BF3 CPU bottleneck? Really?

[moonbogg]

Great replies...ok..i just did an Oman run starting at construction site and made my way over to hotel and got on roof and overlooked the whole map. FPS on the roof is high 50's to 60's. I died a couple times so FPS drops a little there. There were 60-64 players and lots of action, so it was a good stress test I think. On the roof, GPU usage went down to 50% each GPU.

Settings: 1080P All Ultra 4XMSAA no antialiasing POST (looks blurry and stupid)
HT on @ 4.3ghz. I might try different clock speeds to experiment with MIN fps.

Results:
min:59
max:166
average:96

EDIT: Just OCed to 4.6 and did a similar run with these results
min:66
max:179
average:103

So, you think the MIN is a result of CPU bottlekeck? Or am I just expecting too much? I am bothered by the fact that GPU usage is so low and only at those times is FPS low.

Thanks.

 

[2is]

Added cache and the 100mhz clock speed advantage is why it's faster in BF3

That graph really reaffirms the point that it's not the CPU that's causing the problems here. When the FX4100 performs as well as a 2600K it shows just how GPU heavy that game is even at 1680x1050 resolution.

Looks like it's this

People really need to stop posting SP graphs to show CPU usage in BF3 when it's MP that's being discussed. You might as well show us a graph of COD4 or whatever other game you want because it will serve the same purpose, which is none at all.

And yes HT does help in BF3 MP

 

[Grooveriding]

Gents, if you've played BF3 multiplayer, you should know how this works. It's heavily multi-threaded and it's heavily dependent on CPU. Throwing up singleplayer benchmarks from BF3 is about as relevant as showing me a graph from Metro 2033.

Here's the proof:

Back to the OP's question, while disabling HT may help you figure out the CPU load, it surely won't help increase your minimums. Can you give us a FRAPS run from Oman so we can get an idea for min and averages to tell if there's actually something wrong?

This is the person telling you the accurate situation OP. Battlefield 3 MP is heavily CPU dependent, and as he as mentioned, particularly so on the B2K maps. This will likely carry through to all the new maps coming as well.

HT is a big benefit in this game for MP. Overclock your CPU more if possible.

This is probably the only game I can think of off the top of my head where a SB-E is better than a SB. There is a thread on guru or overclock I think where people were doing comparisons of BF3 MP, HT ON vs OFF. HT ON was a huge boon on 64 player maps.

 

[Termie]

Great replies...ok..i just did an Oman run starting at construction site and made my way over to hotel and got on roof and overlooked the whole map. FPS on the roof is high 50's to 60's. I died a couple times so FPS drops a little there. There were 60-64 players and lots of action, so it was a good stress test I think. On the roof, GPU usage went down to 50% each GPU.

Settings: 1080P All Ultra 4XMSAA no antialiasing POST (looks blurry and stupid)
HT on @ 4.3ghz. I might try different clock speeds to experiment with MIN fps.

Results:
min:59
max:166
average:96

EDIT: Just OCed to 4.6 and did a similar run with these results
min:66
max:179
average:103

So, you think the MIN is a result of CPU bottlekeck? Or am I just expecting too much? I am bothered by the fact that GPU usage is so low and only at those times is FPS low.

Thanks.

Awesome results, and just what I expected. You are heavily CPU-limited - not only does your minimum go up by 10%, but your average went up by 7%. That's with a 7% CPU overclock.

These are the kind of results that really help us all understand what's going on. Thanks for putting the time in (and having the high-powered rig to do it with). I do think that for BF3 multiplayer, having much more power than a single 680/7970 is not going to do much at 1080p. By the way, I average (edit:) 60-65fps at the same settings on a single 670 - the question is whether your advantage is entirely due to the CPU, which is about 40% more powerful at 4.6 than my CPU at 3.5 (edit: my guess is not except maybe for mins). If you really wanted to make your head spin, you'd try the exact same test with SLI disabled. You might not like the results, though!

This is the person telling you the accurate situation OP. Battlefield 3 MP is heavily CPU dependent, and as he as mentioned, particularly so on the B2K maps. This will likely carry through to all the new maps coming as well.

HT is a big benefit in this game for MP. Overclock your CPU more if possible.

