gg% SSKrSSKrSSKJrJrJr \"SS5\"SS\R SS5\"SS \R SS5\"S S \R SS5/rS rS r Sr Sr Sr Sr Srg)N) Importance MetricRequestRequestedMetricsParserz0sm__throughput.avg.pct_of_peak_sustained_elapsed sm_sol_pctz@gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsed mem_sol_pctFzJbreakdown:gpu__compute_memory_throughput.avg.pct_of_peak_sustained_elapsedbreakdown_memory launch__waves_per_multiprocessor num_wavescg)N SOLBottleneckr 0nsight-compute-2025.1.1/sections/SpeedOfLight.pyget_identifierr&s rcg)N Bottleneckr r rrget_namer)s rcg)NzHigh-level bottleneck detectionr r rrget_descriptionr,s ,rcg)N SpeedOfLightr r rrget_section_identifierr/s rcSnSnUR5H up4UR5nXR:dMUnUnM" U(a.SSSSSSS.nUHnURU5(dMXgs $ g)NrDRAML1L2)draml1texltsltcfbpfbpa)itemsvalue startswith)breakdown_metricsmax_pipemax_pipe_valuenamemetric pipe_valuetokenstokens r get_max_piper.2sHN)//1 \\^  &'NH 2   E""5))}$ rc/nUR5RS5Hn[U5nURU5 M! UR U5$)N,)r$splitrappendparse)parserbreakdown_metricrequests metric_namerequests rget_breakdown_metricsr9MsLH'--/55c:  , ; << !!rc L[R"U5nURS5RS5nUR 5n[ X5nUR [5nUSnUSnUSnUbUcg[XH5n UR5n UR5n UR5n UR5n SnSnSnSn[RRn/nX:aS nOS nU U:GaU U:GaX:GaX:Gay[RRnUS nU(aUR5U:aUR5nURUR5U[RRS R!U545 UR#5[R$R&:XaS R!U5nOSR!U5nUS- nSnGOURX[RRSR!U U545 URX[RRSR!U U545 SR!U5nSnGO8[(R*"X- 5U:a[RRnX:aPURX[RR,SR!XU545 SR!U5nSnOURX[RR,SR!XU545 [/U 5nU(aSR!U5OSnSR!UU5nSnO7SnSnO2SnX:a [/U 5nU(aUOSnSR!UUU5nSnUR1UUU5nUH$nUR3UUSUSUS US!5 M& UR5U S"- U S"- [7X5S"- S#.5 g)$Nrrrr <Pz;@section:ComputeWorkloadAnalysis:Compute Workload Analysis@z9@section:MemoryWorkloadAnalysis:Memory Workload Analysis@r z.Increase the number of waves per SM towards {}zThis kernel grid is too small to fill the available resources on this device, resulting in only {:.1f} full waves across all SMs.zAll launches of this workload use grids that are too small to fill the available resources on this device, resulting in at most {:.1f} full waves across all SMs.zB Look at @section:LaunchStats:Launch Statistics@ for more details.z Small Gridz{:.3f} < {:.3f}alThis workload exhibits low compute throughput and memory bandwidth utilization relative to the peak performance of this device. Achieved compute throughput and/or memory bandwidth below {:.1f}% of peak typically indicate latency issues. Look at @section:SchedulerStats:Scheduler Statistics@ and @section:WarpStateStats:Warp State Statistics@ for potential reasons.z Latency Issuez{:.3f} - {:.3f} >= {:.3f}zCompute is more heavily utilized than Memory: Look at the {} section to see what the compute pipelines are spending their time doing. Also, consider whether any computation is redundant and could be reduced or moved to look-up tables.zHigh Compute Throughputzto identify the {} bottleneckz,to see where the memory system bottleneck isa:Memory is more heavily utilized than Compute: Look at the {} section {}. Check memory replay (coalescing) metrics to make sure you're efficiently utilizing the bytes transferred. Also consider whether it is possible to do more work per memory access (kernel fusion) or whether there are values you can (re)compute.zHigh Memory ThroughputaCompute and Memory are well-balanced: To reduce runtime, both computation and memory traffic must be reduced. Check both the @section:ComputeWorkloadAnalysis:Compute Workload Analysis@ and @section:MemoryWorkloadAnalysis:Memory Workload Analysis@ sections.zBalanced Throughput workloadsaThis workload is utilizing greater than {:.1f}% of the available compute or memory performance of the device. To further improve performance, work will likely need to be shifted from the most utilized to another unit. Start by analyzing {} in the {} section.zHigh Throughputd)compute_throughput_normalizedmemory_throughput_normalizedmax_throughput_normalized)NvRules get_context range_by_idx action_by_idxfrontendrr3requested_metricsr9r)r$ IFrontendMsgType_MSG_OKMsgType_MSG_OPTIMIZATIONr2Severity_SEVERITY_HIGHformat workload_typeIActionWorkloadType_KERNELmathfabsSeverity_SEVERITY_LOWr.message focus_metricsend_dict_to_childrenmax)handlectxactionfer4metricssm_sol_pct_metricmem_sol_pct_metricbreakdown_memory_metricbreakdown_metrics_memorysm_sol_pct_namemem_sol_pct_namerrbalanced_thresholdlatency_bound_thresholdno_bound_thresholdwaves_thresholdmsg_type focus_metricsbottleneck_sectionnum_waves_metricr rWr) pipe_namepipe_msgmsg_idrXs rapplyrqUs   f %C   a . .q 1F B #F 3Fll,-G - /%&89!%<%D4VU',,.O)..0"((*J$**,K O  //HM ZX&&;9K+K  /K4Y((AAH&{3 $4$:$:$<$N,224 $$&6&;&;&=y'J[J[JrJrueululm|u}&~'')W__-P-PPbiijstGBIIJSTG__#$$o7CTCTCkCkm~nFnFGQSenf&gh$$&6WEVEVEmEmpApHpHITVhpi&jkIPPQhi& YYz/ 04F F((AAH'$$oGDUDUDkDknInPnPQ[j|n}&~GNNOab0$$&6WEVEVElEloJoQoQR]k}o~&@()AB PY:AA)L`NW^^_qs{|/YG(D  #$%=>I )9{W^^_qs{}OP  ZZ'4 0F%   Qa,q/S_`aSbc&-7#-=,7#,=),Z)E)K r)rTrFRequestedMetricsrrrOPTIONALrKrrrrr.r9rqr rrrts2 NNDlSTVceoexexz~AFG^`rt~uHuHJNPUV4k:CVCVX\^cd -6"Zr