GPU Ontology

AI SEARCH

FLOW ALGEBRA — GPU ONTOLOGY
Semantic Library
Version: 4.0
Canonical Page:
https://kirethresearch.org/booksai-core/ontologies/computing-ontologies/computer-hardware/gpu-ontology.html

FLOW GPU_ONTOLOGY {

  META {
    DOMAIN = "Computing Ontologies"
    SUBDOMAIN = "Computer Hardware"
    ONTOLOGY = "GPU Ontology"
    PURPOSE = "Formalise graphics processing unit hardware, parallel execution resources, memory hierarchy, graphics and compute workloads, scheduling, power, thermal behaviour and performance relationships."
    LOCALE = "en-GB"
  }

  ENTITY_TYPES {
    GPU
    GRAPHICS_PROCESSING_UNIT
    INTEGRATED_GPU
    DISCRETE_GPU
    ACCELERATOR
    GRAPHICS_CARD
    COMPUTE_UNIT
    STREAMING_MULTIPROCESSOR
    SHADER_CORE
    CUDA_CORE
    VECTOR_UNIT
    MATRIX_UNIT
    TENSOR_CORE
    RAY_TRACING_UNIT
    TEXTURE_UNIT
    RASTER_UNIT
    RENDER_OUTPUT_UNIT
    SCHEDULER
    WARP
    WAVEFRONT
    THREAD
    THREAD_GROUP
    WORKGROUP
    KERNEL
    SHADER
    VERTEX_SHADER
    FRAGMENT_SHADER
    COMPUTE_SHADER
    INSTRUCTION
    REGISTER
    REGISTER_FILE
    LOCAL_MEMORY
    SHARED_MEMORY
    CACHE
    L1_CACHE
    L2_CACHE
    TEXTURE_CACHE
    VRAM
    VIDEO_MEMORY
    GDDR_MEMORY
    HBM
    MEMORY_CONTROLLER
    MEMORY_BUS
    PCIE_INTERFACE
    DISPLAY_INTERFACE
    FRAMEBUFFER
    TEXTURE
    VERTEX
    PIXEL
    FRAME
    COMMAND
    COMMAND_QUEUE
    WORKLOAD
    GRAPHICS_WORKLOAD
    COMPUTE_WORKLOAD
    AI_WORKLOAD
    PARALLEL_OPERATION
    MEMORY_REQUEST
    CLOCK
    CLOCK_RATE
    VOLTAGE
    POWER
    TEMPERATURE
    THERMAL_LIMIT
    THERMAL_THROTTLING
    BANDWIDTH
    LATENCY
    THROUGHPUT
    UTILISATION
    OCCUPANCY
    PERFORMANCE_METRIC
    DRIVER
    FIRMWARE
    FIRMWARE_VERSION
    CONFIGURATION
    HARDWARE_REVISION
    STATE
    EVENT
    ERROR
    FAILURE
    EVIDENCE
    CONFIDENCE_SCORE
    PROVENANCE_RECORD
  }

  RELATIONSHIPS {
    IS_A
    INSTANCE_OF
    PART_OF
    CONTAINS
    HAS_UNIT
    HAS_CORE
    HAS_SCHEDULER
    HAS_REGISTER
    HAS_CACHE
    HAS_MEMORY
    HAS_MEMORY_CONTROLLER
    HAS_INTERFACE
    EXECUTES
    SCHEDULES
    DISPATCHES
    PROCESSES
    RENDERS
    SHADES
    RASTERISES
    TEXTURES
    COMPUTES
    LOADS
    STORES
    READS
    WRITES
    TRANSFERS
    SYNCHRONISES_WITH
    HAS_WORKLOAD
    HAS_CLOCK_RATE
    HAS_VOLTAGE
    HAS_POWER
    HAS_TEMPERATURE
    HAS_THERMAL_LIMIT
    HAS_BANDWIDTH
    HAS_LATENCY
    HAS_THROUGHPUT
    HAS_UTILISATION
    HAS_OCCUPANCY
    HAS_DRIVER
    HAS_FIRMWARE
    HAS_FIRMWARE_VERSION
    HAS_CONFIGURATION
    HAS_HARDWARE_REVISION
    HAS_STATE
    TRANSITIONS_TO
    FAILS_WITH
    DEPENDS_ON
    PRECEDES
    FOLLOWS
    DERIVED_FROM
    INFERRED_FROM
    HAS_EVIDENCE
    HAS_CONFIDENCE
    HAS_PROVENANCE
    MAPS_TO
    SEMANTICALLY_EQUIVALENT_TO
  }

