Impact of memory speed on gaming performance

[MageTank]

I would like to tag @SteveGrabowski0 and @Curufinwe_wins for motivating me to do these tests. I would also like to point towards this video for my reason behind wanting to test this.

 

So before we begin, I have to make a quick disclaimer. The results posted below are using a Pentium G4400 dual core Skylake CPU. There are some situations where i am bound by the lack of cores, and simply cannot test certain AAA titles because of this fact. Results may vary on stronger CPU configurations, but below are what i have personally experienced with my current setup. I will come back to test this again once I get a Core i7 6700k. Now, on to the configuration.

 

I used the following hardware in my tests, as it is the best of what i have in my possession:

 

Motherboard: Asrock Fatal1ty Z170 Gaming ITX/AC

CPU: Intel Pentium G4400 3.3ghz

Memory: Panram Ninja-V 8GB (2x4GB) 2133mhz CL13 (overclocked to 3200mhz CL15-15-15-30-CR1 1.37vDIMM)

GPU: MSI GTX 770 (Reference design)

Storage: Samsung 850 Pro

 

To get a more accurate representation of the effect memory speed has on gaming performance, i decided to push the memory as far as i could take it within safe voltage limitations without crashing after 12 hours of custom prime95 and memtest86 runs. I managed to get the memory to 3200mhz, CL15-15-15-30-CR1 at 1.37vDIMM, and 1.070 VCCIO, and 1.12 VCCSA. I tightened every single timing, from primaries to tertiary for this setup, and used the exact same settings for each memory speed below 3200, only changing the memory frequency and not the timings to go with it (Except RTL and IO-L, which have to be adjusted alongside the memory speed). For example: 2133mhz would also be using 15-15-15-30-CR1 even though it defaults to CL13 in XMP. 

 

Here is what my current CPU + 3200mhz memory kit does in Aida64's cache and memory benchmark: 

 

Now that we have got that out of the way, lets get on to the gaming results. 

 

Thief: http://imgur.com/a/wR5es

 

From these results, we do see a steady increase in all 3 frame rates. Min, Max, and Average. However, once memory speed reached around 2800mhz, it showed heavy diminishing returns, in which 3066 to 3200 made little to no difference no matter how many times i re-ran the benchmarks. Still, we can look at 2133mhz vs 3200mhz, and see a clear 10-11% difference in minimum and maximum frames, and an 8% difference in average frames. Quite a boost in performance.

 

Tomb Raider: http://imgur.com/a/pH6wO

 

I only did two tests on this game, as it did not yield that large of a difference for me to warrant several different tests in speeds, as the results seem very similar to what we saw with Thief. The difference in minimum frame rates was 10%, but max frame rates were only 5.5% apart. Average frame rates were only 6.8% apart from each other. Though, i can come back to this and test different speeds if you guys feel i should. 

 

Metro 2033: http://imgur.com/a/Uhs8c

 

This is where we see little to no difference between 2133 and 3200mhz memory. Some games simply do not care for the speed difference, or maybe my Pentium was running into problems with the game itself, as no memory configuration changed the end result. 

 

Shadows of Mordor: http://imgur.com/a/LKYmp

 

This is the most confusing of my benchmarks. The results make little to no sense, and are all over the place. For some reason, no matter how many times i re-test this specific game, i cannot get an accurate baseline result, even if i factor in averages, nothing turns out right. One test, i can get 300+ max frames, and on the next, i get only 140 max frames. However, if we compare just the minimum frame rates, 2400mhz seems to be the cap where memory stops making a difference in this specific game. It made roughly a 15-20% difference on average every single time i tested it, over 2133mhz. If anyone knows of a better way to bench this game other than the very inaccurate tool provided in this game, i would appreciate it.

