[Eurogamer] CPU Bottleneck, perhaps its actually RAM?

[Headfoot]

Link: http://www.eurogamer.net/articles/digitalfoundry-2015-intel-core-i3-6100-review

Much digital ink has been spilled on what CPU is fast enough to not bottleneck a particular graphics card. But perhaps we need to broaden that inquiry.

In the link above, in dual-core i3 scenarios (but interestingly, not in quad core i5 scenarios) a reduction in RAM speed can significantly decrease FPS, but that same CPU can close ranks with the i5 with sufficiently fast RAM. This happens with both Haswell and Skylake i3's, which indicates this isnt limited to DDR4 or DDR3.

For example, in Ryse we see Skylake i3-6100 (2666MHZ) = 103.2 FPS, but with 2133 MHZ - 58.5. Other games show smaller drops but still appreciable. Significantly more than the old standby figure of 2-5% that used to be true.

It seems Fallout 4 is not as much of an outlier as we originally believed. I dont know what is causing this gap, but it seems logical to assume that because games are designed to utilize a single unified fast pool of memory on the new consoles, that this design is bleeding into increased memory bandwidth requirements on PC.

IMO the new answer to "is this GPU bottlenecked by this CPU" needs to include consideration for RAM speed.

 

[Ken g6]

[thread=2452438]Repost[/thread]

 

[Headfoot]

That is in the CPU forum. This is more particularly for talking about GPUs getting bottlenecked by cpus with that article as a source. Both for Skylake, Haswell, anything really. Different topic altogether.

 

[Techhog]

Literally anything can be a bottleneck for Ryse. That game is weird.

 

[Headfoot]

Witcher 3 showed some surprising results too, add in CryEngine 3 games (Ryse as a barometer), plus Fallout 4, and we're starting to see a trend. Just one game is easy to discount, especially given how big the drops are in Ryse. But with Fallout showing similar drops, and Witcher 3 showing a solid 6-7 fps drop, I think something bigger is happening.

 

[gamervivek]

i5s did like 50% better in the single threaded draw calls benchmark in digitalfoundry's testing than an i3, so probably cache is that big of a factor and which would also be a factor here.

 

[Mondozei]

RAM helps a lot more if your CPU is bottlenecked. If your CPU is fast enough, RAM will not help, including your GPU.

There are oddities, Fallout 4 seem to be one, but they are the exception that proves the rule.

 

[tolis626]

Since it's a matter that has concerned me in the past few weeks, I thought I might just throw in my experience.

I recently decided to take the time to properly redo my overclocks, so among other things (Like achieving lower voltages on the core and getting to 4.8GHz), I managed to get my RAM to 2400MHz C10 from its stock 2133MHz C11 or my previous overclock of 2200MHz C9. I don't have benchmarks for the 2200MHz setting, but in Battlefield 4 going to 2400MHz got rid of some sub 60FPS dips in my minimum framerates when using >100% resolution scaling. I also saw some increase in Cinebench scores. I haven't played any other games recently, but if I manage to produce any interesting results, I'll make sure to post them.

In short, there definitely is a link between RAM speed and CPU performance, and it's especially prevalent in Haswell and even more so in Skylake.

PS : @Mondozei, my CPU is a 4790k at 4.7GHz so it's definitely fast enough and has speed to spare. So it's not the issue.

 

[mysticjbyrd]

RAM helps a lot more if your CPU is bottlenecked. If your CPU is fast enough, RAM will not help, including your GPU.

There are oddities, Fallout 4 seem to be one, but they are the exception that proves the rule.

This. It helps the I3.

 

[ShintaiDK]

I wished faster DDR4 speed was there for me when I bought. But 2400Mhz is the highest chips they produce. Anything above is OC and I wanted toa void that.

But maybe in 5 years when they finally make real DDR4 4000 chips. It could be a "SSD upgrade" if I still have the Skylake

It is fun to see the memory development over time. K8 times it was all about latency, then with Core 2 memory more or less didn't matter. And today its all about bandwidth.

 

[happy medium]

This is the reason I bought ddr4 3000 with my i3 6100 and will use that ram with the Kabylake and next gen GPU upgrades next year. .

 

[tential]

Since it's a matter that has concerned me in the past few weeks, I thought I might just throw in my experience.

I recently decided to take the time to properly redo my overclocks, so among other things (Like achieving lower voltages on the core and getting to 4.8GHz), I managed to get my RAM to 2400MHz C10 from its stock 2133MHz C11 or my previous overclock of 2200MHz C9. I don't have benchmarks for the 2200MHz setting, but in Battlefield 4 going to 2400MHz got rid of some sub 60FPS dips in my minimum framerates when using >100% resolution scaling. I also saw some increase in Cinebench scores. I haven't played any other games recently, but if I manage to produce any interesting results, I'll make sure to post them.

