1+ {
2+ "device_name" : " NVIDIA RTX 4000 Ada Generation" ,
3+ "kernel_name" : " vector_add" ,
4+ "problem_size" : 80000000 ,
5+ "tune_params_keys" : [
6+ " block_size_x" ,
7+ " nvml_gr_clock"
8+ ],
9+ "tune_params" : {
10+ "block_size_x" : [
11+ 128 ,
12+ 192 ,
13+ 256 ,
14+ 320 ,
15+ 384 ,
16+ 448 ,
17+ 512 ,
18+ 576 ,
19+ 640 ,
20+ 704 ,
21+ 768 ,
22+ 832 ,
23+ 896 ,
24+ 960 ,
25+ 1024
26+ ],
27+ "nvml_gr_clock" : [
28+ 1080
29+ ]
30+ },
31+ "objective" : " time" ,
32+ "cache" : {
33+ "128,1080": {"block_size_x": 128, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23700487666378292, "time": 3.165401004254818, "times": [3.2092161178588867, 3.1691839694976807, 3.1702721118927, 3.1651840209960938, 3.1774721145629883, 3.161087989807129, 3.1631360054016113, 3.167232036590576, 3.1651840209960938, 3.1733760833740234, 3.1651840209960938, 3.1723520755767822, 3.1539199352264404, 3.1692800521850586, 3.1723520755767822, 3.1733760833740234, 3.1600639820098877, 3.1580159664154053, 3.1600639820098877, 3.16211199760437, 3.1723520755767822, 3.1498239040374756, 3.161087989807129, 3.185663938522339, 3.1703040599823, 3.1641600131988525, 3.151871919631958, 3.1549439430236816, 3.155967950820923, 3.1487998962402344, 3.1498239040374756, 3.1599678993225098], "compile_time": 555.5126052349806, "verification_time": 0, "benchmark_time": 106.16594180464745, "f": 1080.0, "energy": 0.23700487666378292, "GFLOPS/W": 0.3375457970575376, "strategy_time": 0, "framework_time": 0, "timestamp": "2026-06-04 07:40:27.255674+00:00"},
34+ "192,1080": {"block_size_x": 192, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23637132654533696, "time": 3.1701670140028, "times": [3.2030720710754395, 3.1744000911712646, 3.1590399742126465, 3.1631360054016113, 3.16211199760437, 3.1641600131988525, 3.1836159229278564, 3.1487998962402344, 3.16211199760437, 3.1651840209960938, 3.1631360054016113, 3.1723520755767822, 3.171328067779541, 3.1795198917388916, 3.1631360054016113, 3.1795198917388916, 3.1712958812713623, 3.16211199760437, 3.171328067779541, 3.1744000911712646, 3.175424098968506, 3.1824638843536377, 3.1723520755767822, 3.1682560443878174, 3.1692800521850586, 3.1763200759887695, 3.1723520755767822, 3.1723520755767822, 3.156991958618164, 3.1774721145629883, 3.1600639820098877, 3.1682560443878174], "compile_time": 588.7568122707307, "verification_time": 0, "benchmark_time": 106.2352117151022, "f": 1080.0, "energy": 0.23637132654533696, "GFLOPS/W": 0.3384505268436427, "strategy_time": 0, "framework_time": 2.051988150924444, "timestamp": "2026-06-04 07:40:27.952794+00:00"},
35+ "256,1080": {"block_size_x": 256, "nvml_gr_clock": 1080, "temperature": 45.43287037037037, "core_freq": 1080.0, "ps_energy": 0.2358061059384572, "time": 3.1730150282382965, "times": [3.2235519886016846, 3.176448106765747, 3.1723520755767822, 3.180511951446533, 3.155967950820923, 3.16211199760437, 3.166111946105957, 3.1733760833740234, 3.1733760833740234, 3.167232036590576, 3.1733760833740234, 3.181567907333374, 3.187648057937622, 3.1692800521850586, 3.1692800521850586, 3.1661760807037354, 3.1856000423431396, 3.1651840209960938, 3.1825919151306152, 3.1723520755767822, 3.1948800086975098, 3.1682560443878174, 3.1682560443878174, 3.16211199760437, 3.1733760833740234, 3.166208028793335, 3.1723520755767822, 3.176448106765747, 3.1703040599823, 3.156991958618164, 3.1590399742126465, 3.1641600131988525], "compile_time": 589.3192640505731, "verification_time": 0, "benchmark_time": 106.42562713474035, "f": 1080.0, "energy": 0.2358061059384572, "GFLOPS/W": 0.3392617832418603, "strategy_time": 0, "framework_time": 1.879106741398573, "timestamp": "2026-06-04 07:40:28.650491+00:00"},