 

Conclusion: 

 

Memory speed does have an impact on gaming performance. I can feel it, and you can see it from the results above. However, the impact differs greatly depending on the title, and perhaps even your hardware configuration. Memory seems to make the biggest difference when it comes to minimum frame rates, an aspect that just might help SLI configurations more than most other configurations. In Thief and Tomb Raider, we saw on average, a 10% difference in minimum frame rates, which is arguably the most important number in gaming. When you factor in costs, take a look at the following kits, and their difference in price:

 

2133mhz: http://www.newegg.com/Product/Product.aspx?Item=N82E16820148857&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=

 

2800mhz: http://www.newegg.com/Product/Product.aspx?Item=N82E16820232087&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=

 

3200mhz: http://www.newegg.com/Product/Product.aspx?Item=N82E16820231939&nm_mc=AFC-C8Junction&cm_mmc=AFC-C8Junction-PCPartPicker,%20LLC-_-na-_-na-_-na&cm_sp=&AID=10446076&PID=3938566&SID=

 

The difference between 2133 and 2800mhz, is only $12. $12 for up to 10% more minimum frame rates in gaming. This is not even including how helpful it would be in non-gaming scenarios. Yes, i know that the difference between the cheapest 2133mhz 8gb stick and 2800mhz 8gb stick is 26.5%, but if you factor in the performance gained into your total budget of lets say, $500, a $12 difference is only 2.5% of your total budget, for 10% more FPS. I also know board costs can differ between H and Z series motherboards, but you also have to take into consideration the performance benefits memory will give you, along side the upgrade paths and overclocking performance of other components such as the CPU. 

 

I would say faster memory is worth the investment, but that i would not worry too much about speeds beyond 2800mhz for gaming. It seems to be the sweet spot for getting the most performance out of your ram. At least, for my current setup anyways. I hope this was helpful to some people, and i encourage you to test this for yourself, as it really is something you will have to see to believe. 

[Lays]

Would you like me to test with 1600 CL9 vs 2400 CL9, or even 2800 CL9?

 

(ddr3)

[MageTank]

Would you like me to test with 1600 CL9 vs 2400 CL9, or even 2800 CL9?

 

(ddr3)

More information is always helpful. I contemplated putting my G3258 back together and testing DDR3, i just don't have the time right now. It took me over a week to do the tests i did above due to my busy schedule. I would appreciate it if you could test DDR3 with that awesome i7 of yours. Would help shed light on the CPU bottlenecks i mentioned above.

[Lays]

More information is always helpful. I contemplated putting my G3258 back together and testing DDR3, i just don't have the time right now. It took me over a week to do the tests i did above due to my busy schedule. I would appreciate it if you could test DDR3 with that awesome i7 of yours. Would help shed light on the CPU bottlenecks i mentioned above.

 

I only have tomb raider & shadows of mordor, and IMO SoM is a POS bench, its inconsistent as fuq

 

I MIGHT be able to do 2933 mhz CL9, I am totally not sure if it will even function long enough to do a tombraider bench though.

I barely got it to run cinebench, so it "hopefully" will run tomb raider.

 

 

Although my secondary and tertiary timings at 2933 9-12-12-25 are looser than 2800 9-12-12-17, so performance difference isn't a big gap.

 

 

 

What do you want me to run, stock CPU with 1600 9-9-9-24 (this is very common speed) then stock 2400 XMP 9-11-11-31, then 2800 9-12-12-17?

 

Then OC CPU @ 5 ghz with same stuff?

 

(Honestly I can probably run tomb raider bench at 5.2 ghz if it's needed)

[SteveGrabowski0]

Thanks @MageTank for your testing. Your results combined with Digital Foundry's make the world of pc building seem a lot murkier now. It was so easy before back when it seemed like RAM speed didn't matter, you just grabbed the cheapest kit from a decent manufacturer that would look good through your window. But now it's starting to look like Z series chipsets offer a lot more value than previously thought. The difference in price in the RAM itself is pretty minimal even at 16GB unless you're just going for top of the line kits, so the main expense here is in the board. All of a sudden H170, H97, and the like don't look like such great values. I mean 10% is a lot and something you would reasonably think most people would be willing to spend $50 extra for.

[MageTank]

I only have tomb raider & shadows of mordor, and IMO SoM is a POS bench, its inconsistent as fuq

 

I MIGHT be able to do 2933 mhz CL9, I am totally not sure if it will even function long enough to do a tombraider bench though.

