Huge disappointment. MacBook Pro used to have NVIDIA GPUs and was an excellent CUDA development environment. They switched to AMD around 2012. I suspect the change was more business related than technical.
Huge disappointment. MacBook Pro used to have NVIDIA GPUs and was an excellent CUDA development environment. They switched to AMD around 2012. I suspect the change was more business related than technical.
I recognize that a lot of people want CUDA instead, but OpenCL support is superior on AMD cards.
AFAIK their OpenCL debugging/profiling tools are still terrible compared to CUDA or AMDs OpenCL though.
NVidia supports 1.2. AMD supports 2.0. Same concept, but I got the numbers all jumbled up. I'll edit the above post for the correct numbers.
http://support.amd.com/en-us/kb-articles/Pages/OpenCL2-Drive...
http://stackoverflow.com/questions/29219307/opencl-2-0-on-nv...
(Hard to "prove a not", so I'll just leave the Stackoverflow page here)
But this time around Apple is writing their own software (Metal). So this time around it's not surprising that Apple doesn't want to pay more for superior software if it's not going to use the software. It sucks for CUDA users, but Apple doesn't care about you.
The RX480 (Polaris) trades blows with the GTX1060 (Pascal) performance-wise, but the 1060 uses about 30% less power at load:
https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1060/...
According to Guru3D the two cards perform almost identically in Ashes of the Singularity and Total War: Warhammer, both AMD-sponsored DX12 games.
And whether Doom 4 is an outlier or just an example of the difference between Vulkan and DX12 is yet to be determined.
My conjecture is that the big difference between Vulkan and DX12 is that nVidia has put more of it's driver engineering magic into DX12 than it has into Vulkan.
Metal won't have that nVidia driver engineering magic, so I believe that DOOM4 is the best benchmark we have of performance and power consumption for Metal. But one benchmark can be misleading for a wide variety of reasons.
http://www.eurogamer.net/articles/digitalfoundry-2016-nvidia...
And conversely, the developers of Doom said their async path is currently only tuned for GCN and will be enabled on Pascal in a future patch.
We've seen from 3DMark Time Spy that Pascal does benefit from async compute, so there's performance being left on the table there.