36+ "320,1080": {"block_size_x": 320, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23672766817357727, "time": 3.1728129982948303, "times": [3.195904016494751, 3.1703040599823, 3.1825599670410156, 3.167232036590576, 3.1682560443878174, 3.1660799980163574, 3.1590399742126465, 3.1703040599823, 3.154911994934082, 3.1763200759887695, 3.1712958812713623, 3.16211199760437, 3.1723520755767822, 3.1651840209960938, 3.167232036590576, 3.1640639305114746, 3.1824638843536377, 3.176448106765747, 3.180543899536133, 3.1703040599823, 3.185663938522339, 3.175424098968506, 3.1682560443878174, 3.1723520755767822, 3.147775888442993, 3.1846399307250977, 3.167232036590576, 3.166208028793335, 3.1876800060272217, 3.1784958839416504, 3.1928319931030273, 3.180543899536133], "compile_time": 590.9995380789042, "verification_time": 0, "benchmark_time": 106.25777393579483, "f": 1080.0, "energy": 0.23672766817357727, "GFLOPS/W": 0.3379410637430903, "strategy_time": 0, "framework_time": 1.7580576241016388, "timestamp": "2026-06-04 07:40:29.349540+00:00"},
37+ "384,1080": {"block_size_x": 384, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23602203720225295, "time": 3.1753220185637474, "times": [3.2153279781341553, 3.176448106765747, 3.1794559955596924, 3.2040960788726807, 3.18668794631958, 3.1774721145629883, 3.1682240962982178, 3.1784958839416504, 3.1549439430236816, 3.171231985092163, 3.176448106765747, 3.1692800521850586, 3.1692800521850586, 3.1702721118927, 3.166208028793335, 3.1712958812713623, 3.1733760833740234, 3.1590399742126465, 3.1723520755767822, 3.1774721145629883, 3.1733760833740234, 3.1846399307250977, 3.1723520755767822, 3.161087989807129, 3.181567907333374, 3.180543899536133, 3.166208028793335, 3.1733760833740234, 3.1784958839416504, 3.176448106765747, 3.1744959354400635, 3.1703040599823], "compile_time": 590.503228828311, "verification_time": 0, "benchmark_time": 106.34980583563447, "f": 1080.0, "energy": 0.23602203720225295, "GFLOPS/W": 0.33895140025185905, "strategy_time": 0, "framework_time": 1.859939657151699, "timestamp": "2026-06-04 07:40:30.048287+00:00"},
38+ "448,1080": {"block_size_x": 448, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23547623911725496, "time": 3.175483025610447, "times": [3.201024055480957, 3.1917760372161865, 3.180543899536133, 3.1825919151306152, 3.1774721145629883, 3.1744000911712646, 3.1784958839416504, 3.175424098968506, 3.180543899536133, 3.1692800521850586, 3.166208028793335, 3.1784958839416504, 3.171328067779541, 3.1744000911712646, 3.167232036590576, 3.171328067779541, 3.16211199760437, 3.1744000911712646, 3.181504011154175, 3.1733760833740234, 3.1723520755767822, 3.166208028793335, 3.1703040599823, 3.161087989807129, 3.1682560443878174, 3.1744000911712646, 3.1703040599823, 3.1887359619140625, 3.176448106765747, 3.175424098968506, 3.1825919151306152, 3.17740797996521], "compile_time": 591.0170101560652, "verification_time": 0, "benchmark_time": 106.32044123485684, "f": 1080.0, "energy": 0.23547623911725496, "GFLOPS/W": 0.33973703801241767, "strategy_time": 0, "framework_time": 1.506708562374115, "timestamp": "2026-06-04 07:40:30.747164+00:00"},
