smilefaker
Banned
- Registriert
- Feb. 2014
- Beiträge
- 1.161
AMD hat die Treiber auch getrimmt, siehe Assassins Creed CPU Limit.
Folge dem Video um zu sehen, wie unsere Website als Web-App auf dem Startbildschirm installiert werden kann.
Anmerkung: Diese Funktion ist in einigen Browsern möglicherweise nicht verfügbar.
kisser schrieb:Zwei Dinge kann man aus dem Test herauslesen:
1. Nvidia hat die DX11 Treiber auf hohe Draw-Call Last getrimmt, AMD nicht (das ist beides ja schon länger bekannt)
2. DX12 und Mantle performen ungefähr gleich
Somit sind die Benchmarks wertlos da falsch eingestellt. Die vorhandenen 2 Core Optimierung ist hier wohl auch ein Vorteil für Mantle. Vor allem da als Option vom User auswählbar.Update: Oxide Games has emailed us this evening with a bit more detail about what's going on under the hood, and why Mantle batch submission times are higher. When working with large numbers of very small batches, Star Swarm is capable of throwing enough work at the GPU such that the GPU's command processor becomes the bottleneck. For this reason the Mantle path includes an optimization routine for small batches (OptimizeSmallBatch=1), which trades GPU power for CPU power, doing a second pass on the batches in the CPU to combine some of them before submitting them to the GPU. This bypasses the command processor bottleneck, but it increases the amount of work the CPU needs to do (though note that in AMD's case, it's still several times faster than DX11).
This feature is enabled by default in our build, and by combining those small batches this is the likely reason that the Mantle path holds a slight performance edge over the DX12 path on our AMD cards. The tradeoff is that in a 2 core configuration, the extra CPU workload from the optimization pass is just enough to cause Star Swarm to start bottlenecking at the CPU again. For the time being this is a user-adjustable feature in Star Swarm, and Oxide notes that in any shipping game the small batch feature would likely be turned off by default on slower CPUs.
"Wir sind gegen Vendor-Lockin... Außer wir machen das. "
Sontin schrieb:Geil ist, wie AMD weiter Craptle als "offene" API verkaufen will. Das ist schon an Lächerlichkeit nicht mehr zu überbieten:
"Wir sind gegen Vendor-Lockin... Außer wir machen das. "
But if you want more technical detail, than here is it:
NV Fermi: Max UAV is limited to 8 -> TIER1
NV Kepler: Max UAV is limited to 8 -> TIER1
NV Maxwellv1: Max UAV is limited to 8 -> TIER1
NV Maxwellv2: SRVs/stage is limited to 2^20 -> TIER2
Intel Gen7dot5/Gen8: the universal hardware binding table is limited to 255 slot -> TIER1
AMD GCN v1/v2/v3...: GCN is designed to a simplified resource model, so this architecture works more like a CPU than a GPU. This will allow unlimited resource binding -> TIER3
The GCN revs are not so different. The base 1.0 version is also good for the most D3D12 features, and for the best binding tier.
The 1.1 version is mostly brings efficiency updates, like multi queue compute with upgraded ACEs. Also device unified addressing is a huge deal, because it will allow the kernel to view the LDS and the device memory as a single addressable space.
The 1.2 version ISA manual is under NDA, so I can't talk about what possible with Tonga. This GPU has some very interesting unannounced capabilities focused to reduce latency (good for VR)
Could you please elaborate?? I didn't think GCN v1.0 or even v.1.1 could support ALL features in DX12 with hardware-only (no emulation!!!). Are you saying even GCN v1.0 supports ALL features in DX12 with hardware-only, no need for emulation?