Question about Ms/sec scaling

[Vue2Octane]

I was comparing the other two systems with the Octane Benchmark Scene and a few other scenes.

My 'Toy' at home:
http://www.amazon.de/dp/B00F9PSB8E/ref= ... _M3T1_dp_1
(GTX780M)

(sorry for german, you get the specs)

and a workstation:

HP Z420, 8GB Ram, Xeon E-5-1620 @3.7GHz, 8 Cores
equipped with GTX 780M, 3GB

Both 64 Bit.

I expected a speed icnrease for Octane by factor 3-ish, because GTX760M = 768 cuda cores and GTX 780M ~ 2200 Cuda Cores. But the Workstation gives a speed increase roughly factor 4.5-6, for some scenes bit more, for some bit less. Does memory bandwidth etc. also contribute here.

The Benchmark scene gives ~1Ms/Sec as loaded for GTX760M, and ~5Ms/sec for GTX780. Pathtracing, alpha shadows on. Just opened the benchmark scene, switched to PT and press 'play button'

I am not annoyed that my expectations are surpassed, just confused.

[Seekerfinder]

Hi Vue2Octane,

and a workstation:

HP Z420, 8GB Ram, Xeon E-5-1620 @3.7GHz, 8 Cores
equipped with GTX 780M, 3GB

Assuming you mean GTX780 (not M) for your workstation, that card has 2304 cores at a base clock speed of 863MHz. This is not the sum-total of a card's processing power but for the sake of argument, let's compare the 2 cards:

GTX760M: 768 cores x 657MHz clock speed = 504576 (it's just a factor for this purpose)
GTX780: 2304 cores x 863MHz (base) clock speed = 1988352

Theoretical increase: 3.94

Like I said, there are other factors and while I don't know them all, temperature control (including throttling) and drivers are two typical ones.

Best,
Seeker

[Vue2Octane]

Thank you for comparing the two cards.

On the Octane Website it says the render speed of Octane scales with Cuda Cores. I assume they mean two identical cards in the same rig.

Good to know that clock speed x cuda cores is overall unit.

Bandwidth, Temperature etc. might then add to this, you are right.

[Seekerfinder]

On the Octane Website it says the render speed of Octane scales with Cuda Cores.

It's true of course but it's not a linear scaling. And, unfortunately, not everything is in the software developer's control. For instance, Octane (and some other apps) were way more efficient - cuda for cuda - in the Fermi architecture than the subsequent Kepler by Nvidia. So there was a 'negative scaling' between those two processor architectures. Thankfully, GPU computing has a very bright future and new, lower energy solutions are being developed all the time. I guess the challenge is for companies like Otoy to stay ahead of the curve.

Best,
Seeker