This is probably the only game I can think of off the top of my head where a SB-E is better than a SB. There is a thread on guru or overclock I think where people were doing comparisons of BF3 MP, HT ON vs OFF. HT ON was a huge boon on 64 player maps.

Thanks for backing me up, Groove! Looks like my hunch was proven above by the OP.

 

[AtenRa]

I have shown you this before. BF3 MP in large open maps 64 player servers is highly CPU bound. It can use 6-8 threads easily.

 

[moonbogg]

I have done testing with SLI disabled, and I have no benchmarks yet, but I will take a couple benchmark runs later today when I get home. I can tell you with certainty (and not just because I want it to be true) that SLI makes a very large difference, especially when the shit hits the fan and during regular gameplay as well, but perhaps not such a big difference during these heavily CPU bound situations. The data will tell the real story. I remember starting out on Oman with a single 670 active and getting 45fps while running toward the buildings from one of the beach spawn points. I have a 120hz monitor so the extra FPS makes the game much smoother and crisp pretty much all the time. I will provide real data at 4.6ghz with SLI off later in a similar run to show this.
If a 6 core would actually help, then I may upgrade platforms when haswell hits, or maybe even ivy-E. I think Sandy-E has been out too long to want to jump on it especially if my dreadful FPS dips are only into the "terrible" 60fps range, lol. These mins only take place on the most CPU intense locations on the most demanding maps. I mean, I have to be standing on the Hotel roof looking out, or the TV station on Sharki. All other times FPS is higher. So I am no so much really complaining, but trying to diagnose the system to understand the bottleneck and what the next upgrade path should be. Sounds like a 6-8 core ivy-E rig.
I would be thrilled to have people do similar benchmark runs with a 6 core Sandy-E or maybe a nehalem for comparison. I can't understand why review sites don't even get their feet wet with this kind of real testing. This is the testing that gamers care about. It is the ONLY testing that is relevant to online gamers. Inconsistencies be damned, give us the data anyway.

 

[SlowSpyder]

A 4.3Ghz 2600K a bottleneck? Not even close. Up your overclock to 4.6Ghz+

But if 4.3GHz isn't a bottleneck, why overclock the chip more?

 

[Termie]

I have shown you this before. BF3 MP in large open maps 64 player servers is highly CPU bound. It can use 6-8 threads easily.

Amazing data, thanks for posting. I'd say that the OC'd 920 actually has the most consistent results in your graph, and the fewest drops.

I have done testing with SLI disabled, and I have no benchmarks yet, but I will take a couple benchmark runs later today when I get home. I can tell you with certainty (and not just because I want it to be true) that SLI makes a very large difference, especially when the shit hits the fan and during regular gameplay as well, but perhaps not such a big difference during these heavily CPU bound situations. The data will tell the real story. I remember starting out on Oman with a single 670 active and getting 45fps while running toward the buildings from one of the beach spawn points. I have a 120hz monitor so the extra FPS makes the game much smoother and crisp pretty much all the time. I will provide real data at 4.6ghz with SLI off later in a similar run to show this.
If a 6 core would actually help, then I may upgrade platforms when haswell hits, or maybe even ivy-E. I think Sandy-E has been out too long to want to jump on it especially if my dreadful FPS dips are only into the "terrible" 60fps range, lol. These mins only take place on the most CPU intense locations on the most demanding maps. I mean, I have to be standing on the Hotel roof looking out, or the TV station on Sharki. All other times FPS is higher. So I am no so much really complaining, but trying to diagnose the system to understand the bottleneck and what the next upgrade path should be. Sounds like a 6-8 core ivy-E rig.
I would be thrilled to have people do similar benchmark runs with a 6 core Sandy-E or maybe a nehalem for comparison. I can't understand why review sites don't even get their feet wet with this kind of real testing. This is the testing that gamers care about. It is the ONLY testing that is relevant to online gamers. Inconsistencies be damned, give us the data anyway.

