|
| 1 | +schema: eliza.ai_eda.ai_workload_manifest.v1 |
| 2 | +updated: 2026-05-21 |
| 3 | +claim_boundary: ai_workload_manifest_only_no_benchmark_performance_or_release_claim |
| 4 | +policy: |
| 5 | + release_use_allowed: false |
| 6 | + performance_claim_allowed: false |
| 7 | + fallback_ops_must_be_reported: true |
| 8 | + e1_npu_claim_requires_zero_fallback_proof: true |
| 9 | + calibrated_power_or_timing_required_for_tops_per_watt: true |
| 10 | +workloads: |
| 11 | + - id: mobile-smoke-tflite-cpu |
| 12 | + category: runtime_benchmark |
| 13 | + benchmark_plan_ref: tflite_cpu |
| 14 | + source: benchmarks/models/generate_mobile_smoke_tflite.py |
| 15 | + license_status: redistributable_generated_fixture |
| 16 | + artifacts: |
| 17 | + - path: benchmarks/models/mobile_smoke.tflite |
| 18 | + sha256: 6bb01f97e01d782b139eecf32fecada42c08a8edc8d2148da4dcf63b22dd972a |
| 19 | + role: model |
| 20 | + input_shape: |
| 21 | + input: [1, 32] |
| 22 | + quantization: fp32 |
| 23 | + expected_ops: |
| 24 | + - FULLY_CONNECTED |
| 25 | + - RELU |
| 26 | + - SOFTMAX |
| 27 | + fallback_ops: [] |
| 28 | + runtime_path: benchmark_model_cpu |
| 29 | + golden_output_tolerance: |
| 30 | + kind: smoke_runtime_only |
| 31 | + max_abs_error: null |
| 32 | + evidence_status: ready_for_cpu_smoke_when_benchmark_model_exists |
| 33 | + blockers: |
| 34 | + - benchmark_model_binary_required_for_real_runtime_numbers |
| 35 | + - id: mobile-smoke-tflite-e1-npu |
| 36 | + category: runtime_benchmark |
| 37 | + benchmark_plan_ref: tflite_e1_npu |
| 38 | + source: benchmarks/models/generate_mobile_smoke_tflite.py |
| 39 | + license_status: redistributable_generated_fixture |
| 40 | + artifacts: |
| 41 | + - path: benchmarks/models/mobile_smoke.tflite |
| 42 | + sha256: 6bb01f97e01d782b139eecf32fecada42c08a8edc8d2148da4dcf63b22dd972a |
| 43 | + role: model |
| 44 | + - path: docs/benchmarks/capabilities/e1_npu_nnapi.proof.template.json |
| 45 | + role: proof_template |
| 46 | + input_shape: |
| 47 | + input: [1, 32] |
| 48 | + quantization: fp32 |
| 49 | + expected_ops: |
| 50 | + - FULLY_CONNECTED |
| 51 | + - RELU |
| 52 | + - SOFTMAX |
| 53 | + fallback_ops: [] |
| 54 | + runtime_path: benchmark_model_nnapi_e1_npu |
| 55 | + requires_zero_fallback_proof: true |
| 56 | + golden_output_tolerance: |
| 57 | + kind: cpu_vs_nnapi_profiled_smoke |
| 58 | + max_abs_error: 0.001 |
| 59 | + evidence_status: blocked_until_e1_nnapi_proof |
| 60 | + blockers: |
| 61 | + - missing_e1_npu_nnapi_proof_json |
| 62 | + - missing_target_benchmark_model_nnapi_transcript |
| 63 | + - id: npu-arch-sim-open-2028 |
| 64 | + category: architecture_sim |
| 65 | + benchmark_plan_ref: npu_arch_sim_open_2028 |
| 66 | + source: benchmarks/sim/run_npu_scale_sim.py |
| 67 | + license_status: local_project_fixture |
| 68 | + artifacts: |
| 69 | + - path: benchmarks/sim/run_npu_scale_sim.py |
| 70 | + role: simulator |
| 71 | + - path: docs/spec-db/npu-2028-target.yaml |
| 72 | + role: target_contract |
| 73 | + input_shape: |
| 74 | + sweep: deterministic_gemm_conv_attention |
| 75 | + quantization: int8 |
| 76 | + expected_ops: |
| 77 | + - GEMM |
| 78 | + - CONV |
| 79 | + - ATTENTION |
| 80 | + fallback_ops: [] |
| 81 | + runtime_path: python_scale_sim |
| 82 | + golden_output_tolerance: |
| 83 | + kind: deterministic_json_metrics |
| 84 | + max_abs_error: 0.0 |
| 85 | + evidence_status: modeled_only_not_release_benchmark |
| 86 | + blockers: |
| 87 | + - modeled_tops_not_calibrated_to_e1_rtl_or_silicon |
| 88 | + - id: npu-arch-sim-sota-2028 |
| 89 | + category: architecture_sim |
| 90 | + benchmark_plan_ref: npu_arch_sim_sota_2028 |
| 91 | + source: benchmarks/sim/run_npu_scale_sim.py |
| 92 | + license_status: local_project_fixture |
| 93 | + artifacts: |
| 94 | + - path: benchmarks/sim/run_npu_scale_sim.py |
| 95 | + role: simulator |
| 96 | + - path: docs/spec-db/competitor-2028-target.md |
| 97 | + role: comparison_target |
| 98 | + input_shape: |
| 99 | + sweep: deterministic_gemm_conv_attention |
| 100 | + quantization: int8 |
| 101 | + expected_ops: |
| 102 | + - GEMM |
| 103 | + - CONV |
| 104 | + - ATTENTION |
| 105 | + fallback_ops: [] |
| 106 | + runtime_path: python_scale_sim |
| 107 | + golden_output_tolerance: |
| 108 | + kind: deterministic_json_metrics |
| 109 | + max_abs_error: 0.0 |