39+ "512,1080": {"block_size_x": 512, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23548762194821526, "time": 3.176128014922142, "times": [3.2020480632781982, 3.1784958839416504, 3.1692800521850586, 3.1825919151306152, 3.191807985305786, 3.1887359619140625, 3.1836159229278564, 3.1733760833740234, 3.1733760833740234, 3.181567907333374, 3.175424098968506, 3.1744000911712646, 3.1600639820098877, 3.1938560009002686, 3.1651840209960938, 3.167232036590576, 3.1733760833740234, 3.1641600131988525, 3.1723520755767822, 3.175424098968506, 3.167232036590576, 3.161087989807129, 3.1877119541168213, 3.167232036590576, 3.1600639820098877, 3.1744000911712646, 3.171328067779541, 3.1897599697113037, 3.181567907333374, 3.1795198917388916, 3.1744000911712646, 3.175424098968506], "compile_time": 591.0639478825033, "verification_time": 0, "benchmark_time": 106.31107306107879, "f": 1080.0, "energy": 0.23548762194821526, "GFLOPS/W": 0.33972061604831333, "strategy_time": 0, "framework_time": 1.6791722737252712, "timestamp": "2026-06-04 07:40:31.446290+00:00"},
40+ "576,1080": {"block_size_x": 576, "nvml_gr_clock": 1080, "temperature": 45.454861111111114, "core_freq": 1080.0, "ps_energy": 0.2353335454896639, "time": 3.1713550239801407, "times": [3.218303918838501, 3.1703040599823, 3.1631360054016113, 3.171328067779541, 3.1846399307250977, 3.155967950820923, 3.167232036590576, 3.1692800521850586, 3.166208028793335, 3.1703040599823, 3.180543899536133, 3.1651840209960938, 3.1641600131988525, 3.1600639820098877, 3.1703040599823, 3.1744000911712646, 3.1641600131988525, 3.171328067779541, 3.1744000911712646, 3.171328067779541, 3.1651840209960938, 3.166208028793335, 3.1744000911712646, 3.1846399307250977, 3.1682560443878174, 3.1631360054016113, 3.1692800521850586, 3.1733760833740234, 3.1733438968658447, 3.1703040599823, 3.1774721145629883, 3.1651840209960938], "compile_time": 585.2512437850237, "verification_time": 0, "benchmark_time": 106.32186336442828, "f": 1080.0, "energy": 0.2353335454896639, "GFLOPS/W": 0.3399430363127457, "strategy_time": 0, "framework_time": 1.5787328593432903, "timestamp": "2026-06-04 07:40:32.139477+00:00"},
41+ "640,1080": {"block_size_x": 640, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.23499261774443087, "time": 3.170297011733055, "times": [3.1938560009002686, 3.191807985305786, 3.1825919151306152, 3.1691839694976807, 3.1846399307250977, 3.1651840209960938, 3.1641600131988525, 3.1692800521850586, 3.1703040599823, 3.16211199760437, 3.1682560443878174, 3.1600639820098877, 3.1682560443878174, 3.166208028793335, 3.1744000911712646, 3.1795198917388916, 3.1703040599823, 3.1784958839416504, 3.152895927429199, 3.1631360054016113, 3.166208028793335, 3.1568639278411865, 3.1744000911712646, 3.1703040599823, 3.171328067779541, 3.1641600131988525, 3.1600639820098877, 3.1733760833740234, 3.1692800521850586, 3.1723520755767822, 3.161087989807129, 3.175424098968506], "compile_time": 588.1906021386385, "verification_time": 0, "benchmark_time": 106.17024963721633, "f": 1080.0, "energy": 0.23499261774443087, "GFLOPS/W": 0.34043622632862874, "strategy_time": 0, "framework_time": 1.7240559682250023, "timestamp": "2026-06-04 07:40:32.835594+00:00"},
