Is still Intel better than AMD for Single-Thread?

[chiliwili69]

As it is known, IL-2 VR is very dependent on the CPU single-thread performance. We have run multiple tests and we saw that the performance (both in Monitor and VR) was correlated with a public benchmark which everybody can run in their PC (Passmark Performance Test). All those hundred of tests are collected and they produce an average value (including all sort of CPU speeds, overclocked and non-overclocked) which they public here:

https://www.cpubenchmark.net/singleThread.html

 

So they are the top performers in single-thread which is what it matters for IL-2 VR. But those values has to be seen with caution since they are just a bulk average.

The interesting thing would be to see the top performers of each CPU model, not the average.

 

To know that you can have the paid version (or use the trial in the free trial period) and you can visualize all the test (baselines) performed for every CPU model.

For example the 8086K (world max is 3318)

 

or the 9900K (world max is 3270):

 

or Ryzen 3700X (world record 3119)

 

or Ryzen 3900X (world record is 3197)

 

So we can see that top performers (the ones on the right of the distribution) are very similar. Perhaps the 8086K is the one which has more in the very high range.

 

To know the frequencies of those baselines you can go to Manage baselines section and search for specific CPU models and order them in CPU freq.

 

In the section "Manage Baselines" you can go to advanced search an select a model (8086K for example) and request to show only baselines with more than 5.0GHz, so you can have an idea of how far each CPU can go for Overclocking:

[chiliwili69]

A colleague has the chance to run the STMark on a Ryzen 7 3800X at 4.4GHz. The result was 3130!!  And it can be overclocked a bit more!

 

Remember that Intel CPUs (8700K, 9700K, 9900K) at 5.1GHz delivers around 3090 and at 5.2GHz deliver 3190.

So, I think this new Ryzen line is going to be quite attractive in single-core.

[icecream]

Yes they are, currently.

 

However, remember that Intel is still on 14nm too..

[ZachariasX]

1 hour ago, icecream said:

Yes they are, currently.

 

However, remember that Intel is still on 14nm too..

It will be for about two years to come. 10 nm chips that are currently shipping clock one GHz less than current 14 nm and it consumes more power. With Granite Rapids, slated for 2022, Intel seems to get competitive again, supposedly an 82 core monster for the future Xeon lineup. My guess is that future Intel consumer parts will profit equally as 16 cores are entering mainstream.

 

It should be seen as Zen4 analog, the coming successor of the latest Ryzen/Epyc lineup.

 

I don‘t really think the „top performer“ really matters that much, as that one is not for sale, but for the consumer it is a lottery among a certain binning. The 8086 and 9900 really don’t differ much in that as they basically have same specs and come from the same process.

As for now, there really seems to be a draw between Intel and AMD in single core, Intel doing 3000 on average and AMD doing 2950, identical for practical purposes.

 

What I find interesting however, how AMD clusters results around a single mean, while Intel shows 2 to 4 peaks, probably indicating different platforms around that CPU. Going Intel, one shouldn‘t be cheap regarding mobo and RAM.

 

 

 

[chiliwili69]

17 hours ago, ZachariasX said:

while Intel shows 2 to 4 peaks

 

I think those 2 peaks on the 8086 are for OC of 5.1GHz and 5.2GHz. With some distribution around depending on RAM.

 

I have counted the number of Passmark tests submitted by users for different OC frequencies and intel model:

 

 

We can see that 8086K and 9900K have similar OC ability, giving the number of samples test for each model (4000 for 9900K and 800 for 8086K)

[Sgt_Joch]

Ryzen 3700x seems to be the best bang for the buck in AMD.  The 3800x/3900x are a lot more expensive for only a modest increase in performance.

[Voyager]

@Sgt_Joch Always check prices on them though. Right now I'm seeing 3800X's go for $40 more than the 3700X, so it's looking like the 3800X has fallen in price while the 3700X is rising with the demand. And that may have changed again by the time people are reading this. 

 

Another aspect to consider is that the AM4 will be compatible with Zen 3, with has potential to improve both IPC and clock speed. 

