Benchmark measurement details¶

Historical measurement. Not a measurement of the latest release.

Measurement identity	Value
Source revision	`eb8a3e6697c8defd7f3a213354088dcc659a2719`
Measured at	2026-09-21T09:00:59.439156Z
Result status	completed
Source document SHA-256	`4ff84fce31a4229bcbc04c1be3747d171457db0fad96c07930cc06074359940e`
Policy	Timing budget acceptance is separate; no budget pass is inferred from this result.
os	macOS-14.8.9-arm64-arm-64bit
architecture	arm64
cpu	arm
toolchain	3.12.10
power_mode	unknown

Download the public measurement data. This is a presentation snapshot, not the full source receipt. It preserves the original report hash and numerical observations; hostnames, local paths, and internal traces are omitted.

Whole-workflow timings. successful_execution is an execution gate, not a pixel-parity claim.
Requested and actual backends are separate. Host controls and missing terminal evidence do not prove native GPU performance.

Full measurements¶

Workload	Subject	Median µs	P90 µs	P95 µs	Samples	Result / correctness	Requested → actual	Terminal receipt
pipeline.quick.transpose-twice.rgb-1024	pillow	4,184.000	Not measured	7,066.333	100	completed; source_target_match: pass	pillow → pillow	Not proven
pipeline.quick.transpose-twice.rgb-1024	python-cpu	6,389.250	Not measured	11,298.000	100	completed; source_target_match: pass	cpu → cpu	Complete
pipeline.quick.transpose-twice.rgb-1024	python-simd	4,375.751	Not measured	50,701.125	100	completed; source_target_match: pass	simd → simd	Complete
pipeline.quick.transpose-twice.rgb-1024	python-gpu	13,528.646	Not measured	17,992.125	100	completed; source_target_match: pass	gpu → gpu	Complete
pipeline.quick.gaussianblur-invert.rgb-1024	pillow	13,910.916	Not measured	24,479.250	100	completed; source_target_match: pass	pillow → pillow	Not proven
pipeline.quick.gaussianblur-invert.rgb-1024	python-cpu	30,318.834	Not measured	41,843.625	100	completed; source_target_match: pass	cpu → cpu	Complete
pipeline.quick.gaussianblur-invert.rgb-1024	python-simd	26,159.791	Not measured	71,222.500	100	completed; source_target_match: pass	simd → simd	Complete
pipeline.quick.gaussianblur-invert.rgb-1024	python-gpu	21,731.167	Not measured	27,549.000	100	completed; source_target_match: pass	gpu → gpu	Complete
pipeline.quick.multiply-screen.rgb-1024	pillow	9,933.771	Not measured	16,858.959	100	completed; source_target_match: pass	pillow → pillow	Not proven
pipeline.quick.multiply-screen.rgb-1024	python-cpu	3,950.979	Not measured	43,318.709	100	completed; source_target_match: pass	cpu → cpu	Complete
pipeline.quick.multiply-screen.rgb-1024	python-simd	6,720.833	Not measured	9,552.166	100	completed; source_target_match: pass	simd → simd	Complete
pipeline.quick.multiply-screen.rgb-1024	python-gpu	35,873.000	Not measured	44,057.542	100	completed; source_target_match: pass	gpu → gpu	Complete
pipeline.quick.invert-mirror.rgb-1024	pillow	4,305.688	Not measured	6,853.375	100	completed; source_target_match: pass	pillow → pillow	Not proven
pipeline.quick.invert-mirror.rgb-1024	python-cpu	4,935.624	Not measured	13,851.625	100	completed; source_target_match: pass	cpu → cpu	Complete
pipeline.quick.invert-mirror.rgb-1024	python-simd	3,118.125	Not measured	50,578.000	100	completed; source_target_match: pass	simd → simd	Complete
pipeline.quick.invert-mirror.rgb-1024	python-gpu	13,645.979	Not measured	18,512.792	100	completed; source_target_match: pass	gpu → gpu	Complete

Workload boundaries¶

Repeat counts, dimensions, modes, and cache states are retained per workload:

pipeline.quick.transpose-twice.rgb-1024¶

Sample unit: timed workflow execution.

{
  "context": {
    "size": [
      1024,
      1024
    ],
    "mode": "RGB",
    "chain_length": 2,
    "operation_class": "geometry",
    "cache_state": "warm",
    "build_profile": "release"
  },
  "measurement_policy": {
    "boundary": "whole_workflow",
    "step_ids": [],
    "metrics": [
      "latency",
      "throughput"
    ],
    "warmup_iterations": 5,
    "measurement_iterations": 20,
    "samples": 5,
    "concurrency": 1,
    "cache_state": "warm",
    "correctness_gate": "source_target_match"
  }
}

pipeline.quick.gaussianblur-invert.rgb-1024¶

Sample unit: timed workflow execution.

{
  "context": {
    "size": [
      1024,
      1024
    ],
    "mode": "RGB",
    "chain_length": 2,
    "operation_class": "neighborhood",
    "cache_state": "warm",
    "build_profile": "release"
  },
  "measurement_policy": {
    "boundary": "whole_workflow",
    "step_ids": [],
    "metrics": [
      "latency",
      "throughput"
    ],
    "warmup_iterations": 5,
    "measurement_iterations": 20,
    "samples": 5,
    "concurrency": 1,
    "cache_state": "warm",
    "correctness_gate": "source_target_match"
  }
}

pipeline.quick.multiply-screen.rgb-1024¶

Sample unit: timed workflow execution.

{
  "context": {
    "size": [
      1024,
      1024
    ],
    "mode": "RGB",
    "chain_length": 2,
    "operation_class": "multi_image",
    "cache_state": "warm",
    "build_profile": "release"
  },
  "measurement_policy": {
    "boundary": "whole_workflow",
    "step_ids": [],
    "metrics": [
      "latency",
      "throughput"
    ],
    "warmup_iterations": 5,
    "measurement_iterations": 20,
    "samples": 5,
    "concurrency": 1,
    "cache_state": "warm",
    "correctness_gate": "source_target_match"
  }
}

pipeline.quick.invert-mirror.rgb-1024¶

Sample unit: timed workflow execution.

{
  "context": {
    "size": [
      1024,
      1024
    ],
    "mode": "RGB",
    "chain_length": 2,
    "operation_class": "geometry",
    "cache_state": "warm",
    "build_profile": "release"
  },
  "measurement_policy": {
    "boundary": "whole_workflow",
    "step_ids": [],
    "metrics": [
      "latency",
      "throughput"
    ],
    "warmup_iterations": 5,
    "measurement_iterations": 20,
    "samples": 5,
    "concurrency": 1,
    "cache_state": "warm",
    "correctness_gate": "source_target_match"
  }
}

Reproduce and interpret¶

Use the benchmark protocol. Correctness, native dispatch, timing regressions, process memory, and artifact size are separate claims.