Terrible performance in Final Fantasy XIV with GTX 970

[JeffD]

Delirium, realize that you won't see %100 usage yet still be cpu bound. resource manager and task manager are fine estimates, but horrible when determining if a games main thread is getting cpu bound.

[Squa11_Leonhart]

Have a read through this comment block on why unity performs so badly, and hopefully (just hopefully) the way Xraid and others explain it, will help you understand whats up.

Another aspect may be in play here too... fill rate may be getting impacted by the amount of data getting dumped. Could be choking down coming through the network layers or through memory (system or graphics). I found on my 4800 series cards I got much more "bang for the buck" raising the clocks on the graphics memory than the core clocks. I also saw diminishing returns as my CPU got closer to a 4GHz overclock--but by revamping the timings to use a lower multiplier and faster FSB I saw slight improvements in fill rate (and load times).

May be something worth tinkering with to see if you can mitigate the slow downs a little. See if you can increase the system and graphic memory clock speeds a bit and see what happens?

Nah Raist, in Limsa especially, there is alot of object physics happening with things blowing about with the wind and the likes, this coupled with many player character models having moving cloth, earings and hair make for a CPU choke, DirectX and the display driver don't get the processing time they need to feed the GPU and so you get low gpu usage, resulting in low frame rate.

So having a 970gtx, I loaded up fraps and went to limsa, using the crowded bazzar as my test subject. My frame rates were being pulled to around 30-35 with an occasional dip below 20. Keep in mind my CPU is a FX6300, a little weaker then the one OP is using, and not overclocked at all.

I then went through all my video settings. The far majority of them did not impact fps at all. Only 2 settings did, and very little at that.. like 3-4 fps. One was shadows, if I turned them all off then I got a slight fps boost. Another was physics which also netted a 3-4 fps boost. The latter is of course linked to the cpu.

So simular to OP, we are not gpu bound, we are heavily cpu bound. Unfortunately there is very little we can do except probably drop resolutions. In the end however it is a problem with the game, not the gpu.

realtime Dynamic shadows are also linked to the CPU. (One reason Skyrims shadows were static and updated on a timer was because the sheer number of them would bottleneck every other aspect of the graphics rendering).

As for your cpu..... anything by amd is a heck of alot slower than the ops cpu where games like FFXIV are concerned.

emmm If it's crowded issue then try limit number of character showing to minimal. This should help you with FPS with crowded area. I know it feels sucks to lower the setting with new cards.. but just try it

Nemi, have already confirmed this resolves the issue: If you aren't rendering the objects or generating physics for them, then the GPU gets all the information it needs to hold 60fps in these locations.

Some more info:
CPU is at 70-80% utilization in Limsa. FPS drops to 32.
CPU is also at 70-80% in Gridana, FPS drops to 40 but then goes back to 60.
Same CPU utilization in Ul'dah, FPS drops to 50.

An FC mate who also has a 970 tried to OC his CPU from 3.8Ghz to 4.3 Ghz. There was no improvement on the FPS. We think it's an optimization issue rather than a GPU or CPU problem.

With the amount of threads FF14 uses (3 and a bit) 80% is the equivalent of maxed usage.

With multicore cpu's you don't watch the accumulated usage, you watch thread queues and the individual threads with tools such as process explorer.

And, its not so much an optimisation issue, as so much as it is that you have too more dynamic content being displayed than your cpu can take.

That said, this game works fairly smoothly until you go below 30fps, and if that happens, (say, during hunts) you can just drop the max objects on display and it'll bring the fps back up to playable.

[JeffD]

Edit: Post getting edited above me. So my 970gtx isn't maxing out. Lowering graphic settings does not improve performance. (2-3fps difference when your looking at a crowd of people as your benchmark is no big change). So on my fx6300, hitting just under 30fps in crowded cities and hunts, I am cpu bound.

[RobbieH]

The only legit performance issue on 900 series gpus in this game is the previously reported issue regarding dungeon portals and boss barriers (an issue in dynamic lighting recompilation on Maxwell gpu's).

LOL, that's an SLI issue.

I have the exact same problem on Dragon Age Inquisition, some towns i get 35 FPS, if i disable shadows it skyrockets the FPS again to 60 or 50 something, just like FFXIV, and guess what, it's going to get fixed, because nVIDIA only cares about AAA games, which is kinda ok since there's a whole lot more player base on those games so the chances of more people with issues are higher, still...

Just like Skyrim, i'd get 30FPS looking at Whiterun from the top of the town before the NVIDIA optimization patch and then poof the whole game was a lot smoother and the "CPU" bottlenecks post stopped appearing, this on the old 570. Yes, Bethesda also released a performance path that pretty much just made the game more stable by allowing the exe to use more than 4GB RAM which was done by the community long before they decided to patch it.

Ironically i didn't upgrade my CPU and the performance went from average 40 to average 60, interesting CPU bottlenecks going on.

Still i'm not worried anymore as i've stopped playing, there's a lot of issues on the 900 series but i'm ok with it as the upcoming drivers will hopefully fix most of the stuff, and i'm also tired of doing the same kind of end game since 2.0 i guess i'll wait for Gold Saucer on 2.5 part 2 so i'll have something else to do.

