Result bundle reference (run_summary.json, schema_version 4)

Every declared block and field of a result bundle’s run_summary.json, as helx show and helx validate bundle read it. Fields with default none are left out of the file when the run did not produce them. run_summary.schema.json next to this page is the same contract as JSON Schema.

Generated by scripts/render_references.py from the models; do not edit by hand.

RunSummaryV4

run_summary.json: the whole bundle, one case per file.

Field	Type	Default	Meaning
schema_version	one of 4	required	Always 4 for this contract.
generated_at	datetime	required	When the bundle was written; timezone-aware, so two bundles can be ordered without guessing the writer’s clock.
config	block ConfigRecord	required	The config the run was made from.
provenance	block Provenance	required	Everything the numbers rest on: software, platform, target and values status.
contract	block ContractBlock	required	The fixed promises of this result.
run	block RunBlock	required	The case: names, memory mode, platform mode, runtime and NPU size.
source	block SourceBlock	required	Where the model came from.
vela	block VelaBlock or null	none	Vela’s compile-time figures; absent when Vela was bypassed.
fvp	block FvpBlock	required	The FVP stage and its measurements.
memory	block MemoryBlock	required	Placement and footprint.
tensors	block TensorEvidence or null	none	Opt-in fixed input bytes, output specifications and reference comparison context.
measurements	block Measurements or null	none	Indexed measured invocations and explicit population statistics; absent before execution.
vela_energy	block EnergyBlock or null	none	Experimental energy, power and efficiency predicted from Vela’s estimates; never a measurement.
energy_errors	dict[one of measured_fvp_pmu, vela_compile_time_estimate, str] or null	none	Energy estimates that were attempted and failed, keyed by basis (measured_fvp_pmu, vela_compile_time_estimate) with the reason. Absent when every attempted estimate succeeded; a basis listed here has no block above.
per_layer	block PerLayerBlock or null	none	Per-layer attribution artifacts, when requested.
per_segment	block PerSegmentBlock or null	none	Fenced per-segment measurements, when requested.
artifacts	block ArtifactsBlock	required	The files the bundle carries.

ConfigRecord

The config the run was made from: where it was, and the name of its copy in the bundle.

Field	Type	Default	Meaning
path	str	required	Resolved absolute path of the config the run was made from.
copied_as	str	required	File name of the copy of that config inside the bundle.

Provenance

Everything a number rests on: software, platform definition, target, and the values’ status.

Field	Type	Default	Meaning
software	block SoftwareVersions	required	Versions of the tools that produced the numbers.
platform	block PlatformProvenance or null	none	The platform definition and the SoC revision behind it.
target	block TargetProvenance or null	none	The realization of the SoC the measurement is attributed to.
values	block ValuesProvenance or null	none	How far the stored hardware values behind the run can be trusted.

SoftwareVersions

The tools that produced the numbers.

Field	Type	Default	Meaning
helx	str or null	none	helx version that wrote the bundle.
helia_npu	str or null	none	helia-npu version (SoC definitions, Vela system config, PMU programs).
vela	str or null	none	Vela compiler version.
python	str	required	Python interpreter version.
fvp_subsystem	str or null	none	Corstone FVP subsystem the runner was built for (corstone_300 or corstone_320), when the FVP stage ran.
fvp_home	str or null	none	The FVP install root (FVP_HOME) the model binary was taken from, so its exact version can be found.

PlatformProvenance

Which platform definition, at which revision, from which kind of source.

Field	Type	Default	Meaning
name	str	required	Platform the run was attributed to: a shipped name or the path the config gave.
revision	str or null	none	The platform history’s current revision label, or the SoC revision when no wrapper history is recorded. Energy-profile identity is recorded separately; compare both when comparing runs.
origin	str or null	none	golden, derived or custom, so a what-if number cannot pass as silicon.
soc_id	str or null	none	Canonical id of the SoC definition the platform rests on, e.g. atomiq110.
soc_revision	str or null	none	Which drop or stepping of the SoC the timing describes: the label with its date, e.g. v8 (2026-09-22).
soc_stage	str or null	none	Where that SoC revision sits in the part’s life: spec, rtl, silicon or released.
energy_profile_id	str or null	none	Selected platform energy-profile id; absent for inline energy or run.energy_model overrides.
energy_profile_revision	str or null	none	Revision of the selected energy characterization.
energy_profile_sha256	str or null	none	SHA-256 of the captured energy-profile file bytes.

TargetProvenance

Which realization of the SoC the measurement is attributed to, and whether its timing stands for the part.

Field	Type	Default	Meaning
id	str or null	none	helia-npu target id with its kind, e.g. atomiq110-fvp-corstone-320 (simulator).
timing_is_representative	bool or null	none	True when cycle counts on this target stand for the SoC’s timing; false for FPGA drops at other clock ratios and for simulators.