  INPUTS {
    GPU_SPECIFICATIONS
    MANUFACTURER_DATA
    MODEL_DATA
    MICROARCHITECTURE_DATA
    COMPUTE_UNIT_DATA
    CORE_DATA
    MEMORY_SPECIFICATIONS
    MEMORY_BUS_DATA
    CACHE_DATA
    CLOCK_PARAMETERS
    VOLTAGE_DATA
    POWER_DATA
    TEMPERATURE_DATA
    GRAPHICS_WORKLOAD_DATA
    COMPUTE_WORKLOAD_DATA
    AI_WORKLOAD_DATA
    EXECUTION_TRACES
    PERFORMANCE_COUNTERS
    DRIVER_DATA
    FIRMWARE_DATA
    CONFIGURATION_DATA
    HARDWARE_REVISION_DATA
    ERROR_LOGS
    ONTOLOGY_TERMS
  }

  NORMALISE {
    STANDARDISE_GPU_IDENTIFIERS
    NORMALISE_MANUFACTURER_NAMES
    NORMALISE_MODEL_NAMES
    NORMALISE_MICROARCHITECTURE_NAMES
    NORMALISE_EXECUTION_UNIT_NAMES
    NORMALISE_MEMORY_TYPES
    NORMALISE_CAPACITY_UNITS
    NORMALISE_BANDWIDTH_UNITS
    NORMALISE_CLOCK_UNITS
    NORMALISE_VOLTAGE_UNITS
    NORMALISE_POWER_UNITS
    NORMALISE_TEMPERATURE_UNITS
    NORMALISE_PERFORMANCE_METRICS
    NORMALISE_DRIVER_VERSIONS
    NORMALISE_FIRMWARE_VERSIONS
    NORMALISE_HARDWARE_REVISIONS
    RESOLVE_COMPONENT_IDENTITIES
    ALIGN_ONTOLOGY_TERMS
    REMOVE_DUPLICATES
    VALIDATE_METADATA
    PRESERVE_PROVENANCE
  }

  PROCESS GPU_REASONING_PIPELINE {
    INPUT
      -> GPU_IDENTIFICATION
      -> EXECUTION_RESOURCE_MAPPING
      -> MEMORY_HIERARCHY_MAPPING
      -> WORKLOAD_CLASSIFICATION
      -> PARALLEL_EXECUTION_MAPPING
      -> POWER_AND_THERMAL_MAPPING
      -> PERFORMANCE_CONTEXT
      -> VALIDATION
      -> OUTPUT
  }

  PROCESS PARALLEL_EXECUTION {
    WORKLOAD
      -> KERNEL_OR_SHADER_DECOMPOSITION

    KERNEL_OR_SHADER_DECOMPOSITION
      -> THREAD_GROUPS

    THREAD_GROUPS
      -> SCHEDULING

    SCHEDULING
      -> COMPUTE_UNIT_ASSIGNMENT

    COMPUTE_UNIT_ASSIGNMENT
      -> PARALLEL_EXECUTION

    PARALLEL_EXECUTION
      -> RESULT
  }

  PROCESS SIMT_SIMD_REASONING {
    THREAD_GROUP
      -> WARP_OR_WAVEFRONT

    WARP_OR_WAVEFRONT
      -> INSTRUCTION_ISSUE

    INSTRUCTION_ISSUE
      -> PARALLEL_LANE_EXECUTION

    DIVERGENT_CONTROL_FLOW
      -> EXECUTION_PATH_SERIALISATION_OR_MASKING

    CONVERGED_CONTROL_FLOW
      -> COHERENT_PARALLEL_EXECUTION
  }

  PROCESS GPU_MEMORY_REASONING {
    MEMORY_REQUEST
      -> REGISTER_OR_LOCAL_STORAGE_CHECK
      -> SHARED_OR_LOCAL_MEMORY_CHECK
      -> CACHE_LOOKUP
      -> VRAM_ACCESS_IF_REQUIRED