In short, there definitely is a link between RAM speed and CPU performance, and it's especially prevalent in Haswell and even more so in Skylake.

PS : @Mondozei, my CPU is a 4790k at 4.7GHz so it's definitely fast enough and has speed to spare. So it's not the issue.

I'm definitely moving to very fast ram on my next upgrade. Can't believe ram speeds matter now, before it was a couple percentage points at most.
I'm not upset I have slow ram, but now I certainly am thinking about upgrading!

I'll need a whole new build soon though and I'll move my old parts to a new server build(I really thought a 15 Bay server would be enough!) so I can at least not have to throw anything out.

 

[bystander36]

I wished faster DDR4 speed was there for me when I bought. But 2400Mhz is the highest chips they produce. Anything above is OC and I wanted toa void that.

But maybe in 5 years when they finally make real DDR4 4000 chips. It could be a "SSD upgrade" if I still have the Skylake

It is fun to see the memory development over time. K8 times it was all about latency, then with Core 2 memory more or less didn't matter. And today its all about bandwidth.

This is what XMP is for (extreme memory profile). DDR4 sticks these days have a built in profile, and to enable it, you just have to push the XMP button. Sometimes you have to mess with some voltage, but that is about it for the most part.

I got a good deal on DDR4 3000mhz and starting to feel relieved that I did.

Edit: unless you are saying the 3000mhz rated sticks weren't out yet.

 

[TheELF]

For example, in Ryse we see Skylake i3-6100 (2666MHZ) = 103.2 FPS, but with 2133 MHZ - 58.5. Other games show smaller drops but still appreciable. Significantly more than the old standby figure of 2-5% that used to be true.

It's because every game nowadays is written for dx12 (well the consoles equivalent) while on the PC we are forced to play them with dx11.
Look at the video which explains the basic benefits of the console APIs,mainly you can stream textures to the GPU while the GPU keeps on rendering frames,if you use riva tuner to monitor GPU usage on a PC you will notice a steep drop in GPU usage whenever you get FPS drops.
Ryse also locks the texture quality depending on vram and you can only change it from the ini,the higher the textures are in mb the bigger the problem.
https://www.youtube.com/watch?feature=player_detailpage&v=H1L4iLIU9xU#t=941

It's not a CPU bottleneck in Ryse neither a RAM issue, RAM just acts as a cache releaving some of the problem.
Look at process hacker in the background,notice I/O total,anytime it spikes gpu usage drops and so do the FPS, also notice how much disk read has risen at the end of the video and this is with the textures forced to the lowest settings.
It's the same reason they now recomend 16GB for batman.
https://www.youtube.com/watch?v=Z0oBZQ_kBwE

 

[beginner99]

Hm so maybe what also needs to be tested is SSD vs HDD and another interesting thing would be Crystalwell parts with the 128 MB L4 cache. Even if the later would be irrelevant for texture streaming.

 

[Head1985]

some my own tests


6700k 4.5ghz
16Gb DDR4 1600 9-10-10 1T/3000Mhz CL 14-15-15 1T
GTx970 1500/8000
win10

War thunder 1.53 CPU test.All max 1920x1080 no AA
1600Mhz

3000Mhz

Watchdogs 1024x768/1680x1050 all max temporal AA
1024x768 1600mhz ram
Avg: 91.961 - Min: 69 - Max: 143
1680x1050 1600Mhz ram
Avg: 85.571 - Min: 67 - Max: 101

1024x768 3000mhz ram
Avg: 111.087 - Min: 85 - Max: 157
1680x1050 3000Mhz ram
Avg: 92.247 - Min: 81 - Max: 104

Crysis3 1024x768/1680x1050 all max SMAA

1024x768 1600Mhz ram
http://abload.de/img/crysis3_2015_11_29_21vkucz.png
1680x1050 1600Mhz ram
http://abload.de/img/crysis3_2015_11_29_21fvuxf.png

1024x768 3000mhz ram
http://abload.de/img/crysis3_2015_11_29_20mvuwn.png
1680x1050 3000Mhz ram
http://abload.de/img/crysis3_2015_11_29_20pauyh.png

Fallout4 1920x1080 all max FOV 90
1600 mhz ram
http://abload.de/img/fallout4_2015_11_29_2ttuun.png
3000mhz ram
http://abload.de/img/fallout4_2015_11_29_2kyu0e.png

Skyrim
1600mhz ram
http://abload.de/img/tesv_2015_11_29_21_30miur6.png
3000mhz ram
http://abload.de/img/tesv_2015_11_29_21_19ucuiu.png

Witcher3 1024x768/1680x1050 allmax DOF and hairworks off
1024x768 1600Mhz ram
Avg: 96.428 - Min: 82 - Max: 111
1680x1050 1600mhz ram
Avg: 75.090 - Min: 63 - Max: 90

1024x768 3000mhz ram
Avg: 109.293 - Min: 90 - Max: 129
1680x1050 3000mhz ram
Avg: 76.151 - Min: 63 - Max: 89

scaling:
Fallout4- 26.5%
watchdogs- 20.7%
War thunder- 15.7%
witcher3- 13.2%
skyrim- 7%
crysis3 5%

 

[pakotlar]

That 58.5 Ryse number for 2133mhz is almost certainly an error. The rest of the deltas are in line with what you'd expect, approximately 80% of the performance if ram was completely the limiting component, with the deltas showing about ~6%.