 

Addendum: @chiliwili69 Somehow the first time I read this, I missed that the I9-9900K values were for chips running at 5.1-5.2GHz. Finding that out suddenly greatly simplifies my CPU decision making. For some reason I'd gotten it into my head that these had the CPUs locked at the base clock speed. Given that a Zen 2 at 4.4 gets about the same single thread score as an I9-9900K at 5.15Ghz and that it's looking like AMD is going to be able to push anywhere between another 10-15% single threaded performance next year with Zen 3, to me it makes the most sense to get either a 3700X or 3800X (depending on price) and the backbone that I should be able to support a Zen 3 chip for later, and plan on dropping in one of their 16 core versions when the AM4 starts to sunset. I'm thinking that should keep my system backbone in upper end of the performance curve for quite some time.

 

And by that point, we may have an engine rebuild that fully supports multi-threading.

 

Thank you,

 

Harry Voyager

Edited October 23, 2019 by Voyager

[Voyager]

@chiliwili69 So got the rebuild to the point where I can run Passmark, and right now I'm getting between 2930 to 3070 Single Thread performance on my R7 3800X, depending on which core it's using.

 

Ryzen Master is reporting the golden core running at as much as 4450 mhz, and the second core at 4400mhz, so I suspect the second core is down-cyling more often. 

 

Note, this is at stock settings, no over clocking, no Precision Boost Overdrive. The only things I've added are the ram is DDR3600 16-19-19-39, the Bios is 1.0.0.4, and I've got a 360mm water cooler attached to it, so temps are at around 56C under load. 

 

I'm going to run at these settings until I've had a chance to get the rest of the computer set up, and am able to run some testing in IL-2. The brief bits I've gotten into it with the Reverb have only been hanger shots, but it looks really pretty. My guess is that I'm going to end up being GPU limited at that resolution, so OCing the CPU may not gain me anything yet, and might ding the resale value of the chip down the road. (The long term plan is to replace it with a Ryzen 4000 in a year or two, and resell the R7 3800X. Or build someone else a system with it, so if I don't need to OC it, I might as well keep it cool and comfortable in the meantime.) 

[Voyager]

So not yet able to get frame rates, but I can report that with a Reverb and a Geforce GTX 1080 Ti Fe I am now GPU limited, rather than CPU limited.

 

I'll have to see about capturing with Fraps next time to see if I can get some numbers.

[chiliwili69]

16 hours ago, Voyager said:

My guess is that I'm going to end up being GPU limited at that resolution, so OCing the CPU may not gain me anything yet, and might ding the resale value of the chip down the road

 

That´s not exactly true, the better framerates achieved by CPU the more room you left to the GPU to do its job. CPU has to build scene and then GPU has to render it.

So, CPU OC helps GPU performance.

 

56 Deg C is a very good temperature with a stress load. You can try to disable 4 of your 8 cores (assuming you only want the PC for IL-2 VR) and achieve even better overclock.

 

I think it is worth to invest some time in determining the max stable and safe overclock your machine can achieve. Once this is done you don´t need to invest any time more and enjoy the chip for his life.

 

Regarding life of CPUs, I don´t think that a stable and safe OC shorten significantly the useful life of a CPU. Maybe extreme OC will reduce life from 20 to 10 years, but in 10 years all CPUs become obsolete.

[SCG_Fenris_Wolf]

Lab tests like Passmark are all fine and dandy, but I'm going to wait for actual comparisons to decide on which route to go. If @Voyager tests his Ryzen with the Reverb upsampling (by putting resolution so low that his GPU is not a limiting factor), it will be interesting to see whether the Ryzen can hold himself in the field.

[Voyager]

@SCG_Fenris_WolfAw, do I have to downsample? It's so pretty 

 

In all seriousness I'll probably end up running it a couple of ways, first at my current settings then at a Rift CV1 comparable resolution.

 

@chiliwili69For the test itself, since patches broke the benchmark track, I'm thinking of using a quick mission, Rhienland Summer, Heavy cloud cover at 1000m and maximum turbulence, and each side with 4x Fighters and 4x Bombers with the highest bomb count bomb loads, and using a 10 mount FRAPS benchmark average with Max and 1% minimum numbers. 