I barely got it to run cinebench, so it "hopefully" will run tomb raider.

 

 

Although my secondary and tertiary timings at 2933 9-12-12-25 are looser than 2800 9-12-12-17, so performance difference isn't a big gap.

 

 

 

What do you want me to run, stock CPU with 1600 9-9-9-24 (this is very common speed) then stock 2400 XMP 9-11-11-31, then 2800 9-12-12-17?

 

Then OC CPU @ 5 ghz with same stuff?

 

(Honestly I can probably run tomb raider bench at 5.2 ghz if it's needed)

Honestly, the OC CPU wont matter. You can even start off at 5ghz if you want. The only thing that needs to be tested, is the difference in frame rates when changing the raw memory speed. I am personally going to do a few more tests and see if cas latency has any effect on gaming, and see if the performance comes from the raw memory bandwidth itself. Like i said before. More information is always helpful, even if it may seem inconsequential.

[Lays]

Here is my normal 24/7 speeds @MageTank

 

I can do another in a few mins with higher mem speeds.

 

( This is 5 ghz CPU, 4 ghz Cache, 2400 mhz RAM 9-11-11-32 2T )  I should be able to do a run at 5.1 CPU, 4.9 Cache ~2860 Mhz RAM @ 9-12-12-17 1T

 

[Prysin]

in short. If mobo can support it. the 10 bucks it cost to get 2400MHz over 1600MHz is DEFINETIVELY WORTH IT.

 

thanks for doing this

@MageTank

[Curufinwe_wins]

@MageTank interesting results and honestly within expectation. Clearly their is a point for this particular cpu where faster ram just doesnt help much, but that speed appears to be much higher than I personally expected (2800 vs 2400). That said I wish my imc was decent on my 5820k then I'd totally try this test myself on quad channel (not that I disagree with Anandtechs article on the matter, but I have to wonder if newer games care more. Plus I have the gpu strength to force cpu bottlenecks at 1080p and maybe on some games at 1440p).

[Prysin]

Thanks @MageTank for your testing. Your results combined with Digital Foundry's make the world of pc building seem a lot murkier now. It was so easy before back when it seemed like RAM speed didn't matter, you just grabbed the cheapest kit from a decent manufacturer that would look good through your window. But now it's starting to look like Z series chipsets offer a lot more value than previously thought. The difference in price in the RAM itself is pretty minimal even at 16GB unless you're just going for top of the line kits, so the main expense here is in the board. All of a sudden H170, H97, and the like don't look like such great values. I mean 10% is a lot and something you would reasonably think most people would be willing to spend $50 extra for.

which is also why i say Haswell i3 is dead...

 

because cheapest haswell i3 Z97 board is notably more then the cheapest AM3+ board with OC headroom and 2400MHz DDR3....

 

i wish people would pull their heads out of their asses and read ALL the forums, not just the CPU or GPU forums.

A PC aint a CPU + GPU. It is everything.

Everything matters. From a good mobo, to a SSD, to the RAM and the PSU. It all fucking matters. when will people realize?

[Prysin]

@MageTank interesting results and honestly within expectation. Clearly their is a point for this particular cpu where faster ram just doesnt help much, but that speed appears to be much higher than I personally expected (2800 vs 2400). That said I wish my imc was decent on my 5820k then I'd totally try this test myself on quad channel (not that I disagree with Anandtechs article on the matter, but I have to wonder if newer games care more. Plus I have the gpu strength to force cpu bottlenecks at 1080p and maybe on some games at 1440p).

preliminary tests show that Fallout4 enjoys faster memory.

[Curufinwe_wins]

preliminary tests show that Fallout4 enjoys faster memory.

I know. (At least on dual channel and up to 2666 at least.)

But quad channel....

[Lays]

@MageTank

 

 

Results are a bit weird IMO, but 5 ghz with 2800 CL9 was a big improvement IMO.

 

 

 

 

4.4 ghz CPU 1600 mhz RAM @ 9-9-9-24

 

 

 

5 Ghz 1600 9-9-9-24 RAM

 

 

 

 

 

 

4.4 ghz 2400 9-11-11-31 (XMP)

 

 

 

5.0 XMP 2400 9-11-11-31

 

I found this one was the weirdest, I think I should of re-ran it because it seems not in line with the other tests.