 

[Headfoot]

some my own tests


6700k 4.5ghz
16Gb DDR4 1600 9-10-10 1T/3000Mhz CL 14-15-15 1T
GTx970 1500/8000
win10

War thunder 1.53 CPU test.All max 1920x1080 no AA
1600Mhz

3000Mhz

Watchdogs 1024x768/1680x1050 all max temporal AA
1024x768 1600mhz ram
Avg: 91.961 - Min: 69 - Max: 143
1680x1050 1600Mhz ram
Avg: 85.571 - Min: 67 - Max: 101

1024x768 3000mhz ram
Avg: 111.087 - Min: 85 - Max: 157
1680x1050 3000Mhz ram
Avg: 92.247 - Min: 81 - Max: 104

Crysis3 1024x768/1680x1050 all max SMAA

1024x768 1600Mhz ram
http://abload.de/img/crysis3_2015_11_29_21vkucz.png
1680x1050 1600Mhz ram
http://abload.de/img/crysis3_2015_11_29_21fvuxf.png

1024x768 3000mhz ram
http://abload.de/img/crysis3_2015_11_29_20mvuwn.png
1680x1050 3000Mhz ram
http://abload.de/img/crysis3_2015_11_29_20pauyh.png

Fallout4 1920x1080 all max FOV 90
1600 mhz ram
http://abload.de/img/fallout4_2015_11_29_2ttuun.png
3000mhz ram
http://abload.de/img/fallout4_2015_11_29_2kyu0e.png

Skyrim
1600mhz ram
http://abload.de/img/tesv_2015_11_29_21_30miur6.png
3000mhz ram
http://abload.de/img/tesv_2015_11_29_21_19ucuiu.png

Witcher3 1024x768/1680x1050 allmax DOF and hairworks off
1024x768 1600Mhz ram
Avg: 96.428 - Min: 82 - Max: 111
1680x1050 1600mhz ram
Avg: 75.090 - Min: 63 - Max: 90

1024x768 3000mhz ram
Avg: 109.293 - Min: 90 - Max: 129
1680x1050 3000mhz ram
Avg: 76.151 - Min: 63 - Max: 89

scaling:
Fallout4- 26.5%
watchdogs- 20.7%
War thunder- 15.7%
witcher3- 13.2%
skyrim- 7%
crysis3 5%

Thanks for the info!

 

[Headfoot]

That 58.5 Ryse number for 2133mhz is almost certainly an error. The rest of the deltas are in line with what you'd expect, approximately 80% of the performance if ram was completely the limiting component, with the deltas showing about ~6%.

It is not, and they specifically say so in that article.

That Ryse figure is no error - performance falls through the floor when running with lower levels of bandwidth, while faster RAM offers 11 per cent more performance on GTA 5 and Far Cry 4. And again, those figures are averages spread out across the benchmark run - it's noticeably higher at any given point during 'in the moment' gameplay.

 

[Techhog]

some my own tests


6700k 4.5ghz
16Gb DDR4 1600 9-10-10 1T/3000Mhz CL 14-15-15 1T
GTx970 1500/8000
win10

War thunder 1.53 CPU test.All max 1920x1080 no AA
1600Mhz

3000Mhz

Watchdogs 1024x768/1680x1050 all max temporal AA
1024x768 1600mhz ram
Avg: 91.961 - Min: 69 - Max: 143
1680x1050 1600Mhz ram
Avg: 85.571 - Min: 67 - Max: 101

1024x768 3000mhz ram
Avg: 111.087 - Min: 85 - Max: 157
1680x1050 3000Mhz ram
Avg: 92.247 - Min: 81 - Max: 104

Crysis3 1024x768/1680x1050 all max SMAA

1024x768 1600Mhz ram
http://abload.de/img/crysis3_2015_11_29_21vkucz.png
1680x1050 1600Mhz ram
http://abload.de/img/crysis3_2015_11_29_21fvuxf.png

1024x768 3000mhz ram
http://abload.de/img/crysis3_2015_11_29_20mvuwn.png
1680x1050 3000Mhz ram
http://abload.de/img/crysis3_2015_11_29_20pauyh.png