 

I'm thinking that should produce the highest calculation load that's easily achievable, and should stress the CPU, and I'm hoping that running 10 minute runs should even out some of the variability. I'll probably have to do several runs to get a really consistent number. 

 

What are your thoughts? Any recommendations? 

 

Thank you, 

 

Harry Voyager

 

Addendum: I'm not really worried about an OC hurting the CPU, just when I go to resell it, that might make buyers bid less than they would otherwise. Though, on the flip side, having a known, tested overclock, if it is good, could also be an incentive for buyers. Different sides of the same coin. I think I'll finish the testing first and see what we find. 

[Archie]

Be very interested in the results Voyager, I'm overdue an upgrade and am seriously looking at AMD for the first time in years.

[SCG_Fenris_Wolf]

1 hour ago, Voyager said:

@SCG_Fenris_WolfAw, do I have to downsample? It's so pretty 

 

In all seriousness I'll probably end up running it a couple of ways, first at my current settings then at a Rift CV1 comparable resolution.

Spoiler

@chiliwili69For the test itself, since patches broke the benchmark track, I'm thinking of using a quick mission, Rhienland Summer, Heavy cloud cover at 1000m and maximum turbulence, and each side with 4x Fighters and 4x Bombers with the highest bomb count bomb loads, and using a 10 mount FRAPS benchmark average with Max and 1% minimum numbers. 

 

I'm thinking that should produce the highest calculation load that's easily achievable, and should stress the CPU, and I'm hoping that running 10 minute runs should even out some of the variability. I'll probably have to do several runs to get a really consistent number. 

 

What are your thoughts? Any recommendations? 

 

Thank you, 

 

Harry Voyager

 

Addendum: I'm not really worried about an OC hurting the CPU, just when I go to resell it, that might make buyers bid less than they would otherwise. Though, on the flip side, having a known, tested overclock, if it is good, could also be an incentive for buyers. Different sides of the same coin. I think I'll finish the testing first and see what we find. 

 

Hi @Voyager, well, if you want to test the CPU only and temporarily remove the GPU from the equation - then it would be best to downsample for that test, yes

[Voyager]

Ryzen R7 3800X stock, with 360mm water cooler

570X motherboard with 1.0.0.4 Bios

64Gb DDR3600, 16-19-19-39 timings

Geforce GTX 1080 Ti Founders Edition

Installed on 2TB Intel 660p M.2 drive 

 

Basic System settings: 

 

Graphics set to High Quality

1980x1200

UI Scale: Auto

Shadows Quality: Medium

Mirrors: Simple

Distant Landscape Detail 4x

Horizon Draw Distance 150km

Landscape fileter: Off

Grass Quality: Normal

Clouds Quality: High

Dynamic Resolution: Full

Anti-aliasing: 2

Gama: 0.9

Full Screen: True

Enable VR HDM: True

SSAO: True

HDR: True

Sharpen: False

Use 4k Textures: True

Distant Buildings: True

 

Map Scenery Distance set to Unlimited

 

Test 1: HMD Reverb, Native Resolution (2228x2180) Quick Mission Map: Rheinland, Summer 1944, 

Weather: Time 12:00 am, Wind 6m/s, Turbulence 3m/s, Weather Heavy, Cloud Level 1000m

Ground Targets: All

 

Flight level 3,000m, at 5,000m

Friendly Flight: 4x Veteran P-51D-15 with Extra Ammo, Gyro gunsight, 150 octane fuel, Bendix MN-25, 4x Veteran B-25J with 8x 500lb bombs

Hostile Flight 4x Novice Fw-190A-8, No mods, 4x Novice He-111H-16 with 32 SC 50 bombs.

 

Fraps Bench Mark: 600 Seconds, starting from un-pause:

Frames, Time (ms), Min, Max, Avg
 30242,    600000,  39,  91, 50.403

 

Note: the GPU was running at 98% through the majority of this test.