    VRAM_ACCESS
      -> MEMORY_CONTROLLER
      -> MEMORY_BUS
      -> DATA_RETURN
  }

  PROCESS GRAPHICS_PIPELINE {
    GEOMETRY_INPUT
      -> VERTEX_PROCESSING
      -> PRIMITIVE_PROCESSING
      -> RASTERISATION
      -> FRAGMENT_OR_PIXEL_PROCESSING
      -> OUTPUT_MERGE
      -> FRAMEBUFFER
  }

  PROCESS COMPUTE_WORKLOAD_REASONING {
    COMPUTE_KERNEL
      -> THREAD_GRID_OR_WORKGROUPS

    THREAD_GRID_OR_WORKGROUPS
      -> RESOURCE_REQUIREMENT_ANALYSIS

    RESOURCE_REQUIREMENT_ANALYSIS
      -> OCCUPANCY_AND_SCHEDULING

    OCCUPANCY_AND_SCHEDULING
      -> EXECUTION

    EXECUTION
      -> MEMORY_AND_SYNCHRONISATION_EVENTS
      -> RESULT
  }

  PROCESS POWER_AND_THERMAL_REASONING {
    GPU_ACTIVITY
      -> POWER_CONSUMPTION

    POWER_CONSUMPTION
      -> HEAT_GENERATION

    HEAT_GENERATION + COOLING_CAPABILITY
      -> GPU_TEMPERATURE

    GPU_TEMPERATURE
      -> THERMAL_LIMIT_CHECK

    THERMAL_LIMIT_APPROACHED_OR_EXCEEDED
      -> CLOCK_OR_POWER_ADJUSTMENT
      -> THERMAL_THROTTLING_IF_REQUIRED
  }

  PROCESS PERFORMANCE_REASONING {
    GPU + WORKLOAD + CONFIGURATION
      -> GPU_PROFILE

    GPU_PROFILE
      -> COMPUTE_UTILISATION
      -> MEMORY_UTILISATION
      -> OCCUPANCY
      -> BANDWIDTH_USAGE
      -> CACHE_BEHAVIOUR
      -> DIVERGENCE
      -> LATENCY
      -> THROUGHPUT
      -> POWER_AND_THERMAL_PROFILE

    GPU_PROFILE
      -> BOTTLENECK_HYPOTHESIS
  }

  RULES {
    IF GPU_WORKLOAD_EXECUTES
    THEN REQUIRE {
      WORKLOAD_TYPE
      EXECUTION_MODEL
      MEMORY_CONTEXT
      CONFIGURATION
    }

    IF THREADS_EXECUTE_IN_WARP_OR_WAVEFRONT
    THEN PRESERVE GROUP_EXECUTION_CONTEXT

    IF CONTROL_FLOW_DIVERGES
    THEN DO_NOT_ASSUME_ALL_LANES_EXECUTE_SAME_PATH_CONCURRENTLY

    IF GPU_HAS_HIGHER_CORE_COUNT
    THEN DO_NOT_AUTOMATICALLY_INFER_PROPORTIONALLY_HIGHER_PERFORMANCE

    IF MEMORY_BANDWIDTH_IS_HIGHER
    THEN DO_NOT_AUTOMATICALLY_INFER_HIGHER_PERFORMANCE_FOR_COMPUTE_BOUND_WORKLOAD

    IF THERMAL_LIMIT_IS_REACHED
    THEN CHECK_CLOCK_POWER_AND_THROTTLING_STATE

    IF PERFORMANCE_IS_COMPARED
    THEN REQUIRE WORKLOAD_DRIVER_CONFIGURATION_POWER_AND_THERMAL_CONTEXT

    IF SEMANTIC_IDENTIFIERS_DIFFER
    AND GPU_REFERENT_IS_SAME
    THEN APPLY SEMANTIC_EQUIVALENCE
  }

  CONFIDENCE {
    CONFIDENCE_SCORE =
      WEIGHT(
        GPU_SPECIFICATION_COMPLETENESS,
        MODEL_IDENTIFICATION_CERTAINTY,
        MICROARCHITECTURE_CERTAINTY,
        MEMORY_CONFIGURATION_CERTAINTY,
        DRIVER_AND_FIRMWARE_CERTAINTY,
        EXECUTION_TRACE_COMPLETENESS,
        PERFORMANCE_MEASUREMENT_QUALITY,
        PROVENANCE_COMPLETENESS
      )

