Kupa GTX 970

[Zoltan Balaton]

Ako pre sedlaka mozes vdaka asynchronnym shaderom dostat up to 30-40% vykonu navyse oproti stavu aky je teraz.

kde tam vidis realny narast niecoho ? ved v dx11 ma furyx skoro polku fps ako 980ti..a to nie je realna hodnota myslim..skorje tam fury brzdena v dx11 ako upovana v dx12

[Leon]

Ako pre sedlaka mozes vdaka asynchronnym shaderom dostat up to 30-40% vykonu navyse oproti stavu aky je teraz.

šak ale 9xx series asynchronne shadery maju, aspon podla toho co tu dal per http://diit.cz/clanek/asynchronni-shade ... icke-karty

co vlastne teda hejtujes akože?

[Zoltan Balaton]

uz aj na ddworld pisu ze 9xx ich maju..tak ja uz neviem..ked to tam tvrdia,asi to bude naozaj tak
http://www.ddworld.cz/aktuality/grafick ... ykonu.html

[Asnod]

That is if Maxwell 2 has only a single Asynchronous Compute Engine tied to 32 Queues

Ano, ma jednu, amd karty maju up to 64 tychto jednotiek. Ak je tam jedna ako moze pracovat asynchronne? Mozem tu postnut cele vysvetlenie v anglictine 10 krat aj tak tu budete tahat zumpa weby kde nieco prelozia koli navstevnosti.

[Zoltan Balaton]

daj sem test,kde uvidim o kolko je rychlejsia 390 od 970..lebo sa tu len keca a dokopy ziadny dokaz

[Leon]

That is if Maxwell 2 has only a single Asynchronous Compute Engine tied to 32 Queues

Ano, ma jednu, amd karty maju up to 64 tychto jednotiek. Ak je tam jedna ako moze pracovat asynchronne? Mozem tu postnut cele vysvetlenie v anglictine 10 krat aj tak tu budete tahat zumpa weby kde nieco prelozia koli navstevnosti.

asi som priyebaný ja alebo neviem?,... ale keď má nvidia 1+32 tak asynchronne shadery nemá ale keď má amd 1+64 tak už asynchronne shadery má? vôbec to neznamená, len, že amd vie pracovať s viac frontami nie? otvoril si tú tabuľku ty vôbec?

[scream]

https://opus4.kobv.de/opus4-zib/files/5036/ZR-14-19.pdf

6 Conclusion
Executing single small-scale workloads on GPU is insufficient to fully utilize the
GPU’s compute performance. A viable approach to take advantage of accelerators,
however, is to execute multiple kernels in a concurrent way. For a synthetic
benchmarking kernel, we found using Nvidia’s Hyper-Q feature – a mechanism
for concurrent offload computations on a shared GPU within multi-threaded/-
process applications – give significant performance gains over a serial execution
of the same amount of work on GPU. However, we encountered some deficiencies
when using Hyper-Q from within multi-threaded programs causing the overall
program performance be up to a factor 3 behind the expectations. We were able
to recover the performance by using a kernel reordering scheme we introduced
earlier for Nvidia Fermi GPUs. Using the Nvidia Visual Profiler, we found that
seemingly not all hardware queues offered by Kepler GPUs are equally favored,
but some of them are processed to completion before the others. Executions using
Hyper-Q within multiple processes behave close to optimal with respect to
concurrency on GPU. For a real-world application implementing a simulation
of a small molecule solvated within a nano-droplet, we achieve about a factor 2
– 3 performance gain over an optimized multi-threaded CPU execution due to
Hyper-Q.

Nvidia nema, nevie, klame, zavadza... sa pytam potom, kde sa vzal vykonovy rozdiel pri paralelnych vypoctoch.

Samozrejme, AMD to ma implementovane lepsie. Otazne ale je ci a do akej miery sa to prejavi v hrach a to nikto nemoze vediet.

Apropo konzolovka project cars s logami nvidie - amd vykonovo 30% nizsie... takze na konzoly by som sa nespsoliehal.

[loleqboleq]

dosť sa hrotí táto debata, nemyslím len tu ale celé toto koumání na internetoch a ja osobne nemám nič proti tomu, ak by sa to stalo, tak by AMD dohnalo stratu v high-ende a možno by to zase odštartovalo konkurenčný boj na úrovni

aj tak sa idem "zasamovraždiť" s 980Ti, lebo pokiaľ by aj ten benefit s AMD prišiel naozaj, tak to príde o koľko..o dva roky ? dovtedy už budeme mať aj tak všetci nové železo pod kapotou

[scream]

ved to som pisal ...ale na to je zas argument, ze nie kazdy si kupuje kartu na rok... 970tku som za tie prachy kupoval s tazkym srdcom lebo stara AMD 7970 sla prec za 140E a pomohol som si len 20-30% v globale...

V project cars je to ale aj cez 50% + a to bol jediny dovod preco som vobec robil upgrade. Inak by som si pockal na pascala so 7970 ktora dostacovala na medium-high pri 1080p vo vacsine hier.

