arrays (qualitative)4 of 53 of 54 of 54 of 54 of 5
Deploy outside Python (qualitative)4 of 52 of 54 of 54 of 51 of 5
Extensibility (qualitative)3 of 53 of 54 of 54 of 53 of 5
+

Radar for AMD EPYC 9B45, card of 2026-11-10, script 988fa118. Each axis is defined in the list below the chart; hover or tap one for the numbers behind it.

+
Show the data behind the radar
AxisraptorNumbaJAXPyTorchNumPy
Ensemble speed1.00 of the radius (mean of 4)0.56 of the radius (mean of 4)0.36 of the radius (mean of 4)0.20 of the radius (mean of 1 of 4)0.35 of the radius (mean of 1 of 4)
Uneven ensembles2.49 s (1.0× the best)19.6 s (7.8× the best)60.5 s (24× the best)––
Regular ensembles2.06 s (1.0× the best)17.8 s (8.6× the best)28.8 s (13× the best)––
Small jobs167 µs (1.0× the best)368 µs (2.1× the best)491 µs (2.9× the best)44.7 ms (260× the best)15.6 ms (92× the best)
Adaptive ODEs3.76 s (1.0× the best)19.9 s (5.3× the best)60.5 s (16× the best)––
Footprint and turnaround0.60 of the radius (mean of 2 of 3)0.36 of the radius (mean of 2 of 3)0.96 of the radius (mean of 2 of 3)–0.97 of the radius (mean of 1 of 3)
Memory–––––
Compile time565 ms (7.8× the best)449 ms (6.2× the best)71.5 ms (1.0× the best)––
Code size22 lines (1.0× the best)58 lines (2.6× the best)25 lines (1.1× the best)–23 lines (1.0× the best)
Capabilities0.68 of the radius (mean of 7 of 9)––0.73 of the radius (mean of 3 of 9)0.60 of the radius (mean of 2 of 9)
Wide values90.5 ms (1.9× the best)––47.5 ms (1.0× the best)105 ms (2.2× the best)
Fused steps259 ms (2.5× the best)––101 ms (1.0× the best)–
Control flow694 ms (1.0× the best)––––
Small linear algebra, per lane41.2 ms (1.0× the best)––299 ms (7.2× the best)–
Small linear algebra, per team–––––
Contractions78.8 µs (1.0× the best)––––
FFT–––––
Scatter9.80 ms (2.2× the best)–––4.36 ms (1.0× the best)
Reductions3.63 ms (1.0× the best)––––
Tensors and precision (qualitative)1.7 of 52.7 of 54.7 of 55.0 of 54.0 of 5
Tensor workloads (qualitative)2 of 53 of 55 of 55 of 54 of 5
Array-style programming (qualitative)2 of 52 of 54 of 55 of 55 of 5
Half precision and below (qualitative)1 of 53 of 55 of 55 of 53 of 5
One-sample authoring (qualitative)5.0 of 52.0 of 54.5 of 53.5 of 51.0 of 5
Single-thread abstraction (qualitative)5 of 53 of 54 of 53 of 51 of 5
Differentiable kernels (qualitative)5 of 51 of 55 of 54 of 51 of 5
Low-level control (qualitative)3.0 of 55.0 of 52.0 of 52.0 of 52.0 of 5
Reach (qualitative)3.7 of 52.7 of 54.0 of 54.0 of 52.7 of 5
Works with your arrays (qualitative)4 of 53 of 54 of 54 of 54 of 5
Deploy outside Python (qualitative)4 of 52 of 54 of 54 of 51 of 5
Extensibility (qualitative)3 of 53 of 54 of 54 of 53 of 5
+ + +

Every measured axis is the per-sample ensemble workload raptor is built for; dense tensor maths, training and half precision are the qualitative axes, where PyTorch and JAX lead. Outer is better; overlaps are complementary, not a ranking.

NVIDIA Warp gives you near-total control of every kernel, short of raw CUDA. raptor hands that control to the engine on purpose, so compaction and the launch shape are decided at run time; the two are different trade-offs, not a ranking.

