TLDR: Yes, it scales linearly (mostly). Also AMD is still not a bargain.
When I first started fooling around with my graphics card’s settings, and trying to quantify performance enhancements from overclocking, I was primarily using 3DMark11 and Unigine Heaven. While that did the job, I don’t think it painted a good picture of the actual results behind the numbers. There were a number of nuances I couldn’t observe, and restarting my computer between runs for consistency was not only slow, but probably marginally effective at best.
And that’s where the free version of 3DMark is actually a benefit. When I first ran the benchmark, I was rather annoyed that I couldn’t skip the Demo loops that ran before each benchmark. I thought of it as a waste of time. But as it turns out, that demo loop means I can change GPU settings on the fly and re-run my benchmark without a reboot. Each demo loop essentially pre-caches all the assets for each run so nothing is ever hard-drive constrained.
I’m not going to go over all the testing methodologies, motivations, etc, so here’s a quick synopsis of everything important before I dump numbers and graphs at you:
For these tests, I’m running my 660 Ti at 114% TDP, my Core i7-3770k at a 4.3 GHz overclock, and varying the core and memory clocks on my GPU by 25 and 100 MHz increments respectively.
I’m not testing for stability here, I’ve found that booting into Windows is usually problem-free while resuming from sleep can be problematic. I suspect the instantaneous voltage when trying to re-load all of Win8’s assets from sleep is insufficient and leads to memory corruption, whereas a cold start allows the GPU to reach a stable voltage before it starts keeping track of anything important. So these numbers are merely indicative of what the Kepler silicon is capable of if you happen to get a good unit, rather than overclocking guidelines.
I wanted to de-couple the GPU results in 3DMark from the CPU results so I could more accurately track performance scaling without worrying about the inherent weighting in 3DM’s “composite” score. So I tracked all benchmark values to ensure that the CPU was pushing out the same performance regardless of the GPU settings. There are some minor variations, but when you zoom out, you can see that the actual difference in performance between runs is a fraction of a percent. So with that out of the way, let’s get to the actual results.
Ice Storm (DX9):
Alright, so this test is just goofy. The DirectX 9 API doesn’t really contain anything that will reasonably tax a modern gaming PC. But what this test is good for is allowing a direct comparison between Android devices (which have a comparable OpenGL API for games) and PCs.
The Composite/Combined score for a bleeding edge Android handset like the HTC One in the Ice Storm benchmark is 11,309. That’s a 1.5GHz Dual-Core Snapdragon S4 processor being pushed to the limit. A Surface Pro Tablet, which is really just an anemic, amputated ultrabook, scores about 3 times higher (numbers from Anandtech). The lowest score for my PC is 159,731. If there was ever any thought that our smartphones are catching up to “real” computers in terms of horsepower, please. Kill those notions now. My “moderate” gaming rig is over 14 times more powerful than one of the top smartphones of 2013.
Also, in regards to the apparent slump in performance at higher clock speeds, I alluded to it in my previous post, but let me use this metaphor:
Imagine the Ice Storm benchmark as riding your bike in first gear, since DX 9 provides so little resistance. You can very easily peddle faster, but after a point the human body just cannot provide any gains. A 5 year old could peddle a zero-resistance bike just as fast as Doped Armstrong, but does that really tell you who’s better in a 2-mile uphill straight?
Cloud Gate (DX10):
There’s really not much to be said here. We’re still seeing some non-linearities… but what’s the point?
Fire Strike (DX11):
This is the real test and as you can see, Fire Strike’s heavy barrage of effects tease out the true potential of hardware. At the upper end of the results (+75MHz CPU, +500 MHz Mem) it is possible to tease out an extra ~10 percent from your hardware. Which means if you’re looking for a linear price-performance ratio, if a factory-overclocked card costs more than a 10 percent premium it’s probably not a good deal.
There’s no real conclusion to any of this, except maybe a consolation: If you can’t get your card to OC very far, you’re not missing out on *too* much.
Tangent: The AMD 7950 was recently seen at the price point of $269, which is what I got my 660 Ti for a few months ago. Toms Hardware and a bunch of other sites consider the 7950 a level above the 660 Ti in terms of performance. But like I said last week… or was it 2 weeks ago… AMD just flat out sucks compared to nVidia in certain applications, including Portal 2 and Battlefield 3, both of which are dear to my heart. So no, AMD cards are STILL NOT a bargain in case you were wondering.
P.S. I know I didn’t label the axes. Get over it.