[Asnod]

That is if Maxwell 2 has only a single Asynchronous Compute Engine tied to 32 Queues

Ano, ma jednu, amd karty maju up to 64 tychto jednotiek. Ak je tam jedna ako moze pracovat asynchronne? Mozem tu postnut cele vysvetlenie v anglictine 10 krat aj tak tu budete tahat zumpa weby kde nieco prelozia koli navstevnosti.

asi som priyebaný ja alebo neviem?,... ale keď má nvidia 1+32 tak asynchronne shadery nemá ale keď má amd 1+64 tak už asynchronne shadery má? vôbec to neznamená, len, že amd vie pracovať s viac frontami nie? otvoril si tú tabuľku ty vôbec?

Asi si pri*ebany

That is if Maxwell 2 has only a single = 1 Asynchronous Compute Engine tied to 32 Queues.

Each ACE is capable of handling eight queues and every GCN based GPU has a minimum of two ACEs. More modern chips such as the R9 285 and R9 290/290X have eight ACEs.

http://www.techpowerup.com/215663/lack- ... tx-12.html

Async Compute is a standardized API-level feature added to Direct3D by Microsoft, which allows an app to better exploit the number-crunching resources of a GPU, by breaking down its graphics rendering tasks. Since NVIDIA driver tells apps that "Maxwell" GPUs supports it, Oxide Games simply created its benchmark with async compute support, but when it attempted to use it on Maxwell, it was an "unmitigated disaster." During to course of its developer correspondence with NVIDIA to try and fix this issue, it learned that "Maxwell" doesn't really support async compute at the bare-metal level, and that NVIDIA driver bluffs its support to apps. NVIDIA instead started pressuring Oxide to remove parts of its code that use async compute altogether, it alleges.

We wonder how much of that support is faked at the driver-level, like async compute. The company is already drawing flack for using borderline anti-competitive practices with GameWorks, which effectively creates a walled garden of visual effects that only users of NVIDIA hardware can experience for the same $59 everyone spends on a particular game.

Viac sa na tuto temu bavit nebudem, BFU ostane BFU. Tiez mam ine veci na robote nez sa tu hadat o tom.

Len tak btw Scream v globale je 390 vykonnejsia vsade od hocakej nereferencnej 7970 o rovnych 50% pokial nieje bottlenecked, coz u teba v 80% situacii neplati. To iste plati aj cca 970 takze este asi tolko k tomuto.

[Leon]

Mňa ale úplne vôbec nezaujíma, či dokáže nejaké amd pracovať s 200 alebo 50 tisíc frontami,... toto tu všetci chápeme. A ty to tu v kuse opakuješ jak somarík a cituješ stále tu istú a jedinú anglickú vetu. Mne úplne stačí ku kľudnému spánku, že GTX 9xx rady dokážu pracovať s grafickou a vypočetnou frontou zároveň, čiže asynchronné shadery majú,... a to je všetko.

[Asnod]

Nie nemaju. Ty to tu ako somarik nedokazes pochopit ze maju Nv karty " 1ks ace" a kazda GCN grafika "2-8" 1ks ace sama o sebe nikdy asynchronne pracovat nedokaze. To ze tu technologiu ta karta "podporuje" neznamena ze z nej dokaze aj benefitovat.

http://images.anandtech.com/doci/9124/Async_DX11.png
http://images.anandtech.com/doci/9124/Async_DX12.png

Este pre noobov logicka otazka. Ako chces benefitovat z hyperthreadingu ak mas na cpu pre kazde core len 1 thread? Uz ked ti to po tomto nedocvakne v tej pochcanej tabulke mas pocet vypocetnich font =32 na 1 compute engine. amd ma 8 vypocetnich font na kazdy compute engine, ACE, pri r9 390 ich mas teda tych vypoctovych font dokopy 64.

[Zoltan Balaton]

nj,tak sa asi zhodneme,ze nv AS ma,ale v podstate ich nevie pouzivat
http://www.cnews.cz/directx-12-async-co ... ell-kometa

[scream]

ja neviem, cital to niekto? http://docs.nvidia.com/cuda/samples/6_A ... HyperQ.pdf

GMU presklada serializovany vstup na co najoptimalnejsi serializovany vystup takym sposobom aby boli dodrazne dependencie. Vdaka tomu je mozne tento prud optimalne prerozdelit pre urcity pocet SMX, tak aby mohli byt tasky spracovane asynchronne. Cele je to o tom aby jeden task necakal na skoncenie druheho, ak su volne vypocetne zdroje a pri idealnom scenari to ozaj funguje.

Tento cirkus si ale vyzaduje tiez nejake latencie, ktore su sice mensie ako pri false dependencies ale su + par dalsich obmedzeni.

AMD to ma riesene velmi priamociaro, jednoducho a efektivne ale sudny clovek nemoze povedat, ze NVIDA nevie spracovat streamy asynchronne, lebo vie, len sposobom ''dookola blizsie''.

Trollovat tu ale celu produkciu kariet NVIDIA na zaklade jedneho bencharku z hry v alfa verzii, ktora graficky vyzera ako rit je prehnane.

Len tak btw Scream v globale je 390 vykonnejsia vsade od hocakej nereferencnej 7970 o rovnych 50% pokial nieje bottlenecked, coz u teba v 80% situacii neplati. To iste plati aj cca 970 takze este asi tolko k tomuto.

7970Ghz vs. GTX970 gaming 28% https://www.techpowerup.com/reviews/MSI ... ng/27.html

[loleqboleq]

z tohoto hľadiska je na tom kepler a maxwell rovnako, či sú tam rozdiely ?