ValuesProvenance

How far the stored hardware values behind this run can be trusted.

Field	Type	Default	Meaning
status	one of confirmed, fitted, unverified or null	none	The least trusted status among the hardware values the run depends on: confirmed, fitted or unverified.
fitted	list[str] or null	none	Values tuned to reproduce RTL or FPGA measurements, by dotted path; absent when there are none.
unverified	list[str] or null	none	Values that are a specification figure, estimate or placeholder, by dotted path; absent when there are none.

ContractBlock

What the bundle promises: one result per config, and the bases every energy number rests on.

Field	Type	Default	Meaning
result	one of one_config_one_result	required	The bundle holds exactly one result for one config.
placement	str	required	Always exact: the placement recorded is the one compiled and run, not a nearest match.
energy_model	str	required	Shape of the energy model: NPU core energy per clocked cycle plus per-bit transfer energy.
core_energy_basis	str	required	Formula for core energy: npu_total_cycles times the reference core energy times the mode scaling.
memory_energy_basis	str	required	Formula for memory energy: completed data beats times per-bit energy times beat width times (1 + logic_overhead).

RunBlock

What was run, on what.

Field	Type	Default	Meaning
case_name	str	required	The case directory name: the model name and the memory mode joined by two underscores.
model	str	required	Model name from the config, or the source file’s stem.
memory_mode	str	required	The Vela [Memory_Mode.*] section compiled: Sram_Only, Shared_Sram, Dedicated_Sram, or const_X__arena_Y__cache_Z for a custom placement.
fvp_memory_mode	str	required	The standard mode the runtime executed; differs from memory_mode only for a custom placement.
mode	str or null	none	Platform operating mode, i.e. the NPU clock: ULP or HP on AT110.
runtime	str or null	none	Inference runtime linked into the runner (tflite_micro, helia_rt or helia_aot), when the FVP stage built.
runtime_backend	str or null	none	Retained heliaRT backend: helia (heliaCORE); unset when the field does not apply.
npu_mac_units	int	required	The NPU’s MAC array size (MAC units per cycle), e.g. 256 for an Ethos-U85-256.
npu_clock_hz	float or null	none	The NPU clock of the mode the run used, Hz; measured cycles divided by it are the inference time.

SourceBlock

Where the model came from and where it was cached.

Field	Type	Default	Meaning
given	str	required	The source as the config or the command line spelled it.
uri	str	required	The resolved source URI (file, http, https or s3).
scheme	str	required	URI scheme of the source: file, http, https or s3.
cached_path	str	required	Local path of the model file the run compiled.
cache_mode	str	required	Whether downloaded models survive the run: persistent (kept under the cache root) or ephemeral.
cache_root	str	required	Root directory of the model cache.
bytes	int	required	Size of the model file, bytes.
sha256	str or null	none	Observed source-model byte digest, distinct from the Vela/runner model digest.

VelaBlock

What Vela reported for the compile. Cycles and bytes are Vela’s estimates, not measurements.

Field	Type	Default	Meaning
profile	str	required	helx’s label for the Vela run: the SoC id and mode, the same string as system_config.
system_config	str	required	The [System_Config.*] section of the generated ini: SoC id and mode, e.g. atomiq110_HP.
accelerator	str or null	none	Vela’s accelerator configuration, e.g. Ethos_U85_256.
memory_mode	str or null	none	Vela’s own memory_mode column: the [Memory_Mode.*] section it compiled with.
arena_cache_configured_bytes	int	required	The –arena-cache-size handed to Vela, bytes.
arena_cache_kib	float or null	none	Vela’s own arena_cache_size column, KiB.
core_clock_hz	float or null	none	NPU clock Vela assumed, Hz.
nn_macs	float or null	none	MAC operations in one inference as Vela counted them.
inference_time_s	float or null	none	Vela’s own inference time, seconds: cycles_total over core_clock_hz, so it carries Vela’s memory model.
cycles_total	float or null	none	Vela’s total for one inference: per operation the slowest of its NPU and memory-access estimates, summed. Its Its memory estimate can exceed hardware costs at HP; compare cycles_npu with each sample’s active cycles.
cycles_npu	float or null	none	Vela’s estimate of NPU busy cycles; compare with measurements.samples[i].cycles.active.
cycles_sram_access	float or null	none	Vela’s estimate of the cycles bound by port 0 (Vela Sram) access.
cycles_dram_access	float or null	none	Vela’s estimate of the cycles bound by port 1 (Vela Dram) access.
sram_total_bytes	float or null	none	Bytes Vela expects to move over port 0 in one inference.
dram_total_bytes	float or null	none	Bytes Vela expects to move over port 1 in one inference.
sram_bandwidth_gbps	float or null	none	Port 0 bandwidth Vela assumed, GB/s.
dram_bandwidth_gbps	float or null	none	Port 1 bandwidth Vela assumed, GB/s.
sram_used_kib	float or null	none	Port 0 memory Vela’s plan occupies (arena plus weight cache), KiB.
dram_used_kib	float or null	none	Port 1 memory Vela’s plan occupies, KiB.
on_chip_flash_kib	float or null	none	Vela’s OnChipFlash usage, KiB; always 0 because helx assigns nothing to that area.
off_chip_flash_kib	float or null	none	Vela’s OffChipFlash usage, KiB; always 0 because helx assigns nothing to that area.
weights_original_bytes	float or null	none	Weight bytes in the input model.
weights_encoded_bytes	float or null	none	Weight bytes after Vela’s encoding and compression.
command_stream_bytes	int or null	none	Size of the Ethos-U command stream Vela emitted, bytes.