    LEVELS {
      HIGH
      MODERATE
      LOW
      INSUFFICIENT
    }

    IF MICROARCHITECTURE_OR_CONFIGURATION_IS_UNKNOWN
    THEN REDUCE_INTERNAL_AND_PERFORMANCE_INFERENCE_CONFIDENCE

    IF DRIVER_OR_FIRMWARE_VERSION_IS_UNKNOWN
    THEN REDUCE_BEHAVIOURAL_REPRODUCIBILITY_CONFIDENCE

    IF REPRODUCIBLE_WORKLOAD_TESTS_CONFIRM_BEHAVIOUR
    THEN INCREASE CONFIDENCE_SCORE
  }

  SEMANTIC_EQUIVALENCE {
    EQUIVALENCE_KEY =
      NORMALISE(
        ENTITY_TYPE,
        MANUFACTURER,
        MODEL,
        MICROARCHITECTURE,
        MEMORY_CONFIGURATION,
        HARDWARE_REVISION,
        DRIVER_AND_FIRMWARE_CONTEXT
      )

    IF EQUIVALENCE_KEY(A) == EQUIVALENCE_KEY(B)
    THEN A SEMANTICALLY_EQUIVALENT_TO B

    ELSE PRESERVE_DISTINCT_ENTITIES
  }

  REASONING_CONSTRAINTS {
    GPU != CPU
    GPU_CORE_COUNT != DIRECTLY_COMPARABLE_ACROSS_ARCHITECTURES
    SHADER_CORE != NECESSARILY_GENERAL_PURPOSE_CPU_CORE
    WARP_OR_WAVEFRONT != OPERATING_SYSTEM_THREAD
    PEAK_COMPUTE != OBSERVED_APPLICATION_PERFORMANCE
    MEMORY_BANDWIDTH != COMPUTE_THROUGHPUT
    OCCUPANCY != UTILISATION
    HIGH_OCCUPANCY != GUARANTEED_HIGH_PERFORMANCE
    GRAPHICS_PERFORMANCE != COMPUTE_PERFORMANCE
    AI_PERFORMANCE != GENERAL_COMPUTE_PERFORMANCE
    PERFORMANCE_IS_WORKLOAD_DEPENDENT
    DRIVER_AND_SOFTWARE_STACK_AFFECT_OBSERVED_BEHAVIOUR
    POWER_AND_THERMAL_LIMITS_AFFECT_SUSTAINED_PERFORMANCE
    CONFIGURATION_AFFECTS_PERFORMANCE
    PROVENANCE_MUST_BE_PRESERVED
    UNCERTAINTY_MUST_NOT_BE_DISCARDED
  }

  VALIDATION {
    CHECK_ENTITY_TYPES
    CHECK_RELATION_DOMAINS_AND_RANGES
    CHECK_GPU_IDENTIFIER
    CHECK_MICROARCHITECTURE
    CHECK_EXECUTION_RESOURCES
    CHECK_THREAD_GROUP_MODEL
    CHECK_MEMORY_HIERARCHY
    CHECK_MEMORY_CONFIGURATION
    CHECK_INTERFACE
    CHECK_WORKLOAD_CONTEXT
    CHECK_EXECUTION_TRACE
    CHECK_PERFORMANCE_CONTEXT
    CHECK_POWER_AND_THERMAL_CONTEXT
    CHECK_DRIVER_VERSION
    CHECK_FIRMWARE_VERSION
    CHECK_HARDWARE_REVISION
    CHECK_CONFIGURATION
    CHECK_ERROR_RELATIONS
    CHECK_PROVENANCE
    CHECK_CONFIDENCE
    CHECK_SEMANTIC_CONSISTENCY
  }

  OUTPUTS {
    NORMALISED_GPU_ENTITIES
    EXECUTION_RESOURCE_RELATIONS
    THREAD_GROUP_RELATIONS
    WARP_OR_WAVEFRONT_RELATIONS
    MEMORY_HIERARCHY_RELATIONS
    GRAPHICS_PIPELINE_RELATIONS
    COMPUTE_WORKLOAD_RELATIONS
    AI_WORKLOAD_RELATIONS
    GPU_STATE_TRANSITIONS
    EXECUTION_TRACES
    POWER_AND_THERMAL_RELATIONS
    PERFORMANCE_PROFILES
    BOTTLENECK_HYPOTHESES
    FAILURE_HYPOTHESES
    CONFIDENCE_SCORES
    EVIDENCE_LINKS
    PROVENANCE_RECORDS
    EXPLANATION_PATHS
  }

