How is it that this 390x card is beating the 980ti and TitanX which are theoretically MUCH more powerful???
I smell terrible Maxwell optimization.
>.>
Great job with OpenCL so far Blender devs, this is very impressive especially considering that OpenCL rendering was just implemented! Even before the new drivers from AMD, the 390x was pulling around 52 seconds which still beats the $1000 TitanX…
But seriously, optimize CUDA rendering, obviously their full potential is not being used, at least for the Maxwell cards.
Cool, thanks. So, can I combine Nvidia and AMD to work together? I have Nvidia card and was considering buying TitanX, but with your test I might go with AMD. Are there any limitations of OPenCL compared to CUDA with Cycles?
Let’s assume for the moment that the efficiency and power of the Hawaii chip and the Maxwell chip are about the same (they’re most certainly not) Should not the Maxwell chip with much higher clocks on the core and memory be able to beat the 390x?
Now in the real world the Maxwell chip is quite a bit more efficient and powerful in the 980ti and TitanX than the Hawaii chip, does that not raise eyebrows when the 390x pretty clearly is beating these cards?
Yes you can combine them (I tested this), no it’s not ideal and would seem to be rather buggy at least for the moment.
OpenCL runs slower on Nvidia cards than CUDA. (I tested this)
For the moment OpenCL can’t do Subsurface scattering (you can enable it but it could crash), transparent shadows, and a couple other things, consult google. There is a very good chance these will be implemented over the next couple versions of Blender.
For the best render-speed for the best price, right now two R9-390x cards will preform a good deal better than a TitanX and cost less.
You will want a 1000 watt PSU ideally for that, gold at least. (I’d personally go with a 1200 watt PSU so it is always silent). You don’t need quite that much power, but that’s playing rather safe which is where I personally like to be.
P.S. They make a lot of heat. also, you’d render everything and everything in 2 tiles, 1 tile per card.
EDIT*
My advise, if you get two 390x cards, upgrade to a water cooling loop and get blocks for those. That’s extreme, but it’s also extremely awesome!
2 FuryX cars would also be stellar, but I’d wait for the dual GPU version of that.
2 Fury Nanos would also be stellar. The nano is released this week.
Aside that I’m not so sure about the much higher clocks on the 980Ti and Titan X since they don’t boost that much on compute heavy loads. When they hit the temperature threshold boost is gone for good.
You just did that test in a 1 min benchmark. If you did a 10 min one the beating would have been legendary.
Both the 980ti and TitanX used exceed the speed of the 390x in terms of GFLPs due to the much much higher clock speeds they are running stably at and the Titan X already does. Not to mention and aside from all that the memory for the 980ti and TitanX is faster.
Yet the 390x is clearly beating both cards.
Clearly there is an optimization problem for Maxwell here.
82 Cº is above the 2nd threshold of GPU Boost, hence your clocks are limited by this threshold.
Si tGPU < 63°C : 1190 MHz @ 1.174V
Si tGPU > 63°C : 1177 MHz @ 1.162V
Si tGPU > 73°C : 1164 MHz @ 1.149V
Si tGPU > 80°C : 1152 MHz @ 1.137V
GPU Boost is also limited by TDP and a compute intensive task like Cycles is more than likely to be clocking your card down.
If you’re using the MSI Afterburner software it is as easy as making an screenshot showing the GPU frequency graph right after finishing the benchmark.
My point still stands, I was just trying to figure a scenario. Maxwell isn’t better or more powerful, that’s just Nv propaganda. My GTX 580 is supposed to be several times slower than Kepler or Maxwell cards and it’s still hanging there.
Also, the memory on both the 980Ti and Titan X is slower. I’d say that at best you can match the 390X bandwidth.
@Esparadrapo the 5xx series are beasts… there was actually a drop in GPU compute performance going from the 5xx series to the 6xx series… the 9xx series / titans now are slightly faster but not by much… it is slightly disappointing sticking with the same card for years.
@John… these cards are more ‘gaming’ cards… for ‘gaming’ purposes, the TitanX probably will win out… In terms of Compute though… AMD seem faster… However, the AMD kernel does not have all the features that the CUDA kernel has… So it is not a fair comparison.