 

Run #2: Same as run #1, with the Reverb set to 1080x1056 via the SteamVR Application Resolution:

Frames, Time (ms), Min, Max, Avg
 52327,    600000,  53,  92, 87.212
 

I think the Reverb is capping the frame rate at 90hz, which may be skewing the results. I've included the FPS csv files and frame time files. On both runs, I went after the bombers first, then after I believed they were down I went and hunted down the fighters. Typically the bomber engagements were above the cloud level, while the fighter duels were right below the cloud layer, so there was a lot of passing through the clouds during vertical maneuvers.

 

 

 

Il-2 2019-11-06 23-19-16-75 frametimes.txt Il-2 2019-11-06 23-46-03-00 frametimes.txt Il-2 2019-11-06 23-19-16-75 fps.txt Il-2 2019-11-06 23-46-03-00 fps.txt

Edited November 7, 2019 by Voyager
Added specs

[chiliwili69]

That´s a nice finding. Thanks for sharing. It clearly demonstrate that you CPU+GPU can not even reach 90fps with native resolution and those settings.

Please, edit your previous post to include PC specs and also graph setting (LOW, Balanced, High, Ultra), so we have complete picture.

 

It would be interesting to see the same run at native resolution but with not clouds.

Just made a graph fps of your test:

[Voyager]

@chiliwili69Done. 

 

At this point it looks like the limiter is my 3D card, however there are significant dips when I'm engaging the bombers, and you can really see that around 70 seconds in on both plots.

 

I'm thinking the drop on the second half of the native resolution run is the impact of being below and in the clouds eating my GPU for lunch. 

 

For me, I think this says next year is time for a GPU upgrade depending on how Ampier and Big Navi turn out, though if they are not a g ugh jump, or prices stay really high I can probably keep going until we see what Intel brings to the high performance table in 2021.

 

I think the other big take-away is having the extended view distance on is fully playable right now. 

 

On the game side, I think this is indicating the next key thing for the devs to multi thread is the aircraft AI. Even though I was topping out in the main thread, I don't think I was more than 30% total CPU utilization. Splitting the building render system into separate threads seems to have been a huge success, at least on my system, and I suspect if they can do the same for the AI, that may break the bottleneck on heavy bombers and big formations even without needing to seriously optimize the AI code. 

[=SFG=capt_nasties]

check this out...this is my live overclocking benchmark in VR on IL2.  I did this to "show my work" but I picked up 40% average frame rate going from an i9900k @4.7 and ram at 2133 (stock mobo settings) vs i9900k @ 5.0 and ram at 4433.....no changes to GPU (2080) I am by no means a pro you tuber but figured this would be interesting and something I wanted to test.  

 

It should be noted that I have discovered that VR uses AVX which is extremely taxing on a CPU...I uncovered this by running an AVX offset on my over clock and saw clock speeds drop to those levels when using VR.  

 

https://youtu.be/syw13Ww5RQE

 

not related directly to the OPs question, but I was pretty surprised by the results...and though you fellas might find it interesting.  

Edited November 7, 2019 by capt_nasties

[Voyager]

@capt_nasties Interesting. I'll have to watch that when I get a chance. So what frame rates did you get to?

 

Think you might be persuaded to run a similar test to the one I did earlier so we can do a psuedo apples to apples? 

[=SFG=capt_nasties]

I was averaging around 88 with the OC I believe...the results are at the end of the video, time stamp is in the description.   would be super interesting to see someone else perform a similar test.  the auto pilot function really makes this possible.  

Edited November 7, 2019 by capt_nasties

[SCG_Fenris_Wolf]

4 hours ago, capt_nasties said:

It should be noted that I have discovered that VR uses AVX which is extremely taxing on a CPU...I uncovered this by running an AVX offset on my over clock and saw clock speeds drop to those levels when using VR.  

 

https://youtu.be/syw13Ww5RQE

 

not related directly to the OPs question, but I was pretty surprised by the results...and though you fellas might find it interesting.  

That's very nice of you to mention,... I have found this out in 2017. It is pretty common knowledge by now. Or so I thought. ??