Or maybe i'll just wait for the DX11 client.

Just curious, would a 12 Thread CPU fix the problem?

[Squa11_Leonhart]

if i disable shadows it skyrockets the FPS again to 60 or 50 something, just like FFXIV, and guess what, it's going to get fixed,

Probably by recompiling the behavior of these shadows. Or by nvidia working with bioware to make it so the shadows are in larger batches... or optimising culling for shadows that aren't actually in view.

Dynamic shadows are a dog to render at a high framerate without using them sparingly, or faking them.

Just like Skyrim, i'd get 30FPS looking at Whiterun from the top of the town before the NVIDIA optimization patch and then poof the whole game was a lot smoother and the "CPU" bottlenecks post stopped appearing, this on the old 570.

LOLWUT? there was no "NVIDIA optimization patch"! the patch (1.3 btw) was the proper usage of SSE (ie, actually turning it on >.>) instead of x87 fpu.

Others might remember TESV Acceleration Layer and Skyboost, these were code injected optimised replacements for the 'unoptimised' code using SSE/SSE2 that were made redundant with 1.3

Ironically i didn't upgrade my CPU and the performance went from average 40 to average 60, interesting CPU bottlenecks going on.

the x87 fpu is a hell of a lot slower on core 2 duo and amd phenom/athlon cpu's.
For basic arithmetic on nahalem and newer, this unit is on par with SSE, but as functions become more complex the performance tapers off

Skyrim 1.0-1.2 was using code that had not been inlined (improves pipeline execution and branch prediction) or sse optimised, which in a basic manner of explanation means the compiler rewrites standard C++ into instruction/machine code (rather then having to use assembly).

Yes, Bethesda also released a performance path that pretty much just made the game more stable by allowing the exe to use more than 4GB RAM which was done by the community long before they decided to patch it.

That was not a performance patch, it was a simple change of the compiler options to enable Large Address Aware, a setting that allows a 32bit application to use the 4GB user address spaces provided to 32bit applications under 64bit operating systems.

It does not allow more than 4GB to be used at all, and Virtual Address Space is often consumed well before the working set reaches that point.

The 4GB patch was only relevant to heavy texture modders anyway.

Just curious, would a 12 Thread CPU fix the problem?

NO.
The game only scales to 3 cores (load balanced of course).

[RobbieH]

Thanks for taking the time to reply, and unlike other times you had better self control this time, kudos for that lol

Probably by recompiling the behavior of these shadows. Or by nvidia working with bioware to make it so the shadows are in larger batches... or optimising culling for shadows that aren't actually in view.

Dynamic shadows are a dog to render at a high framerate without using them sparingly, or faking them.

LOLWUT? there was no "NVIDIA optimization patch"! the patch (1.3 btw) was the proper usage of SSE (ie, actually turning it on >.>) instead of x87 fpu.

Others might remember TESV Acceleration Layer and Skyboost, these were code injected optimised replacements for the 'unoptimised' code using SSE/SSE2 that were made redundant with 1.3

the x87 fpu is a hell of a lot slower on core 2 duo and amd phenom/athlon cpu's.
For basic arithmetic on nahalem and newer, this unit is on par with SSE, but as functions become more complex the performance tapers off

Skyrim 1.0-1.2 was using code that had not been inlined (improves pipeline execution and branch prediction) or sse optimised, which in a basic manner of explanation means the compiler rewrites standard C++ into instruction/machine code (rather then having to use assembly).

That was not a performance patch, it was a simple change of the compiler options to enable Large Address Aware, a setting that allows a 32bit application to use the 4GB user address spaces provided to 32bit applications under 64bit operating systems.

It does not allow more than 4GB to be used at all, and Virtual Address Space is often consumed well before the working set reaches that point.

The 4GB patch was only relevant to heavy texture modders anyway.

NO.
The game only scales to 3 cores (load balanced of course).

I know very well the Acceleration layer, Skyboost did nothing for me, for some reason, the LAA that the community discovered helped a lot with random crashes and mods, not just texture mods, if i recall it even helped with the base game and a few mods and no texture mods, sure some mods have their own textures however it was just follower tweaks and gameplay changes, anyway yes patch 1.3 helped a lot.

LOLWUT? there was no "NVIDIA optimization patch"!

Driver*

http://www.geforce.com/whats-new/art...ivers-released

NO.
The game only scales to 3 cores (load balanced of course).

Btw the CPU i was talking about has 6 cores and 12 Threads.

Then that means it's not CPU bottleneck and instead a bad optimization, do you think DX11 client will fix the problem? I know DX11 tends to be the slowest (vs DX9, DX10 is a mess), sadly DX12 isn't that near so i guess we'll have to rely on DX11, hopefully.

Thanks.

[Squa11_Leonhart]

Thanks for taking the time to reply, and unlike other times you had better self control this time, kudos for that lol

I know very well the Acceleration layer, Skyboost did nothing for me, for some reason, the LAA that the community discovered helped a lot with random crashes and mods, not just texture mods, if i recall it even helped with the base game and a few mods and no texture mods, sure some mods have their own textures however it was just follower tweaks and gameplay changes, anyway yes patch 1.3 helped a lot.