Exactly - unfortunately, it's up to us real-world users to test various configurations, because review sites testing GPUs will just use one platform (like a 2600K), and review sites testing CPUs will just use one GPU (like a 7970). I really like Techspot's game reviews, because they test multiple CPUs in various games, whereas most game reviews only look at GPUs. Unfortunately, they, like every other review site besides HardOCP, has never benchmarked BF3 multiplayer. By the way, a while back HardOCP admitted that their i7-920@3.6 had become a bottleneck in their tri-SLI testing, after repeatedly saying it didn't matter: http://hardocp.com/article/2011/05/03/nvidia_3way_sli_amd_trifire_redux/. Now if only they'd actually do some comparative CPU benchmarking in BF3 multiplayer.

Something tells me that Intel would be none too pleased, especially if results showed the anemic bulldozer killing a stock 2500k, as the graph above illustrates. This is the reason I'll never "upgrade" from my i7-860 to an i5. I'm going to wait for a compelling all-around upgrade, and I hope that's Haswell. By the way, I don't think it will be 6-core, as you suggest above.

I'm looking forward to your single-GPU benchmarks, when you have the chance. I don't doubt that SLI provides a signficantly higher average, but when the s*** hits the fan, my guess is that the CPU will be limiting, and so playability may not be that much greater.

BTW, is there an advantage to a 120Hz monitor if you're never above 120fps, i.e., is it still smoother than 60Hz?

 

[Don Karnage]

But if 4.3GHz isn't a bottleneck, why overclock the chip more?

Cause 4.3ghz is a weak overclock

 

[saratoga172]

Check your drivers. Also did you do a clean install? When I did a clean install of my os, drivers, games etc I noticed much smoother gameplay and higher frames.

I've got sli'd 670's and a 2500k at 4.5ghz. Playing on a Dell u3011 at 2560x1600. My fps average is low 80's to high 70's. And I rarely drop below the 60's on mp. Everything maxed except I think 2xmsaa instead of 4. Can't tell a difference personally.

which monitor are you going on and how are you liking the 120hz rate. Thinking about going with one for gaming.

Sent from my Galaxy Nexus using Tapatalk 2

 

[moonbogg]

I understand the skepticism regarding the benefits of SLI at 1080p. Wait for the benchmark and you'll see. The benefits are significant. Regarding the 120hz monitor, it is the Asus VG278 and yes, it makes a huge ass difference at anything above 60fps, especially in the 80+ fps range. Motion is much more clear, especially panning left to right, the background remains crisp and clear, whereass 60hz the background blurrs. You wouldn't notice it unless you experienced a 120hz monitor with high framerate. It actually reminds me of the good old CRT days when motion was crisp and clear. This was my biggest regret about making the move to an LCD monitor and with 120hz it has been nearly eliminated so long as FPS stays up there. There are skeptics about this, but the difference is very real.
About SLI and the shit hitting the fan, the cards are just CPU limited enough so that when a ton of explosions and especially smoke cover the screen, the CPU still has the power to push the cards to 95-99% usage during those intense, shit hitting moments and as a result the FPS stays basically where ever it was before the intensity started. The rig still has enough CPU power to let the cards throttle UP when the scene gets intense, where a single card will be the bottleneck and FPS will fall to the 40's and sometimes 30's depending on intensity. This is based on my recent observations and memory of what FRAPS was reporting during single card use, but I will provide a real benchmark to compare the two. I am compelled to ask you to trust me for now that SLI makes a big difference in this game even at 1080p.
Oh, I thought i'd mention, despite the bad 3D drivers for the 600 series at the moment, I played BF3 in 3D to test it. GPU usage was pegged at 95-99% the whole time and framerate was astonishing and usually at 60 or above. Thats amazing for 3D. Too bad the drivers broke down and the GPU usage went down to 50% per card and framerate tanked. Usage even went as low as 30% per card at times. This is certainly a driver issue as my CPU has plenty of power to feed both cards for 3D use as your framerates in 3D are cut in half, especially since it worked correctly for a short time.
EDIT: I am not exactly sure how CPU power and scaling effects 3D performance, so I will do benchmarks with 3D as well in single card and SLI and post those numbers.

 

[Termie]

I understand the skepticism regarding the benefits of SLI at 1080p. Wait for the benchmark and you'll see. The benefits are significant. Regarding the 120hz monitor, it is the Asus VG278 and yes, it makes a huge ass difference at anything above 60fps, especially in the 80+ fps range. Motion is much more clear, especially panning left to right, the background remains crisp and clear, whereass 60hz the background blurrs. You wouldn't notice it unless you experienced a 120hz monitor with high framerate. It actually reminds me of the good old CRT days when motion was crisp and clear. This was my biggest regret about making the move to an LCD monitor and with 120hz it has been nearly eliminated so long as FPS stays up there. There are skeptics about this, but the difference is very real.
About SLI and the shit hitting the fan, the cards are just CPU limited enough so that when a ton of explosions and especially smoke cover the screen, the CPU still has the power to push the cards to 95-99% usage during those intense, shit hitting moments and as a result the FPS stays basically where ever it was before the intensity started. The rig still has enough CPU power to let the cards throttle UP when the scene gets intense, where a single card will be the bottleneck and FPS will fall to the 40's and sometimes 30's depending on intensity. This is based on my recent observations and memory of what FRAPS was reporting during single card use, but I will provide a real benchmark to compare the two. I am compelled to ask you to trust me for now that SLI makes a big difference in this game even at 1080p.
Oh, I thought i'd mention, despite the bad 3D drivers for the 600 series at the moment, I played BF3 in 3D to test it. GPU usage was pegged at 95-99% the whole time and framerate was astonishing and usually at 60 or above. Thats amazing for 3D. Too bad the drivers broke down and the GPU usage went down to 50% per card and framerate tanked. Usage even went as low as 30% per card at times. This is certainly a driver issue as my CPU has plenty of power to feed both cards for 3D use as your framerates in 3D are cut in half, especially since it worked correctly for a short time.

I don't doubt that SLI helps, but as your own testing has already shown, you need a lot of CPU to back it up. I briefly had a second 670 to benchmark with, and in my limited BF3 singleplayer testing, I was getting 40% scaling at 1920x1200. My CPU was not enough for it, even in singleplayer, but I never intended to go SLI.

By the way, I totally buy the idea that 120Hz helps a lot, I just thought that you had to be at 120fps to get a benefit - thanks for clearing that up. I just wish the 120Hz screens came in bigger sizes. I'd be pretty interested in trying out 3D too...

 

[Don Karnage]

I don't doubt that SLI helps, but as your own testing has already shown, you need a lot of CPU to back it up. I briefly had a second 670 to benchmark with, and in my limited BF3 singleplayer testing, I was getting 40% scaling at 1920x1200. My CPU was not enough for it, even in singleplayer, but I never intended to go SLI.

By the way, I totally buy the idea that 120Hz helps a lot, I just thought that you had to be at 120fps to get a benefit - thanks for clearing that up. I just wish the 120Hz screens came in bigger sizes. I'd be pretty interested in trying out 3D too...

Did you test at 3.6ghz? Why not ramp that 860 up to 4.0-4.2?

 

[Termie]

Did you test at 3.6ghz? Why not ramp that 860 up to 4.0-4.2?

Nah, just at 3.5. I had to return the 670 via RMA, so I didn't want to risk destroying anything, and it wasn't a stable card anyway.

That being said, I might eventually do some CPU scaling tests with a higher OC and a single card. Tests I did a while back on BF3 showed a pretty nice jump in minimums just going from stock to 3.4 (using 5850 crossfire, which is close to a 670), but at 3.5, I might have enough CPU to keep the 670 happy.

 

[Grooveriding]