 

That chiliwili's benchmark and OC thread isn't on page 1 anymore, and you not knowing this, means that a lot of new guys are probably not aware of this. We need to re-open an OC/performance tuning thread for IL-2's VR.+

 

P.S. We did find out back then that IL-2 stops throttling via CPU at around 4.9GHz on i7 7700Ks. It was the magical border to break, we could maintain >80fps avg with lowest dips at 65fps. Since we have just received the new clouds, everything has changed again - for the better ?

Edited November 7, 2019 by SCG_Fenris_Wolf

[dburne]

I believe AVX is dependent on the application, and IL-2 does support AVX.

[chiliwili69]

9 hours ago, Voyager said:

At this point it looks like the limiter is my 3D card, however there are significant dips when I'm engaging the bombers, and you can really see that around 70 seconds in on both plots.

 

I'm thinking the drop on the second half of the native resolution run is the impact of being below and in the clouds eating my GPU for lunch

 

Although is a nice test, I don´t think this test demonstrates that your GPU is the limiting factor.

 

The fps values (or equivalent frametimes) are the sum of two factors: CPU frametimes + GPU framtimes.

 

At 90Hhz display, to achieve 90fps you need to produce frames every 11.1 ms. 

 

It could be that your new CPU (which I think is not OC) could produce frames at 9ms in this complex scenario, consuming most of the 11.1ms time. 

Then, using a very low resolution in the run#2 1080x1086 (this is like a fullHD monitor, so really nothing for a 1080Ti), the GPU would make his job in just 2ms or less, and you are under 11.1 ms most of the time. (except when CPU needs more than 11ms in bombers shooting)

 

Then, in your run#1 with full native resolution (2228x2180, it is 4 times more pixels) the GPU would need a bit more time to render the scene, perhaps around 8ms to make his job, but 9ms+8ms=17ms is well above the 11.1 limit.  In fact,  1000/17 is 58fps and your run#1 is really around 58fps.

 

You will realize this if you use fpsVR, which give you both CPU and GPU frametimes.

 

So don´t blame you great 1080Ti  ? . Try to overclock you 3800X to a fix frequency.

[Voyager]

@chiliwili69Good point.I hadn't considered that. I think I'll make a start of setting up an OC later this weekend. From what Iv'e been able to glean, Ryzen is not a simple or straight forward thing to push to its limits yet. I actually think I'm also going to start OCing my 1080 Ti to see just how far I can get it to go too. The Reverb resolution definitely loads it for bear, so I suspect I can gain frames from both.

 

@capt_nastiesI don't have Fortress on the Volga, so I wasn't able to run the first mission, so I did end up running the Battle of Kuban A-20 Airfield Attack mission instead, at the 1080 resolution, starting from when the aircraft crossed waypoint 2, and ended up with a min of 63, max of 92 and an average of 87.8. The mission was relatively uneventful though, so not sure how representative it really is. I might try the Mig-3 Scramble from Battle of Moscow instead.

[Alonzo]

Average FPS probably isn't a very useful figure. What you want is percentage of time where you're failing to make 90 FPS. I like the graph -- it gives you a sense of how much time you have smooth VR, how much you don't.

 

@chiliwili69 I still disagree with the way you add up GPU and CPU time per frame. I think due to pipelining, you need them both to be < 11.1ms, not the sum of both to be < 11.1ms. Certainly I have observed a solid 90 FPS on the Rift CV1 where both CPU and GPU frame times were in excess of 8ms each, which would be 16ms when added together.

[chiliwili69]

8 hours ago, Alonzo said:

I still disagree with the way you add up GPU and CPU time per frame.

 

Mmhh...Yes, you are quite right here. I had a wrong thinking about how CPU and GPU frametimes work.

 

Since my only experience with fpsVR was IL-2 I always thought the following:

CPU frametime: The time it takes the CPU to calculate the scene for every cycle considerring flying model, damage model, AI, etc

GPU frametime: the CPU frametime plus the time it takes the GPU to render the scene calculated by the CPU.

 

So, if the above is true, CPU frametimes can never be higher that GPU frametimes. But this is not true. So the above definition is false.

 

Sometimes GPU frametimes are higher than CPU frametimes, but CPU frametimes can be also higher than GPU frametimes. Here just two opposite examples: