I’m aware that AMD/OpenCL support is still in relatively early development, but I was still expecting (perhaps naively) much better performance out of the AMD M395X that you can configure with the 2015 5k iMac.
Running the new BMW benchmark (2 cars), here are the results (using Blender 2.76).
GPU 240x136 (default tile) - 5 min 32 sec
CPU 16x16 - 2 min 53 sec
To give you my own reference point … I was upgrading from the original 2006 Mac Pro (yes, 9 years old) that I had retrofitted with an Nvidia GeForce GTX 570 1.25 GB card. On the same test as above, I get the following:
GPU 240x136 (default tile) - 2 min 37 sec
So not only is the CUDA accelerated Nvidia card over twice as fast the as the AMD M395X, it even beats the i7 CPU time. Which puts me in this awkward position where my aging Mac Pro easily outperforms my brand new iMac.
Can anyone shed any light on this? Can I ever expect great Cycles rendering with this AMD card? How does the computational power of the GTX 570 compare to the AMD M395X? Are they just differently suited for Blender use or is the AMD implementation just so far behind (drivers/OpenCL development) ?
a)
The Nvidia GTX 570 and 580 were exceptionally powerful CUDA cards. Even later NVidia generations struggled to reach the same level of computing power.
b)
The iMacs use notebook hardware. The R9 M395x also is a mobile card (that’s what the “M” stands for). Notebook hardware is optimized for low heat and low energy consumption, but not necessarily for performance. That’s why mobile cards are nowhere near as powerful as their desktop counterparts.
Not using AMD cards myself, but I seem to recall that AMDs like even larger tile sizes. Maybe you could experiment further on that. But don’t expect wonders: I guess you will likely rather use CPU rendering on that machine.
On the bright side of things: Yes, your GTX 570 might be faster than your i7, but it only has 1.25 GB of VRAM, which is really not very much. Your new i7 on the other hand (soon) has 32 GB at its disposal and is not that far away from the GTX in terms of performance.
I was always aware of the “mobile” aspect of the iMac – that it in effect is an oversized laptop with a screen. As a video/graphics pro, this has always given me pause when considering such a machine. However, with the introduction of the extraordinary 5K screen (P3 gamut) and a quad-core Skylake i7 option, I decided to take a chance.
I did do a fair amount of benchmark research about the GPU but unfortunately never came across any Blender results. Nevertheless I still opted for the M395X 4096MB option (at quite a premium, mind you). It was articles like these at Bare Feats that gave me piece of mind. The M395X often beats and otherwise comes quite close to both of the high-end Nvidia cards, and is quite comparable (excluding tests that benefit from dual GPUs) to the much pricier Radeon Dx00 series of cards found in the newer trash-can Mac Pro.
I realize now that I hit a sweet spot by adding the GTX 570 to my aging Mac Pro. It really did surprise me how fast I was suddenly able to push out frames. But as you have also mentioned, I have hit the VRAM limit in many scenes quite often.
Overall, I get the impression that the M395X has quite a bit of power on tap, but it just doesn’t seem to translate to Blender
newbie here. My work got me the 5k iMac as well (well the 2gig M395 2048MB i5 one). It takes my machine forever to render in GPU. So i’m wondering if I missed a step or something, could you help me out?
I have a quick question: My Computer has 8GB of memory and is using an AMD graphic card with 4GB of memory. My spec can only handle 6 millions polygons in Blender 2.76b, beyond that, Sculpting is not possible because it lags very badly. I also notice that once it reaches 6m polys, my computer memory goes down to around 500MB. What I would like to know is this, will upgrading (more rams) help speed up and remove the lag? If not, what does having more RAM do in Blender?
Another thing I notice is that in Dyntopo mode, undoing a brush stroke seems to lag a bit. But in normal Sculpting mode, undoing has no lag whatsoever.
