Frame Rating: GTX 760M Results and NVIDIA Optimus Issues with Haswell
DiRT 3 and Closing Thoughts
We see a very similar pattern of results with DiRT 3 – the GTX 760M system exhibits frame issues that cause a difference in the FRAPS reported performance and the actually, observed performance. Please keep in mind that this is exactly why we developed and have promoted the Frame Rating performance testing methodology, in order to find issues like this that might not otherwise have been discovered.
This phenomenon occurred in our Battlefield 3 testing as well, but did NOT occur in Left 4 Dead 2, so I am not certain yet what is causing the problem as it is application dependent. After talking with NVIDIA about it briefly, they were able to replicate the issues but couldn’t yet say if the bug was with Intel’s GFX driver, NVIDIA’s driver or perhaps something to do with the hardware bandwidth limitations.
Screenshots and Evidence
So what exactly is going on? After diving into the recorded video from our Frame Rating captures, we found these interesting results.
What are you seeing in the first image, and then zoomed in on in the second image, is a kind of frame transfer bug that causes a frame to be partially sent to the frame buffer. That transfer gets stalled or halted momentarily and thus the previous frame (in purple) is sent to the display again until the actual next frame (olive-ish) is fully transferred. While you might miss this in playing the game, if ENOUGH of it happens you will definitely experience this tearing. If you look back at the Frame Times graph for DiRT 3 above, any time you see the blue line jut down suddenly, we are seeing one of these "tears".
Here is another example of the problem we saw with the HD 4600 + GTX 760M combination. This time, the shift happens between the yellow and orange overlay segments.
The tearing of these frames is obviously an issue for NVIDIA and its Optimus technology that needs to be addressed. NVIDIA has been very proud of its frame pacing technology on the desktop side but my guess is that this is an unintended consequence of Optimus. In my gameplay testing the performance of the GTX 760M is so much higher than that of the integrated graphics that going discrete is still a much better solution for a mobile user with a focus on gaming.
It seems to be an Optimus specific issue. When I did my testing with the MSI GT60 that came with the GTX 780M graphics, it did not exhibit any of these problems. It did NOT utilize Optimus technology though – something I thought was odd when first receiving the sample; now it kind of makes more sense. Also, depending on the exact configuration of the notebook, some external outputs can bypass the Optimus data path and connect directly to the discrete GPU’s frame buffer, relieving the need to copy data from discrete to integrated graphics memory. The notebook display though, cannot be bypassed in that way.
We’ll be doing more testing with additional notebooks (both AMD and NVIDIA discrete based) in the next couple of weeks so you can expect some more analysis of this issue. Also, once NVIDIA has some kind of official response on the topic, we’ll update our story as necessary.