[gamegpu.ru] Game CPU performance over time (126 game meta study)

[BrightCandle]

As many of you are aware I have spreadsheet (https://dl.dropboxusercontent.com/u/3638175/GameGPU CPU performance.ods) of gamegpu.ru data that I have captured from their graphs for various CPUs across a wide range of games. I used to have 56 games on there that was into last year, I have updated it to the latest (took 3 hours) and I now have 126 separate games. Some games are represented twice as they were tested in beta, there is also some Unreal 4 engine tests in the data as well.

Gamegpu.ru added some Haswell CPUs so we now have 50 games with the 4770k, 4670k and i3 4330 for comparison to the sandy bridges (2600k, 2500k, i3 2100 and 3670X and 3930k).

How many cores
First up the 6c12t v 4c8t v 4c4t v 2c4t:

* The maximum gain for the i3-2100 v the 2500k is 250% but the graph is cut shorter to show the other differences as that peak tends to dominate.

We can clearly see that an i5 has a substantial improvement over a dual core. It can have a very significant impact when its 2.5x faster. The 2600k however shows marginal improvements over the 2500k, but in a few games we can still see a decent jump, but most of the time its a lot less and around 5%. The 3970X v 2600k is interesting as well, we see small improvements almost all the time and the maximum gain is reasonable but not quite to the 50% theoretical we would expect from perfect scaling.

Haswell doesn't change much, it closes the gap with the 6 core CPU and its i5 is less of a jump over the I3:

Core use over time

I also wondered if there was any a trend towards higher CPU core usage. Unfortunately neither of the hexcores covers the entire range like the 2600k, so instead I have two graphs. The first is the 3930k v 2600k which is for the older games (going back to Guild wars 2). Then the 3970X v 2600k which started with the Wasteland 2 beta.

There isn't a big obvious trend. There does appear to be a climb in the number of games gaining from 6 cores to a performance improvement around 10% in the recent games. The number of games benefiting a lot seems to have dropped but the average gain has improved.

8350 v 2600k

I also wondered how well did the 8350 compare over time to the 2600k. Considering the trend in time to have a slight improvement in 6 core performance improvements we would anticipate that the 8350 would similarly gain on the 2600k over time.

Hmmm there is definitely a trend there but its not looking good for the 8350. The 2600k is pulling an ever growing lead against the 8350 in more and more games except in a few cases (those last few games).

Haswell upgrades from Sandy Bridge

My final comparison for now is how good is Haswell compared to Sandy Bridge. We have 50 results so there is enough for a decent look at just how different they are.

So Haswell is worth on average about 5% in the quad cores and 10% in the dual core. The best gains are decent but most of the time the difference is very minor. When we consider the Sandy Bridge CPUs overclock better this actually could mean most of the time you would be better off with an overclocked SB verses an overclocked Haswell (4.8Ghz v 4.4 is a 9% difference).

Usual caveats apply with gamegpu.ru data. They capture games at particular points, its almost always single player but not always and its across a variety of graphics cards , some of its SLI and some of its just a single high end Nvidia card like a 780 ti. But its still the best overall source for a wide variety of games testing I know of.

 

[Homeles]

Usual caveats apply with gamegpu.ru data. They capture games at particular points, its almost always single player but not always and its across a variety of graphics cards , some of its SLI and some of its just a single high end Nvidia card like a 780 ti. But its still the best overall source for a wide variety of games testing I know of.

Wish TPU would do for CPUs what they do for GPUs. Their gaming benchmark suite is enormous (though not as large as gamegpu.ru's, it appears), and I really like their aggregate charts.

Perhaps the most interesting to me is just how awful Zambezi runs, and how huge of an improvement Vishera was over it, in many cases. There's multiple instances of 40%+ improvements. Renegade X shows a 53% improvement... damn... the difference one year can make.

One thing I should point out... there are seemingly a lot of games that cap out at 60FPS. These would water down the improvement numbers a fair bit. Not trying to say they're not valid or anything... just goes to show that you really need to keep your needs in mind. If you're buying a CPU to play Titanfall, for instance... save your money and buy something less expensive, as you'll still get the same performance.

 

[ochadd]

Thanks for the analysis. Surprised this hasn't got more attention.

 

[Lepton87]

Thanks for the analysis. Surprised this hasn't got more attention.

+seconded

very interesting, I already linked this thread in another. What I wonder about is if INTEL HEDT platform advantage comes down to more cores or more cache because in some games it's the second. Here's a very glaring example
3.1GHz 20MB L3 beating 3.5GHz 15MB L3 in a 2-threaded game

I've been thinking if it's just about the cores not cache and whatnot why Quad-Core HT enabled CPUs aren't really faster then those without HT except in a few cases on the other hand there are a few cases where enabling HT decreases performance even though the scheduler should be intelligent enough to assign cpu heavy threads to real cores and then look at 2C/2T and 2C/4T CPUs, the latter just obliterates the former. 2T pentium at 4.7GHz just equals Core i3 at 3.5GHz.

So the argument that games don't care about virtual cores can be dismissed based on that. I was actually surprised how well 3.5GHz HT performed compared to 4.7GHz non-hyper-threaded CPU. It's 34% more frequency!!!