Nvidia DX12 driver holding back Ryzen

N19h7m4r3 · 2 Apr 2017 at 10:23

Goldblum said:
Seeing as Ryzen, Vega and AM4 are all new they could have incorporated the support in design. If it works over the PCIE and it does mention both on and off die support we could well see it at a consumer level. It might also help with AMDs plan to reduce the amount of frame buffer.

I can see it play into their APUs, and console SoCs, it's their new "universal" interconnect.

So Ryzen APU + Vega GPU on die connected via "Infinity Fabric" with HBCC acting as a L4 cache, and allowing GPU to also access system RAM as needed.

Goldblum · 2 Apr 2017 at 10:29

N19h7m4r3 said:
I can see it play into their APUs, and console SoCs, it's their new "universal" interconnect.

So Ryzen APU + Vega GPU on die connected via "Infinity Fabric" with HBCC acting as a L4 cache, and allowing GPU to also access system RAM as needed.

Yeah for their APU market it could be great. It will be interesting if it is on the desktop products.

N19h7m4r3 · 2 Apr 2017 at 11:19

OH MY! 295X2 is beating the Titan X Pascal in DX12 Rise of the Tomb Raider as well. Where as Fury X gets smashed to a pulp by GTX 1070.

7 mins 47 seconds!

Panos · 2 Apr 2017 at 11:24

N19h7m4r3 said:
OH MY! 295X is beating the Titan X Pascal in DX12 Rise of the Tomb Raider as well. Where as Fury X gets smashed to a pulp by GTX 1070.

7 mins 47 seconds!

295X2 beating TXP on a Gameworks game.

However the point here is, that the AMD cards are working at 100% while the Nvidia cards not on the Ryzen system. The best the TXP was hovering around 60%!!!!!
While TXP and 1070 head on, exactly on the same AMD Ryzen do not operate at 100% showing similar performance also!!!!!!

Goldblum · 2 Apr 2017 at 11:28

N19h7m4r3 said:
OH MY! 295X is beating the Titan X Pascal in DX12 Rise of the Tomb Raider as well. Where as Fury X gets smashed to a pulp by GTX 1070.

7 mins 47 seconds!

Hawaii core always seems to gain well under DX12 and better than AMDs other cores.

Panos · 2 Apr 2017 at 11:57

Goldblum said:
Hawaii core always seems to gain well under DX12 and better than AMDs other cores.

And consider the 295X2 4GB VRAM is very slow GDDR5 at 5gbps. To put in perspective RX480 is using 8GB GDDR5 8gbps memory, TXP GRRD5X of 10gbps

Goldblum · 2 Apr 2017 at 11:59

Panos said:
And consider the 295X2 4GB VRAM is very slow GDDR5 at 5gbps. To put in perspective RX480 is using 8GB GDDR5 8gbps memory, TXP GRRD5X of 10gbps

And the Hawaii core uses less compression techniques to increase throughput

Panos · 2 Apr 2017 at 12:02

Goldblum said:
And the Hawaii core uses less compression techniques to increase throughput

True.
So the benchmark shows that there is some serious issues with the Nvidia driver, which doesn't affect only the Ryzen CPUs but the Intel enthusiast like the 6800/6900/6950X.

Rroff · 2 Apr 2017 at 12:13

Panos said:
True.
So the benchmark shows that there is some serious issues with the Nvidia driver, which doesn't affect only the Ryzen CPUs but the Intel enthusiast like the 6800/6900/6950X.

Not necessarily - it shows that there is something there that needs looking into more - as he even said in the video need to do a lot more benchmarks, etc. to see what is going on.

295X2 has, and it often isn't as crippled in mGPU scaling as other aspects, a lot of capability to shovel pixels/texels, etc. which may be making up some of the difference in some of those tests and so on tons of variables.

Goldblum · 2 Apr 2017 at 12:39

Rroff said:
Not necessarily - it shows that there is something there that needs looking into more - as he even said in the video need to do a lot more benchmarks, etc. to see what is going on.

295X2 has, and it often isn't as crippled in mGPU scaling as other aspects, a lot of capability to shovel pixels/texels, etc. which may be making up some of the difference in some of those tests and so on tons of variables.

Yes the scaling under mGPU is a lot better than traditional CF i think AMD claim over 90% on a pair of 480s but this still doesn't explain what we are seeing

fs123 · 2 Apr 2017 at 12:58

A new Ryzen Review tested with an RX480. Does pretty well against an i7 6900K.

https://thetechaltar.com/amd-ryzen-1800x-performance/2/

nashathedog · 2 Apr 2017 at 13:00

Snippety snip

hyperseven · 2 Apr 2017 at 13:04

nashathedog said:
Snippety snip

Why do you keep posting this?

