Gaussian16 C.01 AVX | Benzene CPU / oneGPU teaching example

bash run.sh : print included realreference and redraw2figures; does not run Gaussian or submitjobs.
yskit gaussian-init : initialize your own authorized Gaussian environment ifneeded.
bash submit.sh p100 : submit6nativeconditions tooneP10016G with2physicalCPUcores.
bash submit.sh v100 : optionallyrunthesame6conditions ononeV10016G; notV10032G.
bash results.sh : after your job(s)finish, independentlyextract ownnumericoutputsandplotownfigures. Oneplatformissufficient;thereferencecontainstwo.

Use a fresh extracted directory for a new calculation. Existing work/platform is protected. Eachjob runsHF/B3LYP Force atfixedbenzenegeometry/6-31G(d),102basisfunctions,42electrons/singlet; eachmethod hasCPU-onlybaseline,oneGPUunprofiled,oneGPUwithruntimelaunchnotifications. All3modesuse2totalphysicalCPUcores;GPUmodesreserveoneascontroller,not2extraCPUcores. %CPU/%GPUCPUcoreidsarerenderedfromactualSlurmaffinity. %Mem=2GBwithSlurmhostmemory8G. CPUbaselineisinthesameGPUnodewithCUDAhidden;itdoesnotrepresentdedicated8173MCPUhardware.

Publicjobtemplatesloadgpumdkit1.5.8Python3.11 and youralreadyinitializedGaussian environment. Existing g16root is respected. GaussianC.01isvendorAVXbinary,notlocallycompiledwiththenotificationhelpercompiler. No program,fullwavefunction,FCHK,RWF,licensedmanual,rawexecutionreceipt,jobIDorprivateaccountpathisinthepublicarchive.

Notifications: thelegacyGaussianOpenACCruntimerequestslibaccnotifymp.so; theclusterhasNVIDIA25.3hostreporterlibaccnotify.so. Thegpu-notifyconditionmakesabyteidenticalprojectlocalcopywiththelegacyreporterfilenameandloadsonlythatdirectoryforthatprocess. No OpenACCmathcore/libcuda/driver/sharedfileisreplaced. ExistingNVIDIAreporterSHA256:38bda4fa9126027d2c091e7dd1d732c551400043ea1dd6f7718f0cc9ac7def59. ActualsuccessfulruntimekernelnotificationsandsameCPU/GPUenergy/forcegatesproveuseofGPUkernels;memoryallocationorloadmessagesalonearenotproof. Eventsreportkernelqueues,notGPUkerneltimes;donottimetheenqueuecallbackasdeviceexecutiontime.

FourCSVtablescontainactualenergies,allforces,CPU/GPUdifferencesandaggregatedkernellaunchnotifications. selected-physics.jsoncontainsownselectednumericfieldsandgeometryonly. Runtime/countervaluesvarywithscheduling,batching,hardwareandcache;theyarenotbitwisephysicalinvariants. CPU/GPUenergy1e-7Ha/force1e-6HaBohr/geometry3e-8Agatesarecheckedunchanged. NativeoutputlogenergyandforcesplusFCHKshell-count/electron-countareindependentlychecked. Ownrecomputedplotsneednotmatchreferencetiming/countplotsbyteforbyte;bashrun.shredrawsthefrozenreferenceexactly.

Thetimingfigurecontainsunprofiledwallobservationsincludingstartupandmonitoring,whilenotificationcountscomefromseparatereporterconditions. This12atomsmallcaseisnotalargesystemspeedupbenchmark. GPUcanbesloweronasmallcase. ItisnotanoptimizationresultorproofthatallGaussianmethods/keywordsareGPUaccelerated.
