Firstly, I'm not sure if this is the right segment to post this topic, but because 3DMark Time Spy benchmark relates a lot to hardware (CPUs and GPUs), I thought I'll post it here. To the Forum Admins, I'm sorry if I post this in wrong section/ segment. Appreciate if you can move this topic to the correct section/ segment if necessary.

3DMark Time Spy DX12 Benchmark came out a few days ago. The results and analysis has been pouring in the net. It also spurred quite a lot of discussion here, both in AMD and Nvidia threads, especially those still using or planning to buy 2nd hand past generation GPUs (Fiji, Maxwell etc) and also those who already bought or planning to buy the new AMD Polaris and Pascal GPUs.

I saw quite a number of user asking this question "Will my CPU bottleneck the new GPU I bought/ planning to buy?". Some still using i5-750 right up to the new i5-6600 Skylake and models in between (have not seen and I don't think there will be anyone with i7-6700K asking this question). I don't have any answer to this question.

But it got me thinking, I already know that overall score is dependent on physics test score (if the GPU is the same), which in turn depends on how powerful your CPU is. But does the CPU performance has any effect on the Graphic test result? So that's what I thought I wanted to test and share here. I hope this topic does not turn into thread to discuss which brand or model is better than the other. Even though I have Nvidia GPU now, I want to see this as a healthy discussion and I hope anyone with ideas or opinions can chip in, no matter if you use AMD, Nvidia or Intel products user

The method of my test was to change the number of my CPU active cores through BIOS and cycle it through the Time Spy bench. Brief details of my setup and methods

My PC spec:
- Intel i7-4930K 6 Core/ 12 Threads Ivy Bridge OC to 4.6Ghz across all cores. Hypertrading, Intel Turbo/ Speedstep enabled. Enermax ELC240 AIO Liquid CPU Cooler.
- Asus X79 Deluxe Mobo, 2 x 8GB Corsair Vengence DDR3 RAM OC 2133 MHz, Adata SP 900 256 GB SSD, Gigabyte GTX 1080 G1 Gaming GPU and Samsung 28" UHD 3840 x 2160 Monitor (through display port), Windows 10 Pro 64-Bit and latest Nvidia 368.81 driver

Test Method
- I use BIOS CPU Settings to change the "Active Core" parameter to 2, 4 or All (6) cores. Other setting remained unchanged.
- After each CPU core change and Windows boot, I will OC the GPU using Afterburner: +140 Mhz clock, + 190 Mhz memory, power limit 108% and temp limit 92C.
- I used HWMonitor to see CPU utilization and will only start the Time Spy bench after the CPU has idled (to minimize or avoid any apps that starts with windows boot from effecting the test run).
- Afterburner and HWMonitor closed before running each bench. After each Time Spy run, I saved the result locally to my PC and also validate online.
- I used 3DMark online compare to put the results of side by side for easy comparison, and the local PC saved file to view details including Monitoring Graphs

So here the results

3DMark Time Spy Default Bench Test Comparison


From left to right: 2C/4T, 4C/8T and 6C/12T

Firstly, please ignore maximum CPU turbo clock for 4C/8T run shown as only 3.4 Ghz. Had this problem since Firestrike time that sometimes 3DMark run did not show the correct turbo clock. Also, 3DMark showed # of core as 6 for all 3, but its the logical processor (4, 8 and 12 threads) indicates active CPU cores.

As I have expected, the overall scores scales with the CPU core/ thread count. But as you can see, regardless running 2, 4 or 6 cores on the CPU, there very little variance in Graphics Test 1 and Test 2 score or fps. The physics score and fps scaled linear as it went from 2-4-6 cores. So only the physics score contributed to overall score variance.

The default Time Spy bench runs on 1440P, so I thought will it change if I bump it up to 4K

3DMark Time Spy Custom 4K Bench Test Comparison


From left to right: 2C/4T, 4C/8T and 6C/12T.

To test 4K res, I had to run Time Spy in custom mode, which does not give any overall score. Anyway, again, very little variance in Graphic test 1 and 2 score/ fps, physics score scaled with CPU core count.

So looking at both result (and this is where I hope you guys can chip in your opinions because I don't know if my opion is right or not), this was my initial thought totally based on this results only; regardless if you have Pentium Dual Core, i5 or i7, the graphics performance remains the same. Only the overall score defers based on how powerful your CPU is. Nothing unexpected and kinda boring.

But when I look at the detailed results and CPU/ GPU monitoring graphs(from 3DMark app), that's where it got interesting. This result is from the same Custom 4K run.


Left to right pictures: 2 Core, 4 Core and 6 Cores Time Spy Custom 4K run results and monitoring graphs.

Comparing the result under the different test category

Graphic Test 1
While GPU was fully loaded, this test put little stress on the CPU. Regardless if its 2, 4 or 6 cores, the CPU speed only fluctuated between 1.2 Ghz to max 3.4 Ghz. However, you might notice that 2C graphs spent more time at 3.4 Ghz compared to 4 and 6 core test. Looked like non of the CPU went into Turbo

Graphic Test 2
Graphics test 2 now began to take effect on CPU. The 2C run CPU speed fluctuated between 2.6 Ghz and Turbo 4.6 Ghz now. Evethough 4C and 6C still didn't go above 3.4 Ghz, you will notice 4C spent a lot more time touching 3.4 GHz compared to 6C. Because the graphic score and fps doesn't vary much across the 3 test, my take is that, the GPU now requires the CPU with lesser core count to work harder to deliver the same level of performance as the higher core-count more powerful CPU.

Physics Test
This is the most interesting. The physics test fps result scaled with CPU core count with all the CPUs maxing out 4.6 Ghz. We look at the fps at period roughly 3/4 way through each test run. 2-Cores:7.72 fps, 4-Cores:15.65 fps and 6-Cores:23.19 fps.

But look at them GPU load. For 2-Core CPU, GPU was sweating at 40% load to deliver 7.7 fps. Whereas the 6-Core CPU almost didn't needed the GPU. A very walk in the park 6% load to deliver 23fps.

I thought, if the Physics test is only to stress the CPU, shouldn't the GPU load remained the same regardless of CPU performance? Is this an indication of CPU Bottlenecking or something else?

What do you guys think?

This post has been edited by adilz: Jul 20 2016, 11:42 PM

goldfries svfn Demonic Wrath Someonesim Acid_RuleZz

khelben skylinelover kerolzarmyfanboy Skylinestar

I believe this discussed many times already. i3 is actually powerful enough to handle mostly game at 1080p, if you pair with powerful gtx card.

if you compare same latest gen i3/i5/i7 i believe the fps will not differ very much unless the game is super heavy cpu usage

if upgrade from old procc to new, it's really hard to tell. I think best way to know whether it bottleneck or not, is to search google.

last time i oso know my cpu bottleneck by searching google. i see ppl dota 2 can get 100-120fps, while my 280x pair with phenom ii x3 710 just around 60fps, when upgraded to i3-4160 fps just around 80 (well although bottleneck but still 80fps is good enough as screen refresh rate just 60hz). so i know both my amd and i3 oso bottleneck my gpu performance.

now i upgraded to i7-4790k, no more bottleneck


btw 1080 so powerful, the graphic score is 3x of my 280x



This post has been edited by CyrusWong: Jul 17 2016, 06:26 AM

My result with 2c/2t


with 4c/8t


another run with 4c/8t


Like totally random in my case.

I can't find any reliable info on how the GPU usage is calculated (seems different from each vendor too). IMO, it might be measuring the driver queue to GPUs. On a slower/less core CPU, the driver/CPU couldn't send the work to the GPU fast enough, resulting in longer queue (i.e. higher GPU usage). But again, this is what I think, no reliable info.

Intel CPU improvement for the past 5 years has been lackluster. The i5 is going stagnant while the i7 is getting more core counts. Thanks to AMD for not having any competition.

Still on i5 2500K @ 4GHz, might need to run higher clock now.

Maybe your 2nd run, CPU & GPU little bit warmer.

So sad, AMD gamble their future on APU, and yet to get significant impact. In the end, some forumers said my i5 2500K @ 4.7GHz are near identical to latest i5 skylake, because Intel no need improve performance too much.

Actually my 2nd run is my 1st run.

I checked my previous Firestrike runs, sometime there some GPU load in the CPU test.

i think it depends on the game title, some games are more dependant on core clock freq, some on number of cores, or even HT. some games are also not that optimized.

there was a youtube video posted by forumer earlier about i5 bottlenecking 1070 in 144hz. i suppose it can be more apparent if >60fps on 1080p. but more tests are needed because that gpu utilization he showed could be just that the game that was tested on wasnt too demanding for the 1070 which has better architecure than 970. a better test would be comparing i7 with i5 on the same GPU.

i dont think 3DMark matters as much as actual game benchmarks.

This post has been edited by svfn: Jul 17 2016, 06:14 PM

svfnAcid_RuleZzSomeonesimCyrusWong All this while, its a generally accepted that you don't need more than 4 cores to run games, and generally prefers faster cores that more cores. Which is true for games we have to date runs on. Many benchmark confirmed this. Games run as fast on the i7-6700K 4-core as it is on 8-core i7-5960X. I don't know if that game engine limitation, or is it DX11 limitation. But as some of you guys pointed out already some reviews showed that with the new gen GPUs, games built for on DX11 already showed sign of CPU bottleneck (like video below showing GTX1070 is underutilized with i5-4960K). Correct me if i'm wrong, my understanding too is that at 1080P, games are more CPU bound than GPU bound (at least with current powerful GPUs)



One of the supposed advantage of DX12 is better utilization of multi-core/ multi-thread CPU. Since Time Spy was design to bench DX12, I thought I'll test to see what the impact would be, in addition to running 4K res. Looks like it does use all the available cores (too bad I sold my Xeon 8 core, otherwise would be interesting to test that too).