FvpBlock

The FVP stage: how far it got, and the measurements when it ran.

Field	Type	Default	Meaning
stage	one of disabled, configured-built, ran	required	How far the FVP stage got: disabled, configured-built, or ran.
requested_acquisition	block AcquisitionPolicy or null	none	Resolved build-time request; null warmup means automatic, not zero. Distinct from observed counts.
enabled	bool	required	Whether the config enabled the FVP stage.
run_enabled	bool or null	none	Whether the config asked to run the runner, not only build it.
subsystem	str or null	none	Corstone FVP subsystem: corstone_300 (its U55 or U65 model, as vela.accelerator says) or corstone_320.
build_type	str or null	none	Runner build type: Debug or Release.
host_transport	str or null	none	How the host collected the runner’s report from the FVP, from the subsystem’s profile.
beat_bits	int or null	none	Width of one AXI data beat on this SoC, bits, so traffic_beats can be turned into bytes.
sram_burst_bytes	int or null	none	AXI SRAM-port burst split the runner’s driver was built with (U85 AXI_LIMIT_SRAM_MAX_BEATS, U55/U65 AXI_LIMIT0/1_MAX_BEATS_BYTES), bytes: the platform’s, or a -DHELX_AXI_SRAM_MAX_BEATS in fvp.cmake_args.
ext_burst_bytes	int or null	none	AXI EXT-port burst split the runner’s driver was built with (U85 AXI_LIMIT_EXT_MAX_BEATS, U55/U65 AXI_LIMIT2/3_MAX_BEATS_BYTES), bytes: the platform’s, or a -DHELX_AXI_EXT_MAX_BEATS in fvp.cmake_args.
sram_port_hash	int or null	none	Address mask that picked the NPU’s SRAM port in the run (the FVP’s sram_hash0): the platform’s, or a -C ...sram_hash0 in fvp.fvp_args; absent when the FVP default (0x100, address bit 8) applied, and on build-only runs, since the FVP takes it at launch. U85 only.
contention	dict[str, block FvpContention] or null	none	Emulated application traffic the run was configured with (run.contention), keyed by memory name; absent when there was none or nothing saw it.
bus_trace	block FvpBusTrace or null	none	The NPU bus trace summary; absent unless fvp.bus_trace was enabled and the FVP ran.
exit_code	int or null	none	Exit code of the FVP process.
warmup_invokes	int or null	none	Inferences the runner ran, uncounted, before the measured one: 1 when the model has a CALL_ONCE init subgraph (resource-variable initialisation runs inside the first inference), else 0 by default. Explicit carry warmups can select another count; AOT reports its actual count, default 0.

AcquisitionPolicy

Build-time lifecycle controls; None warmup preserves the runner’s automatic choice.

Field	Type	Default	Meaning
state	one of carry, single	carry	carry retains variables across invokes; single permits one invoke on the initialized instance.
warmup_invokes	int or null	none	Uncounted warmup invokes. Unset: carry warms up once when CALL_ONCE is present; single uses zero.
measured_invokes	int	1	Measured invokes on one initialized instance; single requires one.

FvpContention

One memory’s emulated application traffic (run.contention): what was asked for and what its port’s timing adapter was programmed with as a result.

Field	Type	Default	Meaning
port	int	required	AXI port it was on in this run: 0 (SRAM class) or 1 (EXT).
share_taken	float	required	Fraction of the memory’s bandwidth other traffic took, 0 to below 1.
extra_latency_ns	float	required	Queueing latency added to reads and writes, ns.
rate_window_cycles	int	required	Timing-adapter window over which the reduced rate is enforced, NPU cycles (the run’s own unless the entry set one).
applies_to_vela	bool	required	Whether Vela compiled for the contended timing too (otherwise only the FVP saw it).
bandwidth	float	required	Beats per NPU cycle the port’s adapter was programmed with after contention (rounded to whole beats per window, as the adapter is built; not reflected if fvp.cmake_args overrides that adapter’s BWCAP or PULSE).
read_latency	int	required	Read latency after contention, NPU cycles.
write_latency	int	required	Write latency after contention, NPU cycles.

FvpBusTrace

The NPU bus trace (fvp.bus_trace): every AXI transaction the NPU made on the FVP, summarised. The timeline, transfer sizes and a Perfetto trace are in the artifacts; the raw trace only when the config kept it.

Field	Type	Default	Meaning
status	one of pass, fail	required	pass when every check agrees.
transactions	int	required	AXI transactions traced.
raw_trace_bytes	int	required	Size of the trace the FVP wrote, bytes.
step_cycles	float or null	none	NPU cycles per FVP step: the FVP advances the NPU in steps and stamps each step’s transactions with one time, so this is the timeline’s finest resolution.
fvp_time_scale	float or null	none	FVP simulated seconds per second of NPU time, from the traced steps and the PMU cycle count (about 20 at a 500 MHz NPU clock on FVP 11.27). The timeline is in NPU time.
npu_segments	int	required	NPU invocations the trace shows: more than one when the model runs operators on the CPU between NPU segments. The timeline leaves the CPU time between them out, as the PMU leaves it out of the NPU’s cycles.
timeline_rows	int	required	Rows in the timeline CSV: one per FVP step, or whole steps merged so there are at most 2000.
sram_port_hash	int	required	Address mask whose parity assigned each SRAM-class transaction to SRAM0 or SRAM1 (the FVP default 0x100 when the run set none).
stripe_bytes	int or null	none	The SRAM interface changes at every multiple of this many bytes: the hash’s lowest set bit; absent for a zero hash, which puts every SRAM-class transaction on SRAM0.
sram_transfers_crossing_stripe	int	required	SRAM-class transactions that cross a multiple of stripe_bytes (with a one-bit hash, as on AT110, that is a transaction starting in one interface’s stripe and ending in the other’s); 0 when every burst fits a stripe.
interfaces	dict[str, block FvpBusTraceInterface]	required	Per interface: sram0, sram1, ext.
activity	block FvpBusTraceActivity or null	none	Which interfaces were active together; absent without both ports’ rates.
checks	dict[str, block FvpBusTraceCheck]	required	The trace against the PMU (data beats per port and direction, SRAM0 reads), against the driver build (largest transfer within the burst), and against itself (time never goes back).

FvpBusTraceInterface

What one NPU memory interface carried over the inference, from the bus trace.

Field	Type	Default	Meaning
read_transfers	int	required	Read transactions.
write_transfers	int	required	Write transactions.
read_bytes	int	required	Bytes read.
write_bytes	int	required	Bytes written.
read_beats	int	required	Read data beats, counting a partial beat as a whole one as the PMU does.
write_beats	int	required	Write data beats on the same basis.
largest_transfer_bytes	int	required	The largest single transaction, bytes: at most the burst the driver was built with.
mean_busy_fraction	float or null	none	Beats moved over the beats the port’s timing adapter allows in the inference’s NPU cycles (the fvp.port_utilization basis, with each interface’s writes counted exactly); absent without the port’s rate.
peak_busy_fraction	float or null	none	The busiest row of the timeline on the same basis. It can exceed 1 when a row is shorter than the port’s timing-adapter rate window, since the adapter caps the beats per window, not per cycle.

FvpBusTraceActivity

How the inference’s cycles divide by which interfaces were active (moving at least 5% of their rate) in each timeline row.

Field	Type	Default	Meaning
sram_only_fraction	float	required	SRAM0 or SRAM1 active, EXT not.
ext_only_fraction	float	required	EXT active, neither SRAM interface.
sram_and_ext_fraction	float	required	EXT and at least one SRAM interface active together.
idle_fraction	float	required	No interface active.

FvpBusTraceCheck

One comparison of the trace with an independent source.

Field	Type	Default	Meaning
status	one of pass, fail	required	pass when the trace agrees.
trace	int	required	The figure from the trace.
expected	int	required	What it is checked against: the PMU’s count, the burst the driver was built with, or 0 for something that should not happen.

MemoryBlock

Footprint: where the regions live and how many bytes each thing takes.

Field	Type	Default	Meaning
const_placement	str or null	none	Physical memory holding the weights.
arena_placement	str or null	none	Physical memory holding the tensor arena (activations).
cache_placement	str or null	none	Physical memory holding Vela’s weight cache.
runtime_model_bytes	int	required	Size of the model the runtime loaded (Vela’s output, or the original when Vela was bypassed), bytes.
tensor_arena_configured_bytes	int	required	Tensor arena size the runner was built with, bytes.
tensor_arena_used_bytes	int or null	none	Arena bytes the runtime reported using.
aot	block AotArenas or null	none	The heliaAOT module’s arenas, when that engine linked the runner.

AotArenas

The heliaAOT module’s arenas, when that engine linked the runner.

Field	Type	Default	Meaning
constant_bytes	int	required	Cold constants and generated staged-constant source storage, including source guards, bytes.
mutable_bytes	int	required	Writable runtime arenas: scratch, persistent and staged-constant destinations, bytes.
total_bytes	int	required	Constant plus mutable arena, bytes.
constant_regions	int	required	Number of constant regions the module declares.
mutable_regions	int	required	Number of writable runtime regions, including staged-constant destinations.

TensorEvidence

Build-bound input bytes and comparison context for opt-in tensor capture.

Field	Type	Default	Meaning
request_sha256	str	required	Canonical tensor-request digest reported by the compiled runner.
inputs_explicit	bool	required	Whether the request supplied input bytes rather than selecting implicit zeros.
inputs	tuple	required	Effective input bytes refilled before every invoke, including warmups.
output_specs	tuple	required	Fixed output metadata read from the actual runner model.
references	tuple	()	One reference set per measured sample, or empty when only capturing outputs.
comparison	ComparisonPolicy	required	Numerical tolerance and NaN policy in stored-value units.

TensorArtifact

A fixed tensor whose raw bytes are retained in the result manifest.

Field	Type	Default	Meaning
spec	TensorSpec	required	Exact model tensor index, name, dtype and fixed shape.
path	str	required	Bundle-relative binary artifact, with size/hash supplied by the manifest.

ReferenceArtifacts

Reference output bytes for one measured invocation, with caller-declared provenance.

Field	Type	Default	Meaning
tensors	tuple	required	All reference outputs in model-index order.
origin	one of files, host_litert, producer	required	Declared source of these reference bytes.
description	str	required	Identifying reference context supplied by the caller.

Measurements

A complete ordered sequence and statistics derived from that sequence.

Field	Type	Default	Meaning
acquisition	block AcquisitionFacts	required	The runner-observed lifecycle and measured count.
samples	tuple	required	All measured invocations in their execution order.
statistics	block MeasurementStatistics	required	Population statistics recomputed from every retained sample.

AcquisitionFacts

Runner-reported lifecycle for the measured invocation sequence, excluding warmup counters.

Field	Type	Default	Meaning
state	one of carry, single	carry	carry retains variables across invokes; single permits one invoke on the initialized instance.
warmup_invokes	int	required	Completed unmeasured invokes before the measured sequence.
measured_invokes	int	1	Measured invokes on one initialized instance; single requires one.
schema_id	one of helx.acquisition	helx.acquisition	Runner acquisition protocol identifier.
version	one of 1	1	Runner acquisition protocol version.
call_once_present	bool or null	required	Interpreter CALL_ONCE detection; null for generated AOT initialization.
initialization	one of before_measurement, interpreter_allocate_and_first_invoke	required	Where runtime state initialization occurred relative to measurement.

MeasurementSample

Counters and derived estimates for one indexed measured invocation, excluding warmups.

Field	Type	Default	Meaning
index	int	required	Zero-based measured invocation index; warmups are excluded.
pmu_profile	one of default, ext_stall, default_ext_rd_stall, dv_align, port_stall, sram_stall, sram_transactions	required	The event selection programmed for this invocation.
output_tensors	tuple	()	Opt-in retained output bytes, in model-index order.
verification	block TensorVerification or null	none	Optional numerical comparison against this sample’s retained reference set.
outputs	tuple	()	Ordered output byte signatures captured after the measured invoke.
cycles	block FvpCycles or null	none	NPU cycle counters from the PMU.
traffic_beats	block FvpTrafficBeats or null	none	AXI data beats per port and direction from the PMU.
traffic_transactions	block FvpTrafficTransactions or null	none	Accepted aggregate SRAM transactions in the same PMU window as traffic_beats; sram_transactions profile.
port_utilization	block FvpPortUtilization or null	none	Average busy fraction of each NPU memory interface over the inference, from traffic_beats and the cycles; absent when the run had no counters, no cycles or no timing-adapter rates.
latency_s	float or null	none	NPU total cycles at the mode’s NPU clock: the latency one inference took on the NPU, seconds.
utilization_fraction	float or null	none	MACs performed over MACs the array could have performed: nn_macs / (npu_total_cycles * npu_mac_units), 0 to 1.
pmu_overflow	int or null	none	The NPU’s PMU overflow flag as the runner reported it; nonzero means a counter wrapped and the counts are short.
energy	block EnergyBlock or null	none	Energy derived from this invocation’s PMU counters.
sanity	block SanityBlock or null	none	Validation checks for this invocation’s counters.
energy_error	str or null	none	Why configured energy could not be estimated for this invocation.