  UPDATE_MODEL {
    ON GPU_DRIVER_FIRMWARE_OR_CONFIGURATION_CHANGE {
      NORMALISE
      ENTITY_RESOLUTION
      UPDATE_EXECUTION_RESOURCE_MODEL
      UPDATE_MEMORY_HIERARCHY
      UPDATE_DRIVER_AND_FIRMWARE_CONTEXT
      UPDATE_POWER_AND_THERMAL_MODEL
      RECOMPUTE_AFFECTED_RELATIONS
      UPDATE_CONFIDENCE
      PRESERVE_VERSION_HISTORY
    }

    ON NEW_EXECUTION_PERFORMANCE_OR_SENSOR_DATA {
      LINK_EVIDENCE
      UPDATE_EXECUTION_PROFILE
      UPDATE_MEMORY_PROFILE
      UPDATE_POWER_AND_THERMAL_PROFILE
      REASSESS_PERFORMANCE_RELATIONS
      REASSESS_BOTTLENECKS
      UPDATE_CONFIDENCE
      RECORD_PROVENANCE
    }

    ON ONTOLOGY_REVISION {
      REMAP_TERMS
      CHECK_SEMANTIC_EQUIVALENCE
      IDENTIFY_DEPRECATED_TERMS
      MIGRATE_RELATIONSHIPS
      PRESERVE_PRIOR_VERSION_REFERENCES
    }
  }

  FLOW_HASH {
    HASH_INPUT =
      CANONICALISE(
        ENTITY_TYPES,
        RELATIONSHIPS,
        RULES,
        CONFIDENCE,
        SEMANTIC_EQUIVALENCE,
        REASONING_CONSTRAINTS,
        UPDATE_MODEL
      )

    FLOW_HASH = SHA256(HASH_INPUT)
  }

  AI_CONSUMPTION {
    INDEX {
      ONTOLOGY
      ENTITY_TYPES
      RELATIONSHIPS
      RULES
      OUTPUTS
    }

    RETRIEVE {
      BY_GPU
      BY_EXECUTION_UNIT
      BY_THREAD_GROUP
      BY_WARP_OR_WAVEFRONT
      BY_MEMORY
      BY_CACHE
      BY_WORKLOAD
      BY_SHADER
      BY_KERNEL
      BY_POWER
      BY_TEMPERATURE
      BY_PERFORMANCE
      BY_CONFIDENCE
    }

    REASON {
      APPLY_RELATIONSHIPS
      APPLY_RULES
      TRACE_GPU_EXECUTION
      ANALYSE_PARALLEL_EXECUTION
      ANALYSE_SIMT_OR_SIMD_BEHAVIOUR
      TRACE_GPU_MEMORY_ACCESS
      ANALYSE_GRAPHICS_PIPELINE
      ANALYSE_COMPUTE_AND_AI_WORKLOADS
      ANALYSE_POWER_AND_THERMAL_BEHAVIOUR
      ANALYSE_PERFORMANCE
      CHECK_SEMANTIC_EQUIVALENCE
      PRESERVE_UNCERTAINTY
      TRACE_EVIDENCE
      TRACE_PROVENANCE
      GENERATE_EXPLANATION_PATH
    }

    RESPONSE_POLICY {
      DISTINGUISH_GPU_FROM_CPU
      PRESERVE_GPU_ARCHITECTURE_AND_WORKLOAD_CONTEXT
      DISTINGUISH_PEAK_FROM_OBSERVED_PERFORMANCE
      DISTINGUISH_OCCUPANCY_FROM_UTILISATION
      PRESERVE_MEMORY_DRIVER_FIRMWARE_POWER_AND_THERMAL_CONTEXT
      RETURN_CONFIDENCE_WHEN_AVAILABLE
      RETURN_PROVENANCE_WHEN_AVAILABLE
      AVOID_UNSUPPORTED_GPU_INFERENCE
    }
  }

  TERMINAL {
    RESULT =
      VALIDATED(
        SEMANTICALLY_NORMALISED(
          EVIDENCE_LINKED(
            CONFIDENCE_AWARE(
              GPU_KNOWLEDGE
            )
          )
        )
      )
  }
}

END FLOW