Built-in data

The package ships a collection of data files, each describing the relative cost of floating-point operations on one specific CPU — either measured by this package's own benchmark suite or derived from published latency numbers. The default FLOP weights are an aggregate over this whole collection, and get_builtin_flop_weights(key_filter=...) (see FLOP weights) lets you aggregate over any subset of it. This page documents what the collection contains and how the keys that select from it are structured.

Key anatomy

Every entry is identified by a .-separated key that mirrors its location in the data tree:

<isa> . <µarch family> . <source type> . <entry>

• ISA: arm or x86.
• µarch family: for arm, the ISA version (v8_x, v9_0, v9_2); for x86, the vendor followed by a microarchitecture generation (amd.2020_zen3, intel.2021_golden_cove_gen_12, ...) — so x86 keys are one level deeper.
• source type: where the numbers come from (see below).
• entry: the individual CPU / data source.

A key_filter is a plain substring match against these keys, so any level works as a filter: "arm", "zen", "benchmarks", "apple_m4", "x86.intel", ...

Source types

Per the methodology, three types of sources are used:

• benchmarks: results of this package's own benchmark suite (counted_float benchmark), run on the machine named by the entry. These cover all FLOP types, including operations without hardware support (sin, pow, ...).
• specs: instruction latencies from official vendor documentation. These only cover operations with hardware instructions (arithmetic, sqrt, conversions) — the transcendental functions are missing.
• analysis_* (under other/ for x86): instruction latencies measured by third parties, notably Agner Fog and uops.info. Like spec sheets, these cover hardware instructions only.

Available entries

Running counted_float show-data (see the CLI reference) prints the full collection as a tree, alongside the weights each entry and each aggregation level resolves to:

The colored bands are the interior (aggregate) nodes — one shade per level of the key hierarchy — while the plain rows are the individual data sources; / marks a weight that is absent and could not be imputed (the trailing flop-type columns are cut from the screenshot; run the command for the full matrix). The same collection in tabular form:

ISA	µarch family	Entries
arm	v8_x	benchmarks: apple_m1_github_actions, apple_m3_max_mbp16, apple_m3_mba15, aws_graviton_2_neoverse_n1_ec2_m6g_large, aws_graviton_3_neoverse_v1_ec2_m7g_large — specs: arm_cortex_a76, arm_cortex_x1, arm_neoverse_n1, arm_neoverse_v1
arm	v9_0	benchmarks: aws_graviton_4_neoverse_v2_ec2_m8g_large, azure_cobalt_100_neoverse_n2_github_actions — specs: arm_cortex_x2, arm_cortex_x3, arm_neoverse_n2, arm_neoverse_v2
arm	v9_2	benchmarks: apple_m4_pro_mbp16, aws_graviton_5_neoverse_v3_ec2_m9g_large — specs: arm_cortex_x4, arm_cortex_x925, arm_neoverse_v3
x86	amd.2017_zen1	benchmarks: amd_epyc_7571_ec2_m5a_xlarge — other: analysis_uops_info_zen1+
x86	amd.2020_zen3	benchmarks: amd_epyc_7763_windows_github_actions, amd_epyc_7r13_ec2_m6a_xlarge — other: analysis_agner_fog_r7_5800x, analysis_uops_info_zen3
x86	amd.2022_zen4	benchmarks: amd_epyc_9r14_ec2_m7a_large, amd_epyc_9v74_linux_github_actions — other: analysis_agner_fog_r9_7900x, analysis_uops_info_zen4, specs_amd
x86	amd.2024_zen5	benchmarks: amd_epyc_9r45_ec2_m8a_large — other: analysis_agner_fog_r7_9800x3d, specs_amd
x86	intel.2017_coffee_lake_gen_8	benchmarks: intel_i7_8550U_windows, intel_i7_8700B_macos_github_actions — other: analysis_agner_fog_coffee_lake, analysis_uops_info_coffee_lake
x86	intel.2019_sunny_cove_gen_10	benchmarks: intel_xeon_8375c_ice_lake_ec2_m6i_xlarge — other: analysis_agner_fog_ice_lake, analysis_uops_info_ice_lake, analysis_uops_info_tiger_lake
x86	intel.2021_golden_cove_gen_12	benchmarks: intel_xeon_8488c_sapphire_rapids_ec2_m7i_xlarge — other: analysis_uops_info_alder_lake_p, specs_intel
x86	intel.2022_raptor_cove_gen_13_14	benchmarks: intel_xeon_8559c_emerald_rapids_ec2_i7i_xlarge — other: specs_intel
x86	intel.2023_redwood_cove_ultra_1	benchmarks: intel_xeon_6975p_granite_rapids_ec2_m8i_xlarge — other: specs_intel

See CPU architecture scope for why these CPUs and no others.

The full, current key list can always be enumerated programmatically:

from counted_float import BuiltInData

>>> list(BuiltInData.get_flop_weights_dict())
['arm.v8_x.benchmarks.apple_m1_github_actions', ...]

Aggregation and missing data

A selection of these entries is combined into one set of weights (what get_builtin_flop_weights() does, and what produces the package's default weights) by hierarchical geometric-mean aggregation, imputing missing values at every level — so the incomplete spec-sheet and third-party entries don't bias the result. See How the final weights are computed for the full explanation.