TensorVerification

Numerical comparison against the matching retained reference set; separate from integrity.

Field	Type	Default	Meaning
reference_index	int	required	Index of the reference set in the bundle’s tensor evidence.
matched	bool	required	Whether all elements met the recorded comparison policy.
tensors	tuple	required	Mismatch counts and exact maximum finite absolute error for each output.

FvpCycles

NPU cycle counters from the PMU; the stall and MAC counters depend on the PMU profile.

Field	Type	Default	Meaning
total	int or null	none	NPU_TOTAL: cycles the NPU was clocked during the inference.
active	int or null	none	NPU_ACTIVE: cycles the NPU was executing work.
idle	int or null	none	NPU_IDLE: cycles the NPU was clocked but idle.
elapsed	int or null	none	NPU elapsed cycles between the driver starting and finishing the inference.
mac_active	int or null	none	MAC_ACTIVE: cycles the MAC array was busy (dv_align and port_stall profiles).
mac_stalled_by_ib	int or null	none	Cycles the MAC array stalled waiting on the input buffer (dv_align profile).
mac_stalled_by_w_or_acc	int or null	none	Cycles the MAC array stalled on weights or the accumulator (dv_align profile).
sram_rd_stall_limit	int or null	none	SRAM read outstanding-limit stall count (port_stall/sram_stall); U85 sums AXI ports, not exclusive lost time.
sram_rd_tran_req_stalled	int or null	none	SRAM read address-request backpressure (port_stall/sram_stall); U85 sums AXI ports, may overlap NPU work.
sram_wr_stall_limit	int or null	none	SRAM write outstanding-limit stall count (sram_stall); summed over AXI ports, not exclusive lost time.
sram_wr_tran_req_stalled	int or null	none	SRAM write address-request backpressure (sram_stall); summed over AXI ports, may overlap NPU work/read stalls. Does not count write-data backpressure; SRAM_WR_DATA_BEAT_STALLED is a separate event.
ext_rd_stall_limit	int or null	none	Cycles port 1 reads were stalled by the outstanding-transaction limit (ext_stall and port_stall profiles).
ext_rd_tran_req_stalled	int or null	none	Cycles a port 1 read request was stalled by the fabric (ext_stall, default_ext_rd_stall and port_stall profiles).
cpu_wall	int or null	none	Cortex-M DWT cycle count across the inference; only the heliaAOT runner emits it.
mcu_systick_ticks	int or null	none	SysTick ticks across the inference; only a runner built with HELX_MCU_WALL_PROBE emits it.

FvpTrafficBeats

AXI data beats per port and direction, as the PMU counted them; fvp.beat_bits is the beat width.

Field	Type	Default	Meaning
sram_rd	int or null	none	Read data beats received on port 0 (Arm AXI_SRAM class).
sram_wr	int or null	none	Write data beats written on port 0.
sram0_rd	int or null	none	Read data beats on the first SRAM-class interface alone (Ethos-U85).
ext_rd	int or null	none	Read data beats received on port 1 (Arm AXI_EXT class).
ext_wr	int or null	none	Write data beats written on port 1.
ext0_rd	int or null	none	Read data beats on the first EXT-class interface (Ethos-U85).
ext1_rd	int or null	none	Read data beats on the second EXT-class interface (Ethos-U85).

FvpTrafficTransactions

Accepted SRAM-class AXI transactions, summed across its interfaces in one invocation window.

Field	Type	Default	Meaning
sram_rd	int or null	none	Accepted read requests, in transactions; null means not counted, zero means counted with no requests.
sram_wr	int or null	none	Accepted write requests, in transactions; not a completed-write count or a queue-depth measurement.

FvpPortUtilization

How busy each NPU memory interface was, averaged over the inference: data beats moved over the beats the port’s timing adapter allows in the NPU cycles the inference took. 1 = saturated at this clock; normally 0 to 1, not clipped (a figure above 1 means the counters exceed the programmed rate). A figure the counters cannot support is absent, and basis says why.

Field	Type	Default	Meaning
sram0	float or null	none	First SRAM-class interface: SRAM0 read beats plus its share of SRAM writes, over total cycles x the port 0 beats per cycle (each SRAM interface is capped at that rate).
sram1	float or null	none	Second SRAM-class interface: the remaining SRAM read beats plus its share of writes, on the same basis.
ext	float or null	none	EXT (port 1): EXT read and write beats over total cycles x the port 1 beats per cycle, treated as one interface (AT110 traffic uses one; basis says when both EXT interfaces carried reads).
sram0_read_share	float or null	none	SRAM0’s share of SRAM read beats: 0.5 is an even split between the two SRAM interfaces.
saturated	list[str] or null	none	Interfaces at or above 98% of their rate (SRAM0, SRAM1, EXT). Their figures stay exact here, and can exceed 1 at very low rates because the timing adapter admits whole transactions; people-facing views show them capped at 100% and marked saturated.
basis	str or null	none	Which counters the figures come from, what was assumed (helx’s PMU programs split SRAM reads, not writes, by interface, so writes are assumed to split like reads), and why any figure is absent.

