PCTuning fórum

dexterav píše:prečo by mali zaplakať, tie fps s[ biedne tam či onam
a hry vychádzajú aj teraz

hnizdo píše: To je dobry vtip S context switchingem prislo jako prvni AMD PR.

Tvrzeni Oxide tu bylo rozebrano jako zcela neduveryhodne. Jejich kod je zjevne umyslne napsan tak, aby vyhovoval AMD a nevyhovoval nvidii, to se opravit neda, a je otazka, jestli je tu vubec vule. To jsem si myslel od zacatku, ale zda se ze zdaleka nejsem jediny, viz ten blog a vyjadreni toho intelaka.

Oculus Rift pouziva uplne jine techniky s uplne jinymi naroky, nez bezna hra. Pro ne se treba AMD architektura hodi. Vzhledem k tomu, ze nvidia ma vlastni VR, musime se pak ptat, jak to s tim pomalym contect switchingem delaji.

Proste vychazis z PR a dojmologickeho gulase promichaneho tak aby vyhovoval, peclive vyloucis protiargumenty, ale zadne dukazy nepredkladas.

Except it isn't exactly "Async". Nvidias hardware can't do that. They only chose to batch most(?) compute shaders and send them to the compute engine instead of executing them through the 3D engine. These two engines can't run in parallel on any Nvidia architecture, so - from Windows point of view - anything goes through the 3D queue and the hardware is just reconfigured, depending on whether the 3D or compute engine was requested.

The compute engine is just far more efficient at executing compute shaders than the 3D engine, that's why Nvidia is still making gains with compute engine use. The compute engine can execute multiple commands in parallel, while the 3D engine can't.

With Nvidia's hardware you have to make absolutely sure than both 3D and compute jobs are batched, cleanly separated by type and not interdependent. Otherwise it will fail horribly.

With what AMDs hardware is providing, compute
shaders are running just fine in parallel with any "classic" 3D
workload. AMD won't care much what you throw at it. Worst case: It will get even faster as the hardware has more work to choose from.

del42sa píše:Tohle IMHO změní až Volta, jakožto nová architektura, Pascal je jen takový mezičlánek mezi současným Maxwellem který neumí full DP výpočty a Voltou, která měla původně přijít hned po Maxwellovi. Je tu tedy "hluché" období cca dvou let, po které to bude Nvidia muset dočasně nějak řešit. Otázka je jak ?