DX12 Performance Discussion And Analysis Thread

[Alessio1989]

Looking at the AMD slides, GCN Gen 4 has an updated hardware scheduler supporting high priority high responsive compute queue. They can be used also on other libraries (like TrueAudio that nobody cares about it) via proprietary extensions.

 

[Ext3h]

Looking at the AMD slides, GCN Gen 4 has an updated hardware scheduler supporting high priority high responsive compute queue. They can be used also on other libraries (like TrueAudio that nobody cares about it) via proprietary extensions.

Yes, and on inquiry, AMD confirmed that these features have also been backported to Fiji via firmware/microcode update.
It's not actually new hardware on this part, "only" a significant improvement on the firmware side.

 

[Jawed]

So Fiji firmware is upgraded by the driver?

 

[Alessio1989]

Yes, and on inquiry, AMD confirmed that these features have also been backported to Fiji via firmware/microcode update.
It's not actually new hardware on this part, "only" a significant improvement on the firmware side.

Any source about that?

I am not aware that AMD drivers are capable to update the GPU firmware. Last time I saw a microcode update was on my intel CPUs last year due a security bug..

 

[Ext3h]

They can't update the BIOS. But GPU initialization includes firmware upload for the ME (GPC) and MEC (HWS).

And micro code updates for CPUs are perfectly normal as well. When you flash a Bios you get the most recent micro code version bundled by default, but the OS will usually upload an updated version if the embedded one is outdated. (This does not persist across reboots, unlike the one loaded from the Bios.)

 

[silent_guy]

I am not aware that AMD drivers are capable to update the GPU firmware. Last time I saw a microcode update was on my intel CPUs last year due a security bug..

There is no complex chip in the world these days that doesn't have tons of small microcontrollers. And except for the boot loader itself, they all have empty code until a driver loads that code.

 

[Alessio1989]

I know it, but any source about updated AMD GCN Gen 3 microcode for scheduling? Because without any relevant source it smells like an urban legend..

 

[Kaotik]

I know it, but any source about updated AMD GCN Gen 3 microcode for scheduling? Because without any relevant source it smells like an urban legend..

Fiji block diagram from hot chips last year attached with 4xACE 2xHWS, original on release had 8xACE no HWS,

Around the same time Dave Baumann commented on it: https://forum.beyond3d.com/threads/...tion-rumor-thread.55600/page-136#post-1868282

 

[AlNom]

Around the same time Dave Baumann commented on it: https://forum.beyond3d.com/threads/...tion-rumor-thread.55600/page-136#post-1868282

Need to use the bottom right corner link in case folks have different PPP:
https://forum.beyond3d.com/posts/1868282/

 

[AlNom]

I know it, but any source about updated AMD GCN Gen 3 microcode for scheduling? Because without any relevant source it smells like an urban legend..

All I recall was this for Tonga



http://www.anandtech.com/show/8460/amd-radeon-r9-285-review/2

 

[Alessio1989]

Yes, but that's not the GCN Gen 4 scheduler feature I am talking about: they call it "quick response queue" and I guess it's all about a high priority compute queue implementation.

EDIT: yes, AMD itself claims they are able to enable via driver update on GCN Gen 2 and later

https://community.amd.com/community/gaming/blog/2016/03/28/asynchronous-shaders-evolved

But I cannot find any sources if they did it yet.

 

[Kaotik]

Yes, but that's not the GCN Gen 4 scheduler feature I am talking about: they call it "quick response queue" and I guess it's all about a high priority compute queue implementation.

EDIT: yes, AMD itself claims they are able to enable via driver update on GCN Gen 2 and later

https://community.amd.com/community/gaming/blog/2016/03/28/asynchronous-shaders-evolved

But I cannot find any sources if they did it yet.

Well at least Fiji is same 4 ACE 2 HWS configuration as Polaris 10 is, so I assume it at least supports those.

 

[Alessio1989]

last version of Intel GPA now includes a better DirectX 12 support: https://software.intel.com/en-us/articles/intel-gpa-release-notes

 

[pharma]

Rise of the Tomb Raider was updated today to version 1.0.638.6, which is a bigger deal than its less-than-snappy name suggests. In DirectX 12 mode Multi-GPU support is now available.

The update adds support for DirectX 12, which “on the right hardware can offer far better performance” than was previously achievable, as well as for Nvidia's VXAO ambient occlusion technology, described by Square Enix as "the world's most advanced real-time AO solution."

http://www.guru3d.com/news-story/rise-of-the-tomb-raider-gets-multi-gpu-directx-12-patch.html

 

[Kaotik]

http://www.guru3d.com/news-story/rise-of-the-tomb-raider-gets-multi-gpu-directx-12-patch.html

I don't know how Guru3D could have missed the earlier DX12 patch

 

[lanek]

I don't know how Guru3D could have missed the earlier DX12 patch

They have not miss it as they use it in the review, could be just an error when write the post.