EnergyBlock

Experimental energy by mode; basis identifies the simulator counter or compiler estimate source.

Field	Type	Default	Meaning
basis	one of measured_fvp_pmu, vela_compile_time_estimate	required	measured_fvp_pmu (from simulator PMU counters) or vela_compile_time_estimate (from Vela’s estimates). Neither identifies measured device energy or qualifies the energy model’s accuracy.
modes	dict[str, block EnergyMode]	required	Energy by platform mode name. The measured block holds only the mode that ran; the Vela-estimate block projects every mode from the same counters.
active_mode	str or null	none	The mode the run executed in.
model	str or null	none	The energy model file, relative to the bundle, when it was kept.

EnergyMode

Experimental energy, power and efficiency estimates for one inference in one operating mode.

Field	Type	Default	Meaning
total_mj	float	required	Energy of one inference in this mode, mJ.
core_uj	float	required	NPU core energy, uJ.
sram_uj	float	required	Transfer energy over port 0 (Vela Sram class), uJ.
ext_uj	float	required	Transfer energy over port 1 (Vela Dram class), uJ.
pj_per_mac	float or null	none	Total energy over MAC operations, pJ per MAC; null for a network that does no MACs.
gmacs_per_w	float or null	none	Efficiency: giga-MACs per second per watt; null for a network that does no MACs.
power_mw	float or null	none	Average power over the inference time, mW. In energy that is measurements.samples[i].latency_s; in vela_energy it is vela.cycles_npu over vela.core_clock_hz, not vela.inference_time_s (which carries Vela’s memory model), except for a graph Vela gives no NPU cycles, where it falls back to vela.inference_time_s.
share_core_fraction	float	required	Core share of the total energy, 0 to 1.
share_sram_fraction	float	required	Port 0 transfer share of the total energy, 0 to 1.
share_ext_fraction	float	required	Port 1 transfer share of the total energy, 0 to 1.

SanityBlock

Outcome of the post-run consistency checks on the PMU report.

Check names are those of helia_estimator.runtime.pmu_report.SanityCheck.

Field	Type	Default	Meaning
status	one of pass, fail	required	pass when every check held, else fail.
checks	dict[str, bool]	required	Each consistency check by name and whether it held.

MeasurementStatistics

Explicit population summaries; counter means and variances retain exact integer ratios.

Field	Type	Default	Meaning
counters	dict[str, IntegerStatistics or null]	required	Exact summaries keyed by cycle or traffic field; null if any sample lacks the field.
latency_s	FloatStatistics or null	none	NPU latency population statistics in seconds, excluding warmups.
energy_total_mj	dict[str, FloatStatistics or null]	required	Modeled total-energy population statistics in millijoules for each mode.

PerLayerBlock

Per-layer attribution artifacts, relative to the bundle.

Field	Type	Default	Meaning
per_layer_csv	str or null	none	Per-layer cycle attribution from Vela’s verbose performance output.
vela_verbose_log	str or null	none	Vela’s stdout carrying the per-layer table.
per_original_op_csv	str or null	none	Per-layer figures folded back onto the original model’s operators.
model_explorer_overlay	str or null	none	JSON overlay for Model Explorer with per-operator figures.
unavailable_reason	str or null	none	Why calibrated attribution was omitted; raw Vela diagnostics remain available when present.

PerSegmentBlock

Fenced per-segment measurement artifacts, relative to the bundle.