The most recent improvement to stability comes from increasing the allocation blocksize, this feature has been integrated in the latest skse
The main case where LAA involved non-texture mods, was where items had duplicated excessively in certain locations.... and geez that was the most annoying bug...

Driver*

http://www.geforce.com/whats-new/art...ivers-released

I don't recall that article, however as it reads, the improvements were mostly indoors. As you can see by the outdoors image, the performance improvement was little or none.
The biggest improvement to whitebrim was the SSE optimisations being enabled in 1.3.

Btw what CPU did you have at that time, i am aware that AMD's processors benefited less from the optimisations due to the creator using Intel's ICC.

Then that means it's not CPU bottleneck and instead a bad optimization, do you think DX11 client will fix the problem? I know DX11 tends to be the slowest (vs DX9, DX10 is a mess), sadly DX12 isn't that near so i guess we'll have to rely on DX11, hopefully.

Thanks.

Not at all, in fact for what effects are available, the optimisation is pretty well done. However there is only so much you can do with model animations and physics before you hit a the limits of what the cpu can do for whats on display.

Assassins Creed Unity is an example of bad optimisation, 50,000 draw calls per frame using DX11 which works best at 10,000.

[RobbieH]

The most recent improvement to stability comes from increasing the allocation blocksize, this feature has been integrated in the latest skse
The main case where LAA involved non-texture mods, was where items had duplicated excessively in certain locations.... and geez that was the most annoying bug...


I don't recall that article, however as it reads, the improvements were mostly indoors. As you can see by the outdoors image, the performance improvement was little or none.
The biggest improvement to whitebrim was the SSE optimisations being enabled in 1.3.

Btw what CPU did you have at that time, i am aware that AMD's processors benefited less from the optimisations due to the creator using Intel's ICC.

Not at all, in fact for what effects are available, the optimisation is pretty well done. However there is only so much you can do with model animations and physics before you hit a the limits of what the cpu can do for whats on display.

Assassins Creed Unity is an example of bad optimisation, 50,000 draw calls per frame using DX11 which works best at 10,000.

Before the 4670k i had a Q9550 at 3.8Ghz and that's where i saw the optimizations.

Still i'm confused, you say the game only scales to 3 cores, and that a 6 core CPU (12 Threads) won't help, so is it a DX9 limitation of some sort? I just can't get over it how the 570 performed better on most CPU intensive areas on same settings, i would get min 35FPS at any time, with the 970 i get min 25. I've removed the OC completely and i have the same exact performance, forced OC to 4.4Ghz makes no difference either, it made an huge difference when i had the 570 and upgraded from the Q9550, even without OC.

Will the DX11 client help with it if done properly?

[Squa11_Leonhart]

Still i'm confused, you say the game only scales to 3 cores, and that a 6 core CPU (12 Threads) won't help, so is it a DX9 limitation of some sort?

Well when you multicore optimise a game, it all depends on what needs to be done.
Direct3D 9 by itself was not designed with api side multithreading, and attempting to force it will result in significant overhead due to state synchronisation -
So all the D3D9 behavior will operate in tandem with the games master thread on on the same cores (if bounced about by load balancing) as other I/O such as input. Audio may be spun off onto another thread, as may I/O such as Network and disk access.

Physics and AI can be multithreaded, but you still end up with the one single master thread having to synchronise everything back together in the end.

The max amount of CPU that a user space thread can use is 100% of a single cpu core, so when looking at that thread in a tool such as process explorer, it might be 13(12.5 rounded up), or 25 or 50% depending on if your cpu has 8, 4 or 2 cores(or hyperthreads).

a kernel space thread, such as from the driver can utilise multiple cores, hence why i've seen nvd3dum.dll!QueryOglResource threads running at 25ish% on 4core/8htt processors where sli is used.

Anyway, as far as i can see, most of the overhead(bottleneck) is coming from D3D9 being fed too much information.

I just can't get over it how the 570 performed better on most CPU intensive areas on same settings, i would get min 35FPS at any time, with the 970 i get min 25.

Can you readily verify that the the same amount of player characters were in the scene, and wearing the same clothes between tests?, this is the problem, as the more people you have wearing unique outfits, the more the api has to update for the differences.

Will the DX11 client help with it if done properly?

If Defferred Rendering is utilised efficiently, then State changes will be minimised, draw calls processed more efficiently and the frame rate should definitely be steadier.

[Syx]

Squall, I distinctly remember touching on the topic of CPU, DX9, multithreading, and bottlenecks over a year ago. Funny how this is issue is continuously broached. There are some website links in my post linked below that might help victims of Squeenix's programming understand the issues being faced at this time by users with enthusiast-tier computers.

http://forum.square-enix.com/ffxiv/t...=1#post1234692

The codermind.com article has been locked down, so here is a webarchive snapshot:

http://web.archive.org/web/201303130...Direct3D9.html

It's all interesting stuff.