42+ "704,1080": {"block_size_x": 704, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.2342263412780191, "time": 3.171537011861801, "times": [3.200000047683716, 3.176448106765747, 3.180543899536133, 3.171328067779541, 3.1651840209960938, 3.1723520755767822, 3.1887359619140625, 3.150752067565918, 3.167232036590576, 3.1784958839416504, 3.174272060394287, 3.164128065109253, 3.176448106765747, 3.1641600131988525, 3.180543899536133, 3.1733760833740234, 3.16211199760437, 3.1682240962982178, 3.180543899536133, 3.1703040599823, 3.1539199352264404, 3.1825599670410156, 3.1682560443878174, 3.1651840209960938, 3.1795198917388916, 3.169152021408081, 3.180543899536133, 3.1610560417175293, 3.1733760833740234, 3.152895927429199, 3.166208028793335, 3.171328067779541], "compile_time": 589.8605068214238, "verification_time": 0, "benchmark_time": 106.29843035712838, "f": 1080.0, "energy": 0.2342263412780191, "GFLOPS/W": 0.3415499706971155, "strategy_time": 0, "framework_time": 1.6239569522440434, "timestamp": "2026-06-04 07:40:33.533408+00:00"},
43+ "768,1080": {"block_size_x": 768, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.2348117851425529, "time": 3.169880025088787, "times": [3.2081921100616455, 3.1733760833740234, 3.175424098968506, 3.1723520755767822, 3.1784958839416504, 3.16211199760437, 3.1723520755767822, 3.1641600131988525, 3.166208028793335, 3.1764159202575684, 3.1600639820098877, 3.161087989807129, 3.1774721145629883, 3.1712958812713623, 3.1702399253845215, 3.1672000885009766, 3.166208028793335, 3.175424098968506, 3.1703040599823, 3.161087989807129, 3.165087938308716, 3.1631360054016113, 3.1774721145629883, 3.1651840209960938, 3.1692800521850586, 3.1651840209960938, 3.161087989807129, 3.1692800521850586, 3.1682560443878174, 3.1631360054016113, 3.1733760833740234, 3.166208028793335], "compile_time": 587.5585596077144, "verification_time": 0, "benchmark_time": 106.18988610804081, "f": 1080.0, "energy": 0.2348117851425529, "GFLOPS/W": 0.34069840213272284, "strategy_time": 0, "framework_time": 1.4678766019642353, "timestamp": "2026-06-04 07:40:34.228656+00:00"},
44+ "832,1080": {"block_size_x": 832, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.22670568248873874, "time": 2.9767550006508827, "times": [3.047424077987671, 2.978816032409668, 2.9767680168151855, 2.974720001220703, 2.9757440090179443, 2.973695993423462, 2.9757440090179443, 2.973695993423462, 2.974720001220703, 2.973695993423462, 2.973695993423462, 2.974720001220703, 2.973695993423462, 2.974623918533325, 2.974720001220703, 2.9726719856262207, 2.974720001220703, 2.974720001220703, 2.973695993423462, 2.974720001220703, 2.9745919704437256, 2.973695993423462, 2.974720001220703, 2.9726719856262207, 2.973695993423462, 2.974720001220703, 2.9745919704437256, 2.974720001220703, 2.973695993423462, 2.974720001220703, 2.9736320972442627, 2.973695993423462], "compile_time": 591.2145189940929, "verification_time": 0, "benchmark_time": 100.32539488747716, "f": 1080.0, "energy": 0.22670568248873874, "GFLOPS/W": 0.35288043564578, "strategy_time": 0, "framework_time": 1.6864659264683723, "timestamp": "2026-06-04 07:40:34.921915+00:00"},