Field	Type	Default	Meaning
segments	int	required	Number of segments the fences cut the model into.
tensor_arena_bytes	int	required	Arena size of the fenced model, bytes.
per_segment_csv	str or null	none	Measured cycles per segment.
fenced_model_tflite	str or null	none	The model with ETHOS_U_STREAM_SPLIT fences inserted.
fenced_vela_tflite	str or null	none	The fenced model after Vela.
fenced_vela_summary_csv	str or null	none	Vela’s summary for the fenced compile.
fenced_per_layer_csv	str or null	none	Per-layer attribution of the fenced compile.
estimated_vs_measured_csv	str or null	none	Vela’s estimate against the measured cycles, per segment.
npu_invocations	int or null	none	NPU invocations measured, one row each in per_segment_csv. Fewer than segments when some fence segments hold only operators Vela left on the CPU (they start no invocation) or Vela removed fences after operators it folded into constants (the segments either side run as one invocation). Each row’s segment_index is the first fence segment it covers and its op range spans all of them.
pmu_profile	one of default, ext_stall, default_ext_rd_stall, dv_align, port_stall, sram_stall, sram_transactions or null	none	The PMU program the fenced run counted with (fvp.pmu_profile): which counters per_segment_csv carries, e.g. the read-stall and MAC-active cycles of port_stall. All per-segment figures come from this one run.
counters_check	one of pass, fail or null	none	pass when every counter the run reports sums over the segments to the run’s total, so no segment line was lost or misread; fail names the counters in counters_not_summing, including any only some segments carry, and all of them when the fenced log had no segment lines. Absent when there was nothing to compare (no run totals).
counters_summed	list[str] or null	none	The counters compared and found equal, by their run-total names (npu_total_cycles is the segments’ ccnt; npu_sram_rd_tran_req_stalled their sram_rd_tran_req_stalled_cycles).
counters_not_summing	list[str] or null	none	Counters, by run-total name, whose segment sum differs from the total or that some segments lack; absent when all agree.
bus_trace_csv	str or null	none	Each NPU invocation’s traffic from the fenced run’s bus trace (fvp.bus_trace), keyed as per_segment_csv is: the FVP steps it spans, transfers, bytes and beats per interface (SRAM0, SRAM1, EXT) and direction, and whether its beats equal its PMU counts. Absent without a trace, or when the trace’s segments could not be paired with the invocations (bus_trace_segments says how many it found).
bus_trace_check	one of pass, fail or null	none	pass when the run printed segment lines, the trace splits into one segment per NPU invocation, each one’s beats equal that invocation’s PMU beat counts, and the whole-trace checks pass (beats against the run’s totals, time order, transfers within the burst). The trace’s segments come from its own timestamps, so this checks the split and the counts independently. Absent without fvp.bus_trace.
bus_trace_segments	int or null	none	NPU segments the fenced trace splits into; equal to npu_invocations when it is read right.
bus_trace_invocations_not_matching	list[int] or null	none	Invocations (npu_invocation) whose beats the trace does not reproduce: every one when the segment counts differ. Absent when all match.
bus_trace_checks_failed	list[str] or null	none	Whole-trace checks of the fenced trace that failed, by their fvp.bus_trace.checks names, or trace_read when the trace could not be read (the PMU results stand without it); absent when all pass.
bus_trace_raw	str or null	none	The fenced run’s raw bus trace, when the config kept it (fvp.bus_trace.keep_raw).
layers_csv	str or null	none	The fenced run layer by layer, one row per NPU invocation: its original operators, measured cycles and counters, traffic per port (from the bus trace when one was taken), busy and stall shares, Vela’s estimate, and its likely limit with the reason.
layers_perfetto_json	str or null	none	The same layers as a timeline to open at ui.perfetto.dev: end to end in fenced NPU time, a lane per likely limit, and a counter track per measured figure.
cycles_by_limit	dict[one of ext_bandwidth, sram_bandwidth, sram_latency, npu_engine, not_counted, int] or null	none	Fenced NPU-active cycles of the layers on each likely-limit lane. The lanes are a heuristic: the busier port when at least 60% busy, else SRAM reads queued in at least 30% of the cycles (SRAM latency, mostly overlapped by other work), else the NPU engine when both ports and the SRAM read stalls were measured, else not_counted.

ArtifactsBlock

Paths of the files the bundle carries, relative to the bundle root; a file outside the bundle is not named.

Field	Type	Default	Meaning
root	str	required	The artifacts directory, relative to the bundle.
vela_tflite	str or null	none	The Vela-compiled model.
vela_summary_csv	str or null	none	Vela’s summary CSV.
vela_config	str or null	none	The generated Vela ini the compile used.
energy_model	str or null	none	The energy model YAML the run used.
runner_axf	str or null	none	The built runner image.
runner_input_tflite	str or null	none	The model the runner was built with.
fvp_stdout_log	str or null	none	The FVP process’s stdout.
fvp_stderr_log	str or null	none	The FVP process’s stderr.
fvp_uart_log	str or null	none	The runner’s UART output, including the PMU report.
runner_build_log	str or null	none	The runner build’s CMake and compiler output, kept out of the terminal.
bus_trace_timeline_csv	str or null	none	Bus trace timeline: bytes and busy fraction per interface per slice of the inference.
bus_trace_transfers_csv	str or null	none	Bus trace transfer sizes: transactions per size, interface and direction.
bus_trace_perfetto_json	str or null	none	Bus trace timeline as a Chrome trace with one lane per interface; open it at ui.perfetto.dev.
bus_trace_raw	str or null	none	The FVP’s raw bus trace, when the config kept it (fvp.bus_trace.keep_raw).