How to read this radar

Each of the seven majors is the plain mean of its minor axes (the detailed radar draws every minor), with equal weights and the same rule for every tool; a minor a tool has no reading on is left out of its mean and the major's marker is drawn hollow. Hover or tap a major to see its minors.

Each measured axis is stretched over its own best-to-worst range, so ring labels differ per axis (code size: at least 4x). Axes are grouped by theme, so read the axes, not the area.

Italic axes are the author's judgement, not measurements: hollow markers, a linear scale, and no marker where a tool has no reading. A hollow marker on a measured axis means a minor is missing. The hover panel has each axis's full wording.

Every number, in the full tables.

- + ``` diff --git a/_generated/radar_detailed.md b/_generated/radar_detailed.md index bc21ce8..d3903e5 100644 --- a/_generated/radar_detailed.md +++ b/_generated/radar_detailed.md @@ -51,11 +51,26 @@
+ + +
+ + +
+
+
+ +
+
+ +
+
+
@@ -67,19 +82,28 @@
-UnevenensemblesRegularensemblesSmall jobsAdaptiveODEsMemoryCompiletimeCode sizeTensorsArray styleHalfprecisionSingle-threadDerivativesControlYour arraysDeployExtensibilityWide valuesFused stepsControlflowSmall linearalgebra, per laneSmall linearalgebra, per teamContractionsFFTScatterReductions +UnevenensemblesRegularensemblesSmall jobsAdaptiveODEsMemoryCompiletimeCode sizeTensorsArray styleHalfprecisionSingle-threadDerivativesControlYour arraysDeployExtensibilityWide valuesFused stepsControlflowSmall linearalgebra, per laneSmall linearalgebra, per teamContractionsFFTScatterReductions
  1. 1 Uneven ensembles faster is better
  2. 2 Regular ensembles faster is better
  3. 3 Small jobs faster is better
  4. 4 Adaptive ODEs faster is better
  5. 5 Memory smaller is better
  6. 6 Compile time lower is better
  7. 7 Code size smaller is better
  8. 8 Tensor workloads (qualitative)
  9. 9 Array-style programming (qualitative)
  10. 10 Half precision and below (qualitative)
  11. 11 Single-thread abstraction (qualitative)
  12. 12 Differentiable kernels (qualitative)
  13. 13 Low-level control (qualitative)
  14. 14 Works with your arrays (qualitative)
  15. 15 Deploy outside Python (qualitative)
  16. 16 Extensibility (qualitative)
  17. 17 Wide values faster is better
  18. 18 Fused steps faster is better
  19. 19 Control flow faster is better
  20. 20 Small linear algebra, per lane faster is better
  21. 21 Small linear algebra, per team faster is better
  22. 22 Contractions faster is better
  23. 23 FFT faster is better
  24. 24 Scatter faster is better
  25. 25 Reductions faster is better
Uneven ensembles
How long a batch takes when every sample stops at its own step, at N = 1,000,000.
Regular ensembles
How long a batch takes when every sample runs the same orbit, at N = 1,000,000.
Small jobs
Wall time of the smallest batch (N = 1,000), where launch overhead matters most. Not re-measured by the 0.7.3 cards: the value is from the earlier card named in the axis note.
Adaptive ODEs
Wall time of an adaptive-step orbit propagator (RK7(8)) at N = 1,000,000. Not re-measured as a separate card in 0.7.3: the value is from the earlier card named in the axis note.
Memory
Peak memory the process holds for the same workload (device memory on a GPU; host memory on a CPU, net of the benchmark harness's control copy).
Compile time
Cold compile of the adaptive kernel on empty caches. Not re-measured by the 0.7.3 cards: the value is from the earlier card named in the axis note.
Code size
Lines of code the task takes, non-blank and non-comment.
Tensor workloads (qualitative)