 

 

 

4.4 2800 Mhz RAM 9-12-12-17 1T super tight secondary & tertiary timings

 

 

 

 

5 Ghz  2800 Mhz RAM 9-12-12-17 1T super tight secondary & tertiary timings

 

 

 

 

 

 

 

5 GHZ CPU, 2400 9-11-11-31 2T RAM:

 

 

 

 

 

5 Ghz CPU, 4.8 Ghz Cache, 2800 Mhz RAM 9-12-12-17 1T

 

[Prysin]

@Lays

In short: safer to buy 2400MHz DDR3 then risk frying your 330 USD CPU to get 2 FPS better minimums

[Lays]

@Lays

In short: safer to buy 2400MHz DDR3 then risk frying your 330 USD CPU to get 2 FPS better minimums

Not gonna fry a cpu by running faster rams with more ram voltage.

The memory controller voltages are not effected by RAM voltage.

The memory controller voltages are system agent, iod/ioa volts and aren't increased by increasing ram voltage, ram voltage is all controlled by the motherboard.

[Prysin]

Not gonna fry a cpu by running faster rams with more ram voltage.

The memory controller voltages are not effected by RAM voltage.

The memory controller voltages are system agent, iod/ioa volts and aren't increased by increasing ram voltage, ram voltage is all controlled by the motherboard.

was referring to CPU at 5GHz rather then stock/turbo

[Lays]

was referring to CPU at 5GHz rather then stock/turbo

I'm still within safe volts for the 5 Ghz, only 1.335v

Although this chip is super above average lol

[MageTank]

Not gonna fry a cpu by running faster rams with more ram voltage.

The memory controller voltages are not effected by RAM voltage.

The memory controller voltages are system agent, iod/ioa volts and aren't increased by increasing ram voltage, ram voltage is all controlled by the motherboard.

I think @Prysin meant that you can get just as decent of a performance boost by going from slow memory, to fast memory, than what you could when trying to push a CPU from 4.7 to 5.0ghz. Granted, as shown above in my post, it depends entirely on the game itself. I imagine an MMO would benefit greatly from CPU clock over memory clock. Either way, its nice to know that investment in faster memory is not a wasted endeavor.

[Lays]

I think @Prysin meant that you can get just as decent of a performance boost by going from slow memory, to fast memory, than what you could when trying to push a CPU from 4.7 to 5.0ghz. Granted, as shown above in my post, it depends entirely on the game itself. I imagine an MMO would benefit greatly from CPU clock over memory clock. Either way, its nice to know that investment in faster memory is not a wasted endeavor.

I think it would probably benefit from memory speed too, because anything that's cpu intensive almost always benefits from le mems

[Curufinwe_wins]

I think it would probably benefit from memory speed too, because anything that's cpu intensive almost always benefits from le mems

So you really didn't show much difference at all 1-3%... Now that was an i7...

 

That pretty much falls in line with anandtech's haswell memory speed comparison. (COM'ON DO A SKYLAKE VERSION)

 

They found for DDR3 on an 4770k, tighter timings seemed to be the bigger thing, and that 1866 CL9 was within 1-2% of the best of the best (or was the best itself on every test).

 

http://www.anandtech.com/show/7364/memory-scaling-on-haswell

[MageTank]

So you really didn't show much difference at all 1-3%... Now that was an i7...

 

That pretty much falls in line with anandtech's haswell memory speed comparison. (COM'ON DO A SKYLAKE VERSION)

 

They found for DDR3 on an 4770k, tighter timings seemed to be the bigger thing, and that 1866 CL9 was within 1-2% of the best of the best (or was the best itself on every test).

 

http://www.anandtech.com/show/7364/memory-scaling-on-haswell

It should be noted that Skylakes memory controller is far more complex than Haswells. There are timings on my Z170 board that even X99 boards do not have. Much finer control, and its still easier. They did something with Skylake that made overclocking less tedious. For some reason, my memory was much more stable when overclocking when compared to my Haswell DDR3 kit, and yet i did the exact same methods to my overclocking.