Fallout4 1920x1080 all max FOV 90
1600 mhz ram
http://abload.de/img/fallout4_2015_11_29_2ttuun.png
3000mhz ram
http://abload.de/img/fallout4_2015_11_29_2kyu0e.png

Skyrim
1600mhz ram
http://abload.de/img/tesv_2015_11_29_21_30miur6.png
3000mhz ram
http://abload.de/img/tesv_2015_11_29_21_19ucuiu.png

Witcher3 1024x768/1680x1050 allmax DOF and hairworks off
1024x768 1600Mhz ram
Avg: 96.428 - Min: 82 - Max: 111
1680x1050 1600mhz ram
Avg: 75.090 - Min: 63 - Max: 90

1024x768 3000mhz ram
Avg: 109.293 - Min: 90 - Max: 129
1680x1050 3000mhz ram
Avg: 76.151 - Min: 63 - Max: 89

scaling:
Fallout4- 26.5%
watchdogs- 20.7%
War thunder- 15.7%
witcher3- 13.2%
skyrim- 7%
crysis3 5%

Why would you even run DDR4 at 1600MHz? Do they even sell it at that speed?

 

[Madpacket]

I would rerun the tests at the slowest DDR4 speeds (2100) for comparison. Good info though.

 

[Madpacket]

Even with these results it's logical to pick up faster RAM provided the cost difference is small. With the consoles having higher effective memory bandwidth and with primary development being on the PS4/Xbox One this trend is only going to become more common.

 

[dark zero]

Literally anything can be a bottleneck for Ryse. That game is weird.

Ryse bottlenecks itself :V
Even with next gen hardware it will get bottlenecked.

 

[RussianSensation]

Fallout4 1920x1080 all max FOV 90[/b]
1600 mhz ram
http://abload.de/img/fallout4_2015_11_29_2ttuun.png
3000mhz ram
http://abload.de/img/fallout4_2015_11_29_2kyu0e.png

This actually just highlights why your emphasis on highlighting Skylake IPC differences vs. previous Intel architectures using FO4 is not relevant to measure true IPC delta of Intel architectures.

Here you have a situation where CPU is < 40% and GPU usage < 70% with 48 fps and another test with DDR4 3000 showing CPU usage < 40% (again!) and GPU usage < 70% (again!) but FPS goes up to 61.

All of this in a scene that has no explosions, no advanced physics effects, no complex AI, no large number of NPCs, no high resolution textures, no high polygon character/models, no next gen lighting/shadows, etc.

What's it's showing is how poorly optimized/outdated the game engine is and sensitivity to memory speed on Skylake vs. other systems isn't logical either since HW-E destroys Skylake in overall memory bandwidth.

Lower latency at DDR4 3000 is also not a good explanation for why Skylake beats HW-E in FO4 or why i7 6700K is so much faster than i7 4790K with DDR4 2400
http://www.hardware.fr/articles/940-5/cpu-ddr4-vs-ddr3-pratique.html

Even with these results it's logical to pick up faster RAM provided the cost difference is small.

True. For someone building a new i7 4790K rig with DDR3, it costs just $71 US for DDR3 2400.

DDR4 3000 is just $90
DDR4 3200 is just $105

 

[RussianSensation]

It seems Fallout 4 is not as much of an outlier as we originally believed.

Yes, it is. Besides your Rise example with an i3, no game in the world I've ever seen shows the same discrepancies for memory speed vs. CPU performance as FO4 does. There are 8000+ games on Steam. 2 games, of which one is only affected on a weak i3, is the very definition of an outlier.

Also, it's pretty clear from the area Digital Foundry chose to benchmark Ryse Son of Rome, it's not even a CPU demanding section because an i3 6100 gets 58.5 fps. It could easily be the case of poor engine optimization in that section of the title.

What is the point of discussing how DDR4 2133 bottlenecks an i3 6100 to 58.5 fps in Ryse when throughout the most demanding sections of the game, an i3 6100 even paired with DDR4 3200 would get destroyed?

Fact is no matter what DDR4 memory you pair with an i3 6100 in Ryse, it'll continue to be a bottleneck for Ryse in CPU demanding sections. For this game you need an i5/i7/FX 8-core to have any hope of maintaining 60 fps.

For those who have been following i5 bottlenecking in CryEngine games, this wouldn't be surprising. CryEngine games LOVE 6-8 threaded CPUs + IPC. If you do not have an i7 in CryEngine games, even if you use DDR4 2800, it's not going to help the i5 overcome the lack of sufficient threading capability:

390x @ 1140/6200
i7-6700K@4.6
i5-6600K@4.6
https://www.youtube.com/watch?v=AbsXPjVcc6w