I have done testing with SLI disabled, and I have no benchmarks yet, but I will take a couple benchmark runs later today when I get home. I can tell you with certainty (and not just because I want it to be true) that SLI makes a very large difference, especially when the shit hits the fan and during regular gameplay as well, but perhaps not such a big difference during these heavily CPU bound situations. The data will tell the real story. I remember starting out on Oman with a single 670 active and getting 45fps while running toward the buildings from one of the beach spawn points. I have a 120hz monitor so the extra FPS makes the game much smoother and crisp pretty much all the time. I will provide real data at 4.6ghz with SLI off later in a similar run to show this.
If a 6 core would actually help, then I may upgrade platforms when haswell hits, or maybe even ivy-E. I think Sandy-E has been out too long to want to jump on it especially if my dreadful FPS dips are only into the "terrible" 60fps range, lol. These mins only take place on the most CPU intense locations on the most demanding maps. I mean, I have to be standing on the Hotel roof looking out, or the TV station on Sharki. All other times FPS is higher. So I am no so much really complaining, but trying to diagnose the system to understand the bottleneck and what the next upgrade path should be. Sounds like a 6-8 core ivy-E rig.
I would be thrilled to have people do similar benchmark runs with a 6 core Sandy-E or maybe a nehalem for comparison. I can't understand why review sites don't even get their feet wet with this kind of real testing. This is the testing that gamers care about. It is the ONLY testing that is relevant to online gamers. Inconsistencies be damned, give us the data anyway.

I can try to do this on the weekend. I would be interested as well. I'll test HT on vs Off and maybe 6 cores vs 4 cores HT OFF.

I used to be able to test 64 player easily as I had my own server that was busy but have since canceled it It is tough to consistently recreate the same benchmark.

I found the best and most consistent was at the beginning of a full 64 player match of Caspian; driving a tank in a circuit around the map and benching that. Will be tricky to do it 3 or 4 times with map rotations and needing a full server though. :awe:

 

[moonbogg]

I can try to do this on the weekend. I would be interested as well. I'll test HT on vs Off and maybe 6 cores vs 4 cores HT OFF.

I used to be able to test 64 player easily as I had my own server that was busy but have since canceled it It is tough to consistently recreate the same benchmark.

I found the best and most consistent was at the beginning of a full 64 player match of Caspian; driving a tank in a circuit around the map and benching that. Will be tricky to do it 3 or 4 times with map rotations and needing a full server though. :awe:

You are the perfect guy to test this whole thing out. Please do it at 1080p for the CPU limiting factor. I will do more tests as I have said earlier, but you can really show the benefit of the extra cores and your cards are a little stronger, so the CPU limit might show a little more with 4 cores as well.
As far as being concerned about consistent benchmarks go, I am actually not overly concerned with that. Simply finding a decently full server on the most intense maps, Gulf of Oman is what I am focusing on, will provide good results. The minimums are always very similar on the same map in the same areas so long as the server is pretty full. I say we focus on Oman and do similar runs around the construction site and looking out over the map from the Hotel. That seems to be the most CPU intense.

EDIT:
If the difference that 2 extra cores provides is so slight that it requires a perfectly repeatable benchmark to reflect the difference, then we may conclude that it is not an issue and 4 cores is the best we could do for now. In that case it would be IPC that is limiting us and noone can help that. But testing will tell. I hope 6 cores does significantly better so I can go rush out and blow some more money!

 

[Rvenger]

Also, Gulf of Omen running from the Hotel to the construction site. I always dip in the mid 50 fps there.

 

[JAG87]

My GOD, what a bunch of monkey advice.

1) Turn the resolution down to 1280x720 and MSAA off.

2) Check fps.

3) If it has gone up, your bottleneck is NOT the CPU. If it stayed the same, your bottleneck is the CPU.

 

[moonbogg]

My GOD, what a bunch of monkey advice.

1) Turn the resolution down to 1280x720 and MSAA off.

2) Check fps.

3) If it has gone up, your bottleneck is NOT the CPU. If it stayed the same, your bottleneck is the CPU.

You sure? I set details to low and they went in one spot from 55 to over 90. So based on that I was hoping that my CPU was not the bottleneck. I am going through a process of elimination though, ok? No need to refer to us as monkeys when we are infact apes. Perhaps increasing details require more CPU power to provide the cards with additional data? So if thats the case, setting details to LOW might not be the best way to investigate a CPU bottleneck with settings on Ultra.

EDIT: heres an example: 2 years ago with BFBC2 I had a dual core CPU and 2 cards. The cards worked at 50% each and my CPU was pegged. FPS increased when I put details on low. Why would it do that? According to you, the FPS should have stayed the same. The reason I suspect: The CPU could not provide the additional info to the cards for them to keep up on high settings. Moving to a quad core fixed it promptly.