Obviously Time Spy bench can't be used as the indication of gaming experience, at least not now. But it would be interesting to know, what if future game DX12 game are developed to utilize all available cores on the CPU, from 2 cores right up to Xeon 10-12 cores, in addition to using a powerful GPU. Will we have a case of CPU/ GPU bottleneck?

If future games do use more than 4 cores, means my i7-4930K could potential still has a lot to offer for years to come. If games remains preferring faster cores or better IPC, than upgrades means having to balance between gaming performance and CPU-core intensive apps (like video editing). But if future games can utilize more than 6 cores, then I'll be saving money to go for the Xeon E5-2987W CPU

Demonic WrathYour explanation is plausible too. If I'm free, maybe I run the same 2/ 4 /6 Core test using the 3DMark API Overhead test. That gives the number of drawcalls and fps for both DX11 and DX12. See if the GPU CPU utilization exhibits the same behavior in the monitoring graphs.

This post has been edited by adilz: Jul 18 2016, 03:53 PM

In games, the CPU prepares tasks to feed the GPUs.

For DX11/OpenGL, there's an "Immediate Context" running on a main thread. This context talks to the driver alone and the driver will send the tasks to the GPU for processing. However, by only using a single thread, it might not have enough work to be submitted to the GPU.

If the game has multithreading support, it will use several threads to prepare "Deferred Contexts". This deferred/secondary context will submit the task to the "Immediate Context" so it can be sent to the driver and then to GPU for processing.

If the game or the driver doesn't support "Deferred Context/Command List", there will be a lot of holes in the "Immediate Context" causing it to be unable to feed the GPU with enough work, resulting in "CPU bottleneck".

This is also why AMD driver that doesn't support DX11 command list has "high driver overhead". (Well, it is supported only for some DX11 games)



In DX12/Vulkan, this "Immediate Context" middleman is removed. All threads can submit tasks in parallel to the driver through the use of queue. The driver will then submit the task to the GPU.



Note: In DX11/DX12/Vulkan etc, the driver still submits the tasks(command list) to the GPU in serial. However, this submission is very efficient. So, it's very low in overhead.

--
CPU bottleneck happens if there is not enough work to feed the GPU, either due to the lack of good threading or driver overhead or CPU is simply not fast enough.

In the case of GPU bottleneck, it means the GPU can't complete all the tasks fast enough.

In the end, it depends on how the developers code the game.

In DX11, some games don't exploit the deferred context too much resulting in lightly threaded code i.e. both CPU and GPU has low utilization.. In this case, faster IPC and clock speed will benefit more than higher core count.

TLDR: It is very hard to just generalize "Will this CPU bottleneck my GPU?". It depends on different games.

I kinda still prefer faster higher clocked processor than multiple. Because it eliminate the potential case where program/games cant utilize more core ( developer fault or simply API dont support ).

Found the video that convinced me to drop my itch for Skylake upgrade. An overclocked old processor can nearly match few gen newer processor's performance is a bit sad IMO.
Also the video noted that a faster RAM may improve performance a lot as well.
An i5 2500K @ 4.6GHz and 2133MHz DDR3 are on par or even outperformed stock i5 6500 and 2666MHz DDR4
An i7 3770K @ 4.4GHz and 2400MHz DDR3 are even on par with i5 6500 @ 4.5GHz and 3200MHz DDR4 in Crysis 3

I decided to do the same test like I did, but this time with Firestrike Extreme. Decided on Extreme because I think its the most likelihood upgrade path for those coming from 1080P and bought the new gen graphic cards. I think I don't need to explain in detail this time.

Another quick note, unlike my Time Spy test, Firestrike Extreme crashed with GTX 1080 OC'ed using 2-core setup. Tried twice to confirm its not GPU problem. ?????????? Indication DX11 is more demanding on CPU compared to DX12 (high level vs low level API)?????? So had to do all test with stock GPU settings.


Comparison left to right: 2-cores vs 4-cores vs 6 cores

Detailed individual test with monitoring graphs


2-Core


4-Core


6-Core

Quick summary of results

- Side by side comparison online between the 3 test showed that overall result differentiated by physics and combined test score because graphic test 1 and 2 almost the same for all. Though not much increase from 4-cores to 6-cores

- Combined Test; while there increase going from 2-cores to 4-cores (16 fps to 22 fps), there's no difference going for 4-cores to 6-cores (both ~ 22 fps)

- Detailed monitoring graph in particular the Physics Test (3rd segment in the graphs): Almost similar to Time Spy DX12 test except; while 2-core GPU load is higher, there's very little difference in 4-core and 6-core.

My non techie conclusion, looked like Firestrike Extreme hits 4-core wall. I don't know is it because of the way the test in coded/ designed by 3DMark OR limitation in DX11 using more that 4 CPU cores?.

If this is true and and taking into my Time Spy test into consideration, can I assume that if budget permits, its better to go at least for a 4-core CPU if you plan to to new rig to make the most of existing games and whilst future proofing.

SomeonesimDemonic WrathsvfnCyrusWongAcid_RuleZz