| 110 | + evidence_status: modeled_only_not_release_benchmark |
| 111 | + blockers: |
| 112 | + - modeled_tops_not_calibrated_to_e1_rtl_or_silicon |
| 113 | + - id: e1-npu-timeloop-arch |
| 114 | + category: dataflow_mapping |
| 115 | + source: benchmarks/sim/run_npu_timeloop.py |
| 116 | + license_status: local_project_fixture_plus_optional_external_tool |
| 117 | + artifacts: |
| 118 | + - path: benchmarks/sim/run_npu_timeloop.py |
| 119 | + role: runner |
| 120 | + - path: benchmarks/sim/configs/e1_npu_timeloop_arch.yaml |
| 121 | + role: architecture_config |
| 122 | + input_shape: |
| 123 | + mapping: local_architecture_config |
| 124 | + quantization: int8 |
| 125 | + expected_ops: |
| 126 | + - GEMM |
| 127 | + - CONV |
| 128 | + fallback_ops: [] |
| 129 | + runtime_path: timeloop_accelergy_optional_backend |
| 130 | + golden_output_tolerance: |
| 131 | + kind: blocked_until_backend_installed |
| 132 | + max_abs_error: null |
| 133 | + evidence_status: blocked_until_timeloop_backend_present |
| 134 | + blockers: |
| 135 | + - timeloop_accelergy_not_required_for_default_local_gate |
| 136 | + - mapping_outputs_not_calibrated_to_e1_rtl |
| 137 | + - id: e1-stablehlo-tiny-mlp |
| 138 | + category: compiler_lowering |
| 139 | + source: compiler/runtime/test_e1_npu_tiny_mlp_e2e.py |
| 140 | + license_status: local_project_fixture |
| 141 | + artifacts: |
| 142 | + - path: compiler/runtime/test_e1_npu_tiny_mlp_e2e.py |
| 143 | + role: e2e_test |
| 144 | + - path: compiler/runtime/e1_npu_stablehlo.py |
| 145 | + role: lowering |
| 146 | + - path: compiler/runtime/e1_npu_runtime.py |
| 147 | + role: runtime |
| 148 | + input_shape: |
| 149 | + input: tiny_mlp_test_fixture |
| 150 | + quantization: int8 |
| 151 | + expected_ops: |
| 152 | + - stablehlo.dot_general |
| 153 | + - stablehlo.add |
| 154 | + - stablehlo.relu |
| 155 | + fallback_ops: [] |
| 156 | + runtime_path: python_e1_npu_runtime_sim |
| 157 | + golden_output_tolerance: |
| 158 | + kind: test_assertions |
| 159 | + max_abs_error: 0.0 |
| 160 | + evidence_status: local_test_fixture_only |
| 161 | + blockers: |
| 162 | + - not_a_target_hardware_benchmark |
| 163 | + - id: e1-quantization-int4-awq-gptq |
| 164 | + category: quantization |
| 165 | + source: compiler/quantization |
| 166 | + license_status: local_project_fixture |
| 167 | + artifacts: |
| 168 | + - path: compiler/quantization/awq_int4.py |
| 169 | + role: awq_calibrator |
| 170 | + - path: compiler/quantization/gptq_int4.py |
| 171 | + role: gptq_calibrator |
| 172 | + - path: compiler/quantization/tests/test_awq_int4_mlp_e2e.py |
| 173 | + role: e2e_test |
| 174 | + input_shape: |
| 175 | + input: local_mlp_calibration_fixture |
| 176 | + quantization: int4 |
| 177 | + expected_ops: |
| 178 | + - eliza.group_scaled_int4_matmul |
| 179 | + - tflite.matmul |
| 180 | + fallback_ops: [] |
| 181 | + runtime_path: python_quantization_fixture |
| 182 | + golden_output_tolerance: |
| 183 | + kind: quantized_vs_float_fixture |
| 184 | + max_abs_error: 0.25 |
| 185 | + evidence_status: local_test_fixture_only |
| 186 | + blockers: |
| 187 | + - no_representative_product_calibration_set |
| 188 | + - no_e1_runtime_power_latency_capture |
| 189 | + - id: e1-quantization-fp8-sparse24 |
| 190 | + category: quantization |
| 191 | + source: compiler/quantization |
| 192 | + license_status: local_project_fixture |
| 193 | + artifacts: |
| 194 | + - path: compiler/quantization/fp8_e4m3_calibration.py |
| 195 | + role: fp8_calibrator |
| 196 | + - path: compiler/quantization/sparse_2_4.py |
| 197 | + role: sparse_calibrator |
| 198 | + - path: compiler/quantization/tests/test_calibrators.py |
| 199 | + role: calibrator_tests |
| 200 | + input_shape: |
| 201 | + input: local_calibration_fixture |
| 202 | + quantization: fp8_e4m3_and_sparse_2_4 |
| 203 | + expected_ops: |
| 204 | + - eliza.fp8_matmul |
| 205 | + - eliza.sparse_2_4_matmul |
| 206 | + fallback_ops: [] |
| 207 | + runtime_path: python_quantization_fixture |
| 208 | + golden_output_tolerance: |
| 209 | + kind: quantized_vs_float_fixture |
| 210 | + max_abs_error: 0.25 |
| 211 | + evidence_status: local_test_fixture_only |
| 212 | + blockers: |
| 213 | + - no_representative_product_calibration_set |
| 214 | + - no_e1_runtime_power_latency_capture |
0 commit comments