 

@patrickjp93 has been saying they changed the internal latencies of Skylake, so maybe the answer lies within that fact? I don't really know for certain. I can only show what i've experienced above. I can try with my G3258 and DDR3 memory if you want. It would just take me another week or so to do it, as i am very busy this week.

[Lays]

It should be noted that Skylakes memory controller is far more complex than Haswells. There are timings on my Z170 board that even X99 boards do not have. Much finer control, and its still easier. They did something with Skylake that made overclocking less tedious. For some reason, my memory was much more stable when overclocking when compared to my Haswell DDR3 kit, and yet i did the exact same methods to my overclocking.

 

@patrickjp93 has been saying they changed the internal latencies of Skylake, so maybe the answer lies within that fact? I don't really know for certain. I can only show what i've experienced above. I can try with my G3258 and DDR3 memory if you want. It would just take me another week or so to do it, as i am very busy this week.

 

 

Did you adjust RTLs?

Your board may/may not allow changing them, but it is round trip latency from CPU sending shit to ram and then back to CPU I believe.

A lot of boards set them higher than they need to be when increasing Mhz, and lowering them will give even more performance.

[MageTank]

Did you adjust RTLs?

Your board may/may not allow changing them, but it is round trip latency from CPU sending shit to ram and then back to CPU I believe.

A lot of boards set them higher than they need to be when increasing Mhz, and lowering them will give even more performance.

Yes, i have my RTL's super tight. If you loosen the IO-L Offset on DDR4, you can get tighter IO-L and RTL's. I took my IO-L offsets from 21, to 26, and managed to get my IO-L (both channels) down to 1, and RTL's down from 54/54, to 52/52. Going lower fails to post, but this current setup has survived 12 hours of prime95 512-4096mb FFT's and 12 hours of memtest86 (16 passes, all 13 tests).

 

If you want some screenshots of my memory settings from within my BIOS, i will gladly share them for anyone to compare with. Maybe we might be able to find something to test against.

[Lays]

Yes, i have my RTL's super tight. If you loosen the IO-L Offset on DDR4, you can get tighter IO-L and RTL's. I took my IO-L offsets from 21, to 26, and managed to get my IO-L (both channels) down to 1, and RTL's down from 54/54, to 52/52. Going lower fails to post, but this current setup has survived 12 hours of prime95 512-4096mb FFT's and 12 hours of memtest86 (16 passes, all 13 tests).

 

If you want some screenshots of my memory settings from within my BIOS, i will gladly share them for anyone to compare with. Maybe we might be able to find something to test against.

 

 

Here are some stuffs you can look at to compare, they are running much higher volts tho. (Like 1.6-2.2v on ddr4 depending on what memory chips)

 

http://forum.hwbot.org/showthread.php?t=144801

 

http://forum.hwbot.org/showthread.php?t=148402

 

http://forum.hwbot.org/showthread.php?t=148427

 

http://forum.hwbot.org/showthread.php?t=147417

 

http://www.overclock.net/t/1268061/ocn-ram-addict-club-gallery/6920#post_24612829

[MageTank]

Here are some stuffs you can look at to compare, they are running much higher volts tho. (Like 1.6-2.2v on ddr4 depending on what memory chips)

 

http://forum.hwbot.org/showthread.php?t=144801

 

http://forum.hwbot.org/showthread.php?t=148402

 

http://forum.hwbot.org/showthread.php?t=148427

 

http://forum.hwbot.org/showthread.php?t=147417

 

http://www.overclock.net/t/1268061/ocn-ram-addict-club-gallery/6920#post_24612829

I am well aware of the sorcery Alex and Bullant are doing on their kits. However, i am absolutely terrified to try that voltage on my setup. I am going to wait until i see what kind of impact that high vDIMM will have on this cheap PANRAM kit of memory that i have. Once i get my hands on the 32GB (2x16GB) of ram i need for my ITX board, i might torture this PANRAM kit for the fun of it. Until then, i will play it safe.