Gamegpu.ru CPU data distilled

[BrightCandle]

I have gone through all the games that gamegpu.ru have tested and produced a spreadsheet containing key processors and average frame rates. That is the 3970X, 3930k (where 3970X is missing), 2600k, 2500k, 8350, 8150 (where 8350 result is missing) and the i3-2100.

Gamegpu.ru uses canned benchmarks and they do the tests at very high or ultra quality at 1080p. They have used 2 different GPUs over the time period from September last year. The results last year were done on a 690 and the more recent results are on a Titan.

Given the caveats that they aren't necessarily the best tests I am providing a summary of the data. You can find the original spreadsheet with all the numbers (https://dl.dropboxusercontent.com/u/3638175/GameGPU CPU performance.ods) but its just a lot of data points to churn through.

The table below is a summary of the results, showing the percentage of performance of the processor on the left compared to the processor on the right. So a result of 103% means the 39X0k is 3% faster on average compared to a 2600k. But in 25% of games it is 7% faster and at most its 26% quicker.

               39X0k v 2600k  2600k v 2500k  2500k v i3-2100  2600k v 8X50
Average        103%           105%           129%             141%
Maximum        126%           121%           235%             215%
25% Quartile   103%           104%           123%             134%
75% Quartile   107%           107%           144%             170%

The 39X0k v 2600k is a comparison of 6 cores to 4 cores + HT. The 2600k v 2500k is a comparison of hyperthreading. The 2500k v i3-2100 is a test of true quad core verses a dual core with HT. 2600k v 8X50 is a test of Intels mainstream top chip verses AMDs top chip.

 

[fixbsod]

So looks like the 2 extra cores in 39x0k vs 2600k gives less perf (3%) than 4 cores w/HT compared to w/out HT (5%). Until they start dropping lots of 6+ core cpus it looks like software will forever be 4 core only. This whole 'consumers' don't need it BS needs to stop, it perpetuates the cycle -- software would be optimized >4 cores if there were more CPUs with > 4 cores on the market.

 

[ShintaiDK]

So looks like the 2 extra cores in 39x0k vs 2600k gives less perf (3%) than 4 cores w/HT compared to w/out HT (5%). Until they start dropping lots of 6+ core cpus it looks like software will forever be 4 core only. This whole 'consumers' don't need it BS needs to stop, it perpetuates the cycle -- software would be optimized >4 cores if there were more CPUs with > 4 cores on the market.

Yes, Amdahls law and increased development cost would magically go away. :whiste:

 

[Ajay]

Yes, Amdahls law and increased development cost would magically go away. :whiste:

Amdahl's law isn't the problem here. Development/QA costs are the problem as you point out. Allot of code in game engines can be multithreaded, it's just a b*tch to actually do correctly with high performance code. For example - 'server: I need the state of every object now', 'code: objects 12 & 14 return null because they haven't completed a threaded call', so more syncing and/or lock operations are strewn throughout a cascade of functions and that can kill any net performance gain.

Haswell provide a possible solutions to this problem, we'll know over the next couple of years if it works well enough to change development processes towards including more threading.

 

[BrightCandle]

Personally I think the results are quite interesting. We can make some fairly well supported statements that contradict some arguments we see made here often. Like

A 6 core chip does have gaming benefits, but due to their lower clock speed in about 25% of games they are slightly slower. However on average and certainly with 25% of games there is some moderate benefit to 6 cores over a 2600k. If you look at the data you will see modern games show more positive impact than old ones. Over clocking is a good idea here on a 3930k as the 3.2Ghz base clock is impacting the results on games that aren't scaling well.

A 2600k is worth it for gamers over the 2500k. Only in a single case does modern ht slow down the processor, on average it gets you half the benefits of spending on a hexacore for a lot less money.

A dual core i3 despite its clock speed is almost always dramatically slower than a quad core. Dual cores are no longer good enough for gaming.

The 8350 rarely performs well. In a couple of specific games it does well but in general its much slower than an Intel quad core and often behind a dual core. Its not a good budget gamers CPU.

 

[parvadomus]

Amdahl's law isn't the problem here. Development/QA costs are the problem as you point out. Allot of code in game engines can be multithreaded, it's just a b*tch to actually do correctly with high performance code. For example - 'server: I need the state of every object now', 'code: objects 12 & 14 return null because they haven't completed a threaded call', so more syncing and/or lock operations are strewn throughout a cascade of functions and that can kill any net performance gain.

Haswell provide a possible solutions to this problem, we'll know over the next couple of years if it works well enough to change development processes towards including more threading.

Haswell has nothing to do here. Synchronization in a multithreaded program can be done by software.
And Amdahl's law is almost irrelevant for games, lazy programmers are the problem, and that will change if they want to extract the true potential of the slow Jaguar cores in PS4 and XBOX.

 

[USER8000]

Any chance you could do an FX8350 vs FX8150 comparison too??

 

[bononos]

Surprised to see that the 8350 wasn't that good after tabulating the numbers.

It'll be interesting to see if there is going to be a generational jump in performance for 6/8x cores in the near future because of the ps4/xbox720 effect.

 

[Lepton87]

The 39X0k v 2600k is a comparison of 6 cores to 4 cores + HT. The 2600k v 2500k is a comparison of hyperthreading. The 2500k v i3-2100 is a test of true quad core verses a dual core with HT. 2600k v 8X50 is a test of Intels mainstream top chip verses AMDs top chip.

You will get flawed results from this, I'm gonna give just one example to give people an idea why

2600kvs2500k

Frequency difference
Cache difference

Check 2600K with and without HT to measure what you try to.

 

[USER8000]

Surprised to see that the 8350 wasn't that good after tabulating the numbers.

It'll be interesting to see if there is going to be a generational jump in performance for 6/8x cores in the near future because of the ps4/xbox720 effect.

It includes the FX8150 numbers,so that is why it would be interesting to see the FX8350 against the FX8150.

 

[Lepton87]

It includes the FX8150 numbers,so that is why it would be interesting to see the FX8350 against the FX8150.

Sometimes Vishera is so much faster then BD in games that mixing and matching them on the same chart was a very bad idea, it's not that unusual for vishera to be 20% faster in some games.