Tensor workloads (qualitative): how naturally a tool expresses dense array maths (batched linear algebra, convolutions, broadcasting) and how much of it its tuned libraries cover. 5 = this is its home ground.
Array-style programming (qualitative)
Array-style programming (qualitative): how naturally you write whole-array expressions eagerly, the NumPy way. 5 = the array expression is the programming model. Scores are the author's judgement.
Half precision and below (qualitative)
Half precision and below (qualitative): FP16, BF16, FP8 and FP4 arithmetic and tensor cores (FP32 and FP64 are measured on the other axes). 5 = these formats and tensor cores are first-class paths. Scores are the author's judgement.
Single-thread abstraction (qualitative)
Single-thread abstraction (qualitative): how fully you can write the code for one sample, one step, and let the tool run it over the batch: threads, launches, early stopping and scheduling included. 5 = you write one sample and nothing else. Scores are the author's judgement.
Differentiable kernels (qualitative)
Differentiable kernels (qualitative): how directly you get forward- and reverse-mode derivatives of your own kernels, on GPU and on CPU. 5 = built in for every kernel written in the tool's own vocabulary.
Low-level control (qualitative)
Low-level control (qualitative): how much of the GPU's execution you control from the tool's own language: thread indexing, shared memory, atomics, launch shape, synchronisation (raw CUDA/C++ not counted). 5 = you write what each thread does and how it is launched. On a CPU the same question is asked of threads, SIMD and memory layout. Scores are the author's judgement.
Works with your arrays (qualitative)
Works with your arrays (qualitative): how freely arrays move to and from other frameworks without copies. 5 = zero-copy exchange is the norm.
Deploy outside Python (qualitative)
Deploy outside Python (qualitative): whether a compiled kernel can run from programs that are not Python. 5 = a compiled artefact any program can load.
Extensibility (qualitative)
Extensibility (qualitative): how readily new hardware, backends and features can be added without losing speed. 5 = a documented plugin path across many targets. Scores are the author's judgement.
Wide values
Wall time of samples that hold a vector of values at the feature card's float64 row: our best arm against the best arm of each library the row records.
Fused steps
Wall time of several integration steps fused into one launch at the feature card's float64 row: our best arm against the best arm of each library the row records.
Control flow
Wall time of branches and early exits inside a kernel at the feature card's float64 row: our best arm against the best arm of each library the row records.
Small linear algebra, per lane
Wall time of one small matrix per lane at the feature card's float64 row: our best arm against the best arm of each library the row records.
Small linear algebra, per team
Wall time of one matrix shared by a team of lanes at the feature card's float64 row: our best arm against the best arm of each library the row records.
Contractions
Wall time of small tensor contractions at the feature card's float64 row: our best arm against the best arm of each library the row records.
FFT
Wall time of a batched FFT at the feature card's float64 row: our best arm against the best arm of each library the row records.
Scatter
Wall time of adding values into bins by key at the feature card's float64 row: our best arm against the best arm of each library the row records.
Reductions
Wall time of summing over samples at the feature card's float64 row: our best arm against the best arm of each library the row records.
-