Panos · 2 Apr 2017 at 13:06

Rroff said:
Not necessarily - it shows that there is something there that needs looking into more - as he even said in the video need to do a lot more benchmarks, etc. to see what is going on.

295X2 has, and it often isn't as crippled in mGPU scaling as other aspects, a lot of capability to shovel pixels/texels, etc. which may be making up some of the difference in some of those tests and so on tons of variables.

However why the TXP is not working in DX12 with Ryzen/Broadwel-E, showing exactly the same performance as the GTX1070 for that matter? While the AMD cards don't show this issue working at 100% regardless?

Rroff · 2 Apr 2017 at 13:09

Goldblum said:
Yes the scaling under mGPU is a lot better than traditional CF i think AMD claim over 90% on a pair of 480s but this still doesn't explain what we are seeing

If you look at the main DX12 run they did -on both the first CPU core (which is running the main game loop) has high utilisation but then look at CPU4 on the nVidia side - either Ryzen lacks the IPC to handle the nVidia scheduling thread there or its suffering from some compatibility issue with SMT. Hence the performance drop.

Panos said:
However why the TXP is not working in DX12 with Ryzen/Broadwel-E, showing exactly the same performance as the GTX1070 for that matter? While the AMD cards don't show this issue working at 100% regardless?

Looks like the same issue to me - individual Ryzen cores just aren't fast enough, or for some reason can't be utilised effectively enough, to handle the nVidia driver CPU threading hence that is the bottleneck regardless of GPU performance.

A lot of the time you can see it - that extra thread additional to the main game loop thread going high CPU often when it is hitting 80-90% utilisation the overall CPU and GPU utilisation is dropping which isn't manifest on the AMD GPU side where most of the cores are <60% utilisation aside from the main game loop thread.

EDIT: I should say though this needs looking into in more depth - there is still a lot of unknown variables and what I said above might be wrong or only part of what is going on.

Goldblum · 2 Apr 2017 at 14:45

Rroff said:
If you look at the main DX12 run they did -on both the first CPU core (which is running the main game loop) has high utilisation but then look at CPU4 on the nVidia side - either Ryzen lacks the IPC to handle the nVidia scheduling thread there or its suffering from some compatibility issue with SMT. Hence the performance drop.

Looks like the same issue to me - individual Ryzen cores just aren't fast enough, or for some reason can't be utilised effectively enough, to handle the nVidia driver CPU threading hence that is the bottleneck regardless of GPU performance.

A lot of the time you can see it - that extra thread additional to the main game loop thread going high CPU often when it is hitting 80-90% utilisation the overall CPU and GPU utilisation is dropping which isn't manifest on the AMD GPU side where most of the cores are <60% utilisation aside from the main game loop thread.

EDIT: I should say though this needs looking into in more depth - there is still a lot of unknown variables and what I said above might be wrong or only part of what is going on.

But that still doesnt explain why this issue is only on Nvidia GPUs, if this was a issue with just Ryzen the same would be seen no matter which GPU was used.
The Division test using 1060 and 470 showed similar issues

Rroff · 2 Apr 2017 at 14:49

AMD doesn't have the same functionality in the drivers - nVidia does a lot of software stuff from hooking DX functions in memory, command lists, scheduling (they do some stuff in software that AMD does in hardware) - which can be a bit of a mixed blessing - potentially you can actually lower CPU usage in some cases via the optimisation and on faster CPUs and faster GPUs can actually get better results ironically from being able to optimise throughput to the GPU but when the CPU isn't fast enough this can cause significant slowdown. What the actual issue is here isn't clear but that would be my guess from the behaviour of that extra busy thread alongside the main game loop one and why this is causing an issue if it is the issue with Ryzen isn't clear also - it could be being penalised by some kind of scheduling/SMT conflict or a utilisation bottleneck somewhere due to being busier with the additional functionality that thread is doing, etc.

Goldblum · 2 Apr 2017 at 14:57

It might be just due to the shortcuts Nvidia took on their DX12 driver to get around their architecture shortfalls, seeing as it only effects this API. It will be interesting when the true cause is identified

LoadsaMoney · 2 Apr 2017 at 14:58

N19h7m4r3 said:
OH MY! 295X2 is beating the Titan X Pascal in DX12 Rise of the Tomb Raider as well. Where as Fury X gets smashed to a pulp by GTX 1070.

7 mins 47 seconds!

Getting smashed!

TheRealDeal · 2 Apr 2017 at 15:04

LoadsaMoney said:
Getting smashed!

You can't laugh mate, you bought 2 of the pro variant

. Should have bought a 295

. Hawaii the architecture that just keeps giving.

Edit: I do thank you and all the Fury users for being test subjects so they can work out all the flaws for Vega which i think will be a real HBM architecture.