 

[mczak]

Any source about that?

I am not aware that AMD drivers are capable to update the GPU firmware. Last time I saw a microcode update was on my intel CPUs last year due a security bug..

FWIW amd gpus using microcode (and different firmware files for different parts of the chip) isn't exactly a big secret. You can find the firmware for instance in the linux kernel repo (https://git.kernel.org/cgit/linux/kernel/git/firmware/linux-firmware.git/tree/amdgpu) - this should generally be the same as what's used by the closed source driver. The history is a bit lacking, you'd have to poke in the kernel driver parts to maybe learn a bit more about it (there's sometimes files with numbers in them, maybe the driver only loads one of them and so on, I don't know off-hand).
For the MEC and MEC2 firmware, these are all 256kB for GCN3 + GCN4 parts, so max size doesn't look like an issue to me for any gcn3 or newer part (if you look at the files, you can actually see they are only used 1/4 or so, the rest is just empty/repeating bytes). GCN1 (in the radeon directory) only seems to have ~8kB MEC files, whereas GCN2 had ~16kB ones.
(There's way more firmware for these chips, there's ce, mc, me, pfp, rlc, sdma, smc, uvd, vce parts apart from MEC, except for APUs which ditch mc as that's presumably part of the cpu...)

edit: btw, using firmware for gpus is an OLD feature. On the radeon line of gpus, it actually predates radeons... rage 128 was the first one (afaik) to feature firmware for the cp (2kB firmware file). Albeit all the way up from rage 128 to r520, there was just this firmware file for the command processor (all very similar too).

 

[Deleted member 13524]

Too bad that absolutely no one seems to be testing with FX CPUs anymore.
And too bad that DX12 didn't appear 3 years ago, either..


I actually have ordered a 10-core/20-thread Xeon and while it clocks rather low at 3GHz it's still put aside because I'll lose performance in DX11 games over my current 3.9GHz 4820K.
I definitely have to switch to it when I install AotS.

 

[lanek]

Too bad that absolutely no one seems to be testing with FX CPUs anymore.
And too bad that DX12 didn't appear 3 years ago, either..


I actually have ordered a 10-core/20-thread Xeon and while it clocks rather low at 3GHz it's still put aside because I'll lose performance in DX11 games over my current 3.9GHz 4820K.
I definitely have to switch to it when I install AotS.

Thats the problem, in most games quad core is the average of thread used, and only a few will use more than 6.. so you loose nearly 1ghz... You cant oc it a littllle beat on this plateform ?

One question if you permit it, why dont have take a LGA2011v3 version of 8-10cores that you can OC if needed for games ( the 6950X ) ? http://www.guru3d.com/articles_pages/core_i7_6950x_6900k_6850k_and_6800k_processor_review,21.html

I stilly have my 6cores 4930K.. I m really tempted to take a 8-10 cores, but honesty, i can OC this 4930K pretty high ( in the 5.5ghz range ), so on the end, im not so far behind the 8 cores.

 

[Deleted member 13524]

Thats the problem, in most games quad core is the average of thread used, and only a few will use more than 6.. so you loose nearly 1ghz... You cant oc it a littllle beat on this plateform ?

One question if you permit it, why dont have take a LGA2011v3 version of 8-10cores that you can OC ( the 6950X ) ? http://www.guru3d.com/articles_pages/core_i7_6950x_6900k_6850k_and_6800k_processor_review,21.html

I stilly have my 6cores 4930K.. I m really tempted to take a 8-10 cores, but honesty, i can OC this 4930K pretty high ( in the 5.5ghz range ), so on the end, im not so far behind the 8 cores.

I'm counting on the multithreaded nature of DX12 to show better performance with more cores, regardless of the number of cores you have, so I was counting on up(side?)grading when I'm done with the more demanding DX11 titles I want to play (right now I'm on MGS V).
I can't overclock because it's a LGA2011 Xeon E5 so not even the BCLK straps will work like the Core i7 variants. Regardless, the CPU was really cheap on e-bay (180€ IIRC) so I'll probably sell the current 4820k for less money.
I'm also attracted to the high core count because I'd like to use software video encoding when using Steam In-Home Streaming, which has a much better quality than AMD's hardware solution.

I'm not only changing the CPU for the games though. I'll be doing lots of Solidworks FEM, renders and OpenSIM runs for my phd so I'm "joining the useful and the pleasant" as we say in my country



But back to DX12 yeah I'm hoping for better results down the line with the low-clocked 10-core than with the higher-clocked 4-core.