{ "report_date": "2026-08-28", "workload": { "m": 8192, "n": 8192, "k": 8192, "macs": 549755813888, "conventional_operations": 1099511627776, "vendor_counting_convention": "one multiply-accumulate equals two operations", "comparison_contract": "INT8 inputs with INT32 accumulation/output unless a row says otherwise", "input_bytes": 134217728, "int32_output_bytes": 268435456, "minimum_logical_device_traffic_bytes": 402653184 }, "sutro_model": { "energy_coefficient_J_per_score_unit": 1e-15, "score_unit": "one source operand times one abstract address-grid step", "omitted_terms": [ "destination writes", "arithmetic", "memory capacity and physical hierarchy", "HBM transfers", "control and clocking", "leakage and time" ], "schedules": [ { "name": "current record cubically scaled lower bound", "score": 8898635366400, "energy_J": 0.0088986353664, "qualification": "not a generated 8K program; assumes fixed average distance" }, { "name": "retuned sa_cache", "tile_i": 256, "tile_j": 128, "tile_candidates_evaluated": 196, "search_scope": "all ordered pairs of the 14 divisors of 8192 within the literal generalized sa_cache family", "validation": "the same formula reproduces sa_cache(16,8,4)=73602 and naive(16)=340704", "paid_reads": 2205465706496, "score": 118953083721334, "energy_J": 0.118953083721334, "fJ_per_conventional_operation": 108.187199404105, "movement_only_TOPS_per_W_equivalent": 9.24323769825, "improvement_over_naive": 196.727009216 }, { "name": "retuned square tiled", "tile": 64, "score": 317312291631114, "energy_J": 0.317312291631114 }, { "name": "fixed 4x4 tiled", "score": 2136290066409474, "energy_J": 2.136290066409474 }, { "name": "naive baseline", "score": 23401284397481984, "energy_J": 23.401284397481984 } ], "retuned_sa_cache_derivation": { "distance": "d(a)=ceil(sqrt(a))", "range_sum": "D(l,h)=F(h)-F(l-1)", "prefix_sum": "F(x)=q(q+1)(4q-1)/6+(x-q^2)(q+1), q=floor(sqrt(x))", "score_formula": "N^3*d(1)+N^2*(N-1)*d(2)+(N^3/Tj)*D(sB)+(N^3/(Ti*Tj))*D(sC)+(N/Tj)*D(A)+(N/Ti)*D(B)+D(C)", "address_regions": { "sA": [1, 1], "TMP": [2, 2], "sB": [3, 130], "sC": [131, 32898], "A": [32899, 67141762], "B": [67141763, 134250626], "C": [134250627, 201359490] }, "reads_per_cell": { "sA": 549755813888, "TMP": 549688705024, "sB": 4294967296, "sC": 16777216, "A": 64, "B": 32, "C": 1 }, "score_units_by_region": { "sA": 549755813888, "TMP": 1099377410048, "sB": 4514010628096, "sC": 66997983379456, "A": 23475391008000, "B": 21448665777152, "C": 867899704694, "total": 118953083721334 }, "next_tile_candidates": [ {"tile_i": 128, "tile_j": 128, "score": 121105737911286}, {"tile_i": 256, "tile_j": 64, "score": 121681290169654} ], "qualification": "exact closed-form evaluation for this schedule family; not a proof over all algorithms and no 8K IR is materialized" }, "retuned_sa_cache_breakdown_J": { "bulk_A_loads": 0.0234753910, "bulk_B_loads": 0.0214486658, "C_exit": 0.0008678997, "near_ALU_A_stage": 0.0005497558, "B_stage": 0.0045140106, "accumulator_stage": 0.0010993774, "output_tile_stage": 0.0669979834 }, "endpoint_sensitivity": { "formula": "2.5e-15 * (S + 160*R)", "S": 118953083721334, "R": 2205465706496, "energy_J": 1.179568991901735 } }, "nvidia_estimates": [ { "sku": "B200 Blackwell", "dense_INT8_TOPS": 4500, "power_W": 1000, "power_scope": "maximum configurable TDP", "ideal_time_s": 0.0002443359172835556, "energy_J": 