Radar for AMD EPYC 9B45, card of 2026-11-04, script 103c6e00. Each axis is defined in the list below the chart; hover or tap one for the numbers behind it.

-
Show the data behind the radar
AxisraptorNumbaJAXPyTorchNumPy
Uneven ensembles3.69 s (1.0× the best)20.9 s (5.6× the best)60.4 s (16× the best)––
Regular ensembles2.08 s (1.0× the best)17.9 s (8.5× the best)28.8 s (13× the best)––
Small jobs167 µs (1.0× the best)368 µs (2.1× the best)491 µs (2.9× the best)44.7 ms (260× the best)15.6 ms (92× the best)
Adaptive ODEs3.76 s (1.0× the best)19.9 s (5.3× the best)60.5 s (16× the best)––
Memory–––––
Compile time565 ms (7.8× the best)449 ms (6.2× the best)71.5 ms (1.0× the best)––
Code size22 lines (1.0× the best)58 lines (2.6× the best)25 lines (1.1× the best)–23 lines (1.0× the best)
Tensor workloads (qualitative)2 of 53 of 55 of 55 of 54 of 5
Array-style programming (qualitative)2 of 52 of 54 of 55 of 55 of 5
Half precision and below (qualitative)1 of 53 of 55 of 55 of 53 of 5
Single-thread abstraction (qualitative)5 of 53 of 54 of 53 of 51 of 5
Differentiable kernels (qualitative)5 of 51 of 55 of 54 of 51 of 5
Low-level control (qualitative)3 of 55 of 52 of 52 of 52 of 5
Works with your arrays (qualitative)4 of 53 of 54 of 54 of 54 of 5
Deploy outside Python (qualitative)4 of 52 of 54 of 54 of 51 of 5
Extensibility (qualitative)3 of 53 of 54 of 54 of 53 of 5
Wide values90.5 ms (1.9× the best)––47.5 ms (1.0× the best)105 ms (2.2× the best)
Fused steps259 ms (2.5× the best)––101 ms (1.0× the best)–
Control flow694 ms (1.0× the best)––––
Small linear algebra, per lane41.2 ms (1.0× the best)––299 ms (7.2× the best)–
Small linear algebra, per team–––––
Contractions78.8 µs (1.0× the best)––––
FFT–––––
Scatter9.80 ms (2.2× the best)–––4.36 ms (1.0× the best)
Reductions3.63 ms (1.0× the best)––––
+

Radar for AMD EPYC 9B45, card of 2026-11-10, script 988fa118. Each axis is defined in the list below the chart; hover or tap one for the numbers behind it.

+
Show the data behind the radar
AxisraptorNumbaJAXPyTorchNumPy
Uneven ensembles2.49 s (1.0× the best)19.6 s (7.8× the best)60.5 s (24× the best)––
Regular ensembles2.06 s (1.0× the best)17.8 s (8.6× the best)28.8 s (13× the best)––
Small jobs167 µs (1.0× the best)368 µs (2.1× the best)491 µs (2.9× the best)44.7 ms (260× the best)15.6 ms (92× the best)
Adaptive ODEs3.76 s (1.0× the best)19.9 s (5.3× the best)60.5 s (16× the best)––
Memory–––––
Compile time565 ms (7.8× the best)449 ms (6.2× the best)71.5 ms (1.0× the best)––
Code size22 lines (1.0× the best)58 lines (2.6× the best)25 lines (1.1× the best)–23 lines (1.0× the best)
Tensor workloads (qualitative)2 of 53 of 55 of 55 of 54 of 5
Array-style programming (qualitative)2 of 52 of 54 of 55 of 55 of 5
Half precision and below (qualitative)1 of 53 of 55 of 55 of 53 of 5
Single-thread abstraction (qualitative)5 of 53 of 54 of 53 of 51 of 5
Differentiable kernels (qualitative)5 of 51 of 55 of 54 of 51 of 5
Low-level control (qualitative)3 of 55 of 52 of 52 of 52 of 5
Works with your arrays (qualitative)4 of 53 of 54 of 54 of 54 of 5
Deploy outside Python (qualitative)4 of 52 of 54 of 54 of 51 of 5
Extensibility (qualitative)3 of 53 of 54 of 54 of 53 of 5
Wide values90.5 ms (1.9× the best)––47.5 ms (1.0× the best)105 ms (2.2× the best)
Fused steps259 ms (2.5× the best)––101 ms (1.0× the best)–
Control flow694 ms (1.0× the best)––––
Small linear algebra, per lane41.2 ms (1.0× the best)––299 ms (7.2× the best)–
Small linear algebra, per team–––––
Contractions78.8 µs (1.0× the best)––––
FFT–––––
Scatter9.80 ms (2.2× the best)–––4.36 ms (1.0× the best)
Reductions3.63 ms (1.0× the best)––––
+ + +

Every measured axis is the per-sample ensemble workload raptor is built for; dense tensor maths, training and half precision are the qualitative axes, where PyTorch and JAX lead. Outer is better; overlaps are complementary, not a ranking.

NVIDIA Warp gives you near-total control of every kernel, short of raw CUDA. raptor hands that control to the engine on purpose, so compaction and the launch shape are decided at run time; the two are different trade-offs, not a ranking.