45+ "896,1080": {"block_size_x": 896, "nvml_gr_clock": 1080, "temperature": 45.189338235294116, "core_freq": 1080.0, "ps_energy": 0.22485707456937543, "time": 3.0062319934368134, "times": [3.012608051300049, 3.0208001136779785, 3.005311965942383, 3.0269439220428467, 3.0074880123138428, 3.0105600357055664, 3.0064640045166016, 3.004415988922119, 3.009536027908325, 2.9951999187469482, 3.0003199577331543, 3.0023679733276367, 3.0054399967193604, 3.0054399967193604, 3.0115840435028076, 3.0146560668945312, 2.9972479343414307, 3.0023679733276367, 3.004415988922119, 2.9962239265441895, 3.0003199577331543, 3.004415988922119, 2.998271942138672, 2.999295949935913, 3.009536027908325, 3.0197761058807373, 3.006335973739624, 2.9962239265441895, 3.004415988922119, 3.0064640045166016, 3.0146560668945312, 3.0003199577331543], "compile_time": 586.9424073025584, "verification_time": 0, "benchmark_time": 100.3955970518291, "f": 1080.0, "energy": 0.22485707456937543, "GFLOPS/W": 0.35578155658748245, "strategy_time": 0, "framework_time": 1.7974716611206532, "timestamp": "2026-06-04 07:40:35.611083+00:00"},
46+ "960,1080": {"block_size_x": 960, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.22746058879003217, "time": 3.021076038479805, "times": [3.066879987716675, 3.0146560668945312, 3.0156800746917725, 3.0208001136779785, 3.0167040824890137, 3.018752098083496, 3.0299839973449707, 3.033087968826294, 3.0003199577331543, 3.0167040824890137, 3.026911973953247, 3.0115840435028076, 3.01145601272583, 3.0115840435028076, 3.012608051300049, 3.008512020111084, 3.012608051300049, 3.0443520545959473, 3.0146560668945312, 3.0218238830566406, 3.0167040824890137, 3.016576051712036, 3.008512020111084, 3.0197761058807373, 3.0146560668945312, 3.0443520545959473, 3.0310399532318115, 3.0208001136779785, 3.018752098083496, 3.015615940093994, 3.012608051300049, 3.0453760623931885], "compile_time": 589.8595550097525, "verification_time": 0, "benchmark_time": 101.14786308258772, "f": 1080.0, "energy": 0.22746058879003217, "GFLOPS/W": 0.3517092803881187, "strategy_time": 0, "framework_time": 1.6564596444368362, "timestamp": "2026-06-04 07:40:36.303781+00:00"},
47+ "1024,1080": {"block_size_x": 1024, "nvml_gr_clock": 1080, "temperature": 45.0, "core_freq": 1080.0, "ps_energy": 0.22918135136547868, "time": 3.0677370131015778, "times": [3.1139841079711914, 3.092479944229126, 3.078144073486328, 3.0791358947753906, 3.0740480422973633, 3.066879987716675, 3.039232015609741, 3.054527997970581, 3.073024034500122, 3.0515201091766357, 3.047424077987671, 3.072000026702881, 3.063807964324951, 3.078144073486328, 3.0801920890808105, 3.073024034500122, 3.0842878818511963, 3.0515201091766357, 3.047424077987671, 3.0453760623931885, 3.049407958984375, 3.0351359844207764, 3.047424077987671, 3.052544116973877, 3.0566399097442627, 3.0617599487304688, 3.0904319286346436, 3.088383913040161, 3.0760319232940674, 3.078144073486328, 3.1037440299987793, 3.0617599487304688], "compile_time": 587.9706707783043, "verification_time": 0, "benchmark_time": 102.63033676892519, "f": 1080.0, "energy": 0.22918135136547868, "GFLOPS/W": 0.3490685412375586, "strategy_time": 0, "framework_time": 1.669283490628004, "timestamp": "2026-06-04 07:40:36.996083+00:00"}}
48+ }
0 commit comments