0.2443359172835556, "source": "https://dam-cdn.nvd.orangelogic.com/AssetLink/y441155802qub41q118b2852i557jem5.pdf", "evidence": "peak-throughput/TDP estimate; not measured" }, { "sku": "B300 Blackwell Ultra", "dense_INT8_TOPS": 153.5, "power_W": 1100, "power_scope": "maximum configurable TDP", "ideal_time_s": 0.007162942200495114, "energy_J": 7.879236420544625, "source": "https://dam-cdn.nvd.orangelogic.com/AssetLink/1k0p832eq8r5ca0u5383ie5o4tp3bst1.pdf", "evidence": "peak-throughput/TDP estimate; not measured" }, { "sku": "Rubin", "dense_INT8_TOPS": 250, "power_W": 1800, "power_scope": "datasheet application-comparison assumption, not formally labeled TDP", "ideal_time_s": 0.004398046511104, "energy_J": 7.916483719987201, "source": "https://www.nvidia.com/en-us/data-center/vera-rubin-nvl72/", "power_source": "https://dam-cdn.nvd.orangelogic.com/AssetLink/v5rf2icnf86o26e464tf6djn23r8ibhe.pdf", "evidence": "preliminary peak-throughput/assumed-power estimate; not measured" } ], "blackwell_b200_precision_estimates": [ { "precision": "NVFP4", "stored_bits_per_input_element": 4, "dense_throughput_TOPS": 9000, "power_W": 1000, "ideal_time_s": 0.0001221679586417778, "energy_J": 0.1221679586417778, "minimum_input_bytes_excluding_scales_and_output": 67108864, "semantics": "block-scaled four-bit inputs; conversion, scale metadata, and higher-precision output are omitted" }, { "precision": "INT8", "stored_bits_per_input_element": 8, "dense_throughput_TOPS": 4500, "power_W": 1000, "ideal_time_s": 0.0002443359172835556, "energy_J": 0.2443359172835556, "minimum_logical_bytes_with_INT32_output": 402653184 }, { "precision": "TF32", "stored_bits_per_input_element": 32, "dense_throughput_TOPS": 1100, "power_W": 1000, "ideal_time_s": 0.0009995560252509091, "energy_J": 0.9995560252509091, "minimum_logical_bytes_with_FP32_output": 805306368, "semantics": "FP32 storage and accumulation with reduced-precision TF32 multiplication" }, { "precision": "FP32", "stored_bits_per_input_element": 32, "dense_throughput_TOPS": 75, "power_W": 1000, "ideal_time_s": 0.014660155037013333, "energy_J": 14.660155037013332, "minimum_logical_bytes_with_FP32_output": 805306368, "semantics": "literal four-byte FP32 path" } ], "blackwell_b300_precision_estimates": [ { "precision": "NVFP4", "stored_bits_per_input_element": 4, "dense_throughput_TOPS": 14000, "power_W": 1100, "ideal_time_s": 0.00007853654484114286, "energy_J": 0.08639019932525714, "qualification": "individual-GPU table rounds dense throughput to 14 PFLOPS; 8-GPU aggregate implies 13.5 PFLOPS per GPU and 0.0895898363373 J" }, { "precision": "TF32", "stored_bits_per_input_element": 32, "dense_throughput_TOPS": 1100, "power_W": 1100, "ideal_time_s": 0.0009995560252509091, "energy_J": 1.0995116277760002, "semantics": "FP32 storage and accumulation with reduced-precision TF32 multiplication" }, { "precision": "FP32", "stored_bits_per_input_element": 32, "dense_throughput_TOPS": 75, "power_W": 1100, "ideal_time_s": 0.014660155037013333, "energy_J": 16.126170540714664, "semantics": "literal four-byte FP32 path" } ], "formula_for_peak_power_rows": "energy_J = conventional_operations / (dense_TOPS * 1e12) * power_W", "semantic_warning": "The repository verifier uses unbounded Python integers; the score is value-independent, but the INT8 table is not semantic equivalence." }