How to read this radar

Each of the seven majors is the plain mean of its minor axes (the detailed radar draws every minor), with equal weights and the same rule for every tool; a minor a tool has no reading on is left out of its mean and the major's marker is drawn hollow. Hover or tap a major to see its minors.

Each measured axis is stretched over its own best-to-worst range, so ring labels differ per axis (code size: at least 4x). Axes are grouped by theme, so read the axes, not the area.

Italic axes are the author's judgement, not measurements: hollow markers, a linear scale, and no marker where a tool has no reading. A hollow marker on a measured axis means a minor is missing. The hover panel has each axis's full wording.

Every number, in the full tables.

- + From 421169a11a215ce9b3b505452e53d5670aa553e0 Mon Sep 17 00:00:00 2001 From: Alessandro Masat <232780120+amasat01@users.noreply.github.com> Date: Tue, 10 Nov 2026 14:00:57 +0000 Subject: [PATCH 4/4] Cards: regenerate against eagle's final 0.7.5 head --- _generated/SOURCES.json | 2 +- _generated/perf_switch.md | 6 +++--- 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/_generated/SOURCES.json b/_generated/SOURCES.json index 9f49472..04f2d9c 100644 --- a/_generated/SOURCES.json +++ b/_generated/SOURCES.json @@ -1,5 +1,5 @@ { - "generated_utc": "2026-11-10T13:24:52Z", + "generated_utc": "2026-11-10T14:00:38Z", "sources": [ { "path": "benchmarks/device_names.json", diff --git a/_generated/perf_switch.md b/_generated/perf_switch.md index da9c16a..3cc6bfa 100644 --- a/_generated/perf_switch.md +++ b/_generated/perf_switch.md @@ -70,8 +70,8 @@
-
SimulationAICPUCPU and GPUGPU
-
SimulationAICPUCPU andGPUGPU
+
SimulationAICPUCPU and GPUGPU
+
SimulationAICPUCPU andGPUGPU

first-class: a glowing icon with its name not first-class: a faint icon on the middle band, no name number: the grade, 0 to 5; blank when no source confirms one

@@ -88,7 +88,7 @@
Compare all tools
Where each tool has a documented, first-class way to do each capability: CPU, GPU or both. An empty cell is not first-class. Every filled cell links to its source. The small number is the grade, 0 to 5; a blank means no source confirms one.
CapabilityraptorNVIDIA WarpJAXPyTorchCuPyNumbaNumPy
Hand-tuned kernels0both 5GPU 42GPU 4both CPU GPU 50
Fixed-step ODEsboth 5both 3both 4both 3GPU 3both CPU GPU 3CPU 3
Differentiable sensitivitiesboth 5both 4both 4both 4000
Zero-copy hand-over (DLPack)both 5both 5both 4both 5GPU 5both CPU GPU 4CPU 4
Per-sample models and RHSboth 4both 4both 5both 42both CPU GPU 51
Half precision0both 4both 4both 5GPU 3GPU 3CPU 3
Dense linear algebra, large solvers2both 3both 5both 5GPU 5CPU 3CPU 5
Ensembles and Monte Carloboth 5both 4both 5both 4GPU 4both CPU GPU 4CPU 4
Divergent workloadsboth 512
Events and early exitboth 5CPU11
Adaptive ODEs, embedded pairsboth 5both 311
Deployment outside Pythonboth 5both 3both 3both 420
Host and GPU steps in one graphGPU 41GPU 31GPU 300
Differentiable simulation into learningboth 5both 4both 4both 4000
Ensemble data for trainingboth 4both 5both 4both 5GPU 5both CPU GPU 4CPU 4
Hybrid, physics-informed modelsboth 5both 4both 4both 4000
Dense tensor training0both 5both 5000
Low-precision inference0both 3both 3000
Array-style programs0both 4both 5GPU 52CPU 5

What each tool has a documented, first-class way to do, on CPU and on GPU; speed is in the bars and the radar.

An icon lights up only when the tool's own documentation shows a designed feature or API for it; a workaround (masking every lane, raw CUDA strings, running the slowest sample to the end) does not count. Every covered cell links to its source. The tools are complementary; overlap is not a ranking.

- +