 

[Termie]

You sure? I set details to low and they went in one spot from 55 to over 90. So based on that I was hoping that my CPU was not the bottleneck. I am going through a process of elimination though, ok? No need to refer to us as monkeys when we are infact apes. Perhaps increasing details require more CPU power to provide the cards with additional data? So if thats the case, setting details to LOW might not be the best way to investigate a CPU bottleneck with settings on Ultra.

I was going to leave this alone, because his advice really wasn't that friendly, but here's the problem with how you interpreted it:

- he said reduce resolution and MSAA, which are solely dependent on GPU
- you interpreted it as reducing details to low, which is almost all cases (except obvious ones like HBAO) have a very significant CPU component

If you really wanted to try his advice (and I'm not sure it applies with BF3), run ultra/noMSAA at 720p. If it runs faster, then you were GPU limited. But you already know based on your testing that you're CPU limited at the settings you want to play at, so I'm not sure there's a point to this exercise.

 

[moonbogg]

I was going to leave this alone, because his advice really wasn't that friendly, but here's the problem with how you interpreted it:

- he said reduce resolution and MSAA, which are solely dependent on GPU
- you interpreted it as reducing details to low, which is almost all cases (except obvious ones like HBAO) have a very significant CPU component

If you really wanted to try his advice (and I'm not sure it applies with BF3), run ultra/noMSAA at 720p. If it runs faster, then you were GPU limited. But you already know based on your testing that you're CPU limited at the settings you want to play at, so I'm not sure there's a point to this exercise.

Thanks for clearing that up, and yes, I misunderstood how the CPU loading works regarding details vs resolution. It is more than worthwhile to try the experiment and I will try it ASAP and report the numbers. Many people simply claim that BF3 is poorly optimized and that SLI scaling is poor. I want to eliminate guesswork and get to the bottom of it. I am not interested in being right. I am interested in solving the problem.

 

[sm625]

Nah, just at 3.5. I had to return the 670 via RMA, so I didn't want to risk destroying anything, and it wasn't a stable card anyway.

That being said, I might eventually do some CPU scaling tests with a higher OC and a single card. Tests I did a while back on BF3 showed a pretty nice jump in minimums just going from stock to 3.4 (using 5850 crossfire, which is close to a 670), but at 3.5, I might have enough CPU to keep the 670 happy.

But why would there be such steep dropoffs when the gpu is overclocked? That data just doesnt look right to me.

 

[Termie]

But why would there be such steep dropoffs when the gpu is overclocked? That data just doesnt look right to me.

And that's the challenge of benchmarking BF3 multiplayer. No run is exactly the same. That's why you have to do multiple runs and look for patterns.

There is clearly a pattern in my data reflecting higher minimums with a higher CPU clock, and what appears to be a CPU bottleneck on averages as well. That was the point of my testing. I agree that the averages when the GPUs were highly overclocked look way off, but those are the data points I got.

 

[BD231]

If you run 3 monitors with dual cards you get 80-90% gpu useage in multi-player on BF3 in SLi, try it for yourself, it's not the CPU.

 

[moonbogg]

If you run 3 monitors with dual cards you get 80-90% gpu useage in multi-player on BF3 in SLi, try it for yourself, it's not the CPU.

I am sure you are right. For me though, I prefer to use a single 120hz screen and haul as much ass as possible on that screen. With multiple monitors or any higher res than 1080p, I wouldn't feel comfortable with 2gb cards. I would definitely go the 4gb route as 2gb is already pushed and exceeded in some cases at 2560x1600.

Looking forward to getting home to crank out some benchmarks. I just might grab a sick water cooler for that damn CPU and smack its ass till its happy at 5.0

EDIT: Just did a little testing. I found the most CPU intensive and the most intense part of the entire game I am aware of, sharki on the TV station roof with 64 players. Just standing and overlooking the map.

SLI: 55-60
Single Card: 35-40

SLI: Resolution at 720p, no AA: 60FPS

So I see there is certainly a CPU bottleneck, but even with it the FPS is significantly higher with SLI enabled. CPU usage went up with SLI enabled.

CPU usage with SLI and HT on: 65%
CPU usage with single card and HT on: 50%

FPS while running around sharki with SLI: 80-100
Same thing with single card: 45-60