[Part 2] Measuring CPU Draw Call Performance

[tomatone]

It's just bizarre man. Your Haswell Pentium with an AMD GPU scores up with the fully fledged i7's paired with an NVidia GPU (NV's driver being "optimized" for such synthetics).

Maybe people are not aware of the difference in FPS when the window is in focus and out of focus.

FPS: 11.84-11.91 window focused, 9.59-9.61 not focused, 12.12-12.15 fullscreen

 

[MajinCry]

Maybe people are not aware of the difference in FPS when the window is in focus and out of focus.

It'd be obvious which results were off-focus; they'd be in single digits or ~12fps if on a modern, >=4ghz intel CPU.

What we need now, is for someone with Ryzen to test the benchmark on Win 7 with SMT disabled.

 

[tomatone]

When I'm running test off-focus FPS hovering around 18.

 

[Atari2600]

Sweet.

Does the fact that a Ryzen would appear to have the same draw rate as a Deneb really not set off any alarm bells with you?

edit: My bad, I was comparing W10 to W7 - even so. Its only 15% faster (clock normalised).

Given what is known about improvements elsewhere that still means the benchmark is likely inappropriate for use right now.

 

[MajinCry]

Does the fact that a Ryzen would appear to have the same draw rate as a Deneb really not set off any alarm bells with you?

edit: My bad, I was comparing W10 to W7 - even so. Its only 15% faster (clock normalised).

Given what is known about improvements elsewhere that still means the benchmark is likely inappropriate for use right now.

Assembled a few scores for the architectures. FPS / (Clock * 10) = Score.


Phenom II: 12.1 / 34 = 0.355882353

Piledriver: 11.13 / 36 = 0.309166667

Ryzen 1800X: 16.53 / 40 = 0.41325


Wolfdale 8300: 12.12 / 30 = 0.404

Lynnfield 860: 17.15 / 36 = 0.486111111

Sandybridge 2310: 13.37 / 31 = 0.431290323

Sandybridge 2500k: 19 / 45 = 0.422222222

Haswell 4771: 17.78 / 39 = 0.455897436

Skylake 6600: 14.77 / 35 = 0.422


Keep in mind that the draw call disparity is way more pronounced in real world (non synthetic) scenarios. The Lynnfield scores a bit over 50% better than Phenom II, but in Fallout 4's Corvega factory, when overlooking Lexington, it will have >3x better framerates. Tested that with a user over on the ENB forums.

 

[tbob18]

E5-1660 @ 4.6ghz
R9 390 @ stock

 

[MajinCry]

E5-1660 @ 4.6ghz
R9 390 @ stock

Please fill out the rest of your specs, kinda good to have all the details.

 

[CatMerc]

CPU: Core i5 6600K @ 4.2GHz
GPU: R9 390 @ 1025MHz
GPU Driver: 17.2.1
OS: Win 10 64bit

Ships: 1
Rocks: 16k
Draw Calls: 16,022
FPS: 23~ FPS

 

[MajinCry]

Updated results, added graph with scores.

 

[bigboxes]

CPU: i7 4790K @ 4.0GHz
GPU: MSI GTX 970 Gaming 4G
GPU driver: 378.78

Ships: 1
Rocks: 16000
Draw calls: 16022
FPS: ~31

 

[Magic Hate Ball]

CPU: Ryzen 1800X @ 4.0ghz SMT Disabled RAM 3200
GPU: Sapphire Fury Tri-X OC @ 1050mhz
GPU driver: 17.3.1

Ships: 1
Rocks: 16000
Draw calls: 16022
FPS: ~17.72

If I let the core affinity alone and it sits on two separate CCX's the FPS drops to ~14.5

 

[MajinCry]

CPU: Ryzen 1800X @ 4.0ghz SMT Disabled RAM 3200
GPU: Sapphire Fury Tri-X OC @ 1050mhz
GPU driver: 17.3.1

Ships: 1
Rocks: 16000
Draw calls: 16022
FPS: ~17.72

If I let the core affinity alone and it sits on two separate CCX's the FPS drops to ~14.5

Ooh. I'll add it to the list 'n' graphs. If you set your RAM to 2133Mhz, what happens to your framerate, for both one CCX, and two CCX's used?

 

[Magic Hate Ball]

Ooh. I'll add it to the list 'n' graphs. If you set your RAM to 2133Mhz, what happens to your framerate, for both one CCX, and two CCX's used?

16.65 or so with 2133 and same CCX.

13.68 with different CCX at 2133 still on RAM.

 

[MajinCry]

16.65 or so with 2133 and same CCX.

13.68 with different CCX at 2133 still on RAM.

With the 3200Mhz RAM, it's almost at Haswell levels of draw call perf. Cor. It's a shame the 6 and 4 core parts will have equally split CCX's.

 

[Magic Hate Ball]

With the 3200Mhz RAM, it's almost at Haswell levels of draw call perf. Cor. It's a shame the 6 and 4 core parts will have equally split CCX's.

Actually, strike that, I rebooted back into Windows after setting my RAM back to 3200 and now it's a tiny bit higher at 17.83, odd.

That is for same CCX, everything the same as my first test but slightly higher FPS steady.

EDIT: Oh, and it's just a Regular Fury, not a Fury X, despite the deceptive model name

 

[Excessi0n]

CPU: 6700K @ 4.9GHz, 3200MHz RAM
GPU: GTX 660Ti
GPU Driver: 376.53
OS: Windows 10 Pro x64

Ships: 1
Rocks: 16000
Draw Calls: 16022
FPS: 31.03

 

[DrMrLordX]

CPU: Ryzen 7 1800X @ 4.0 GHz, Performance Mode/OC Mode, DDR4-3200 14-14-14-32, SMT ON
GPU: Powercolor R9 390 1010 MHz GPU/1500 MHz RAM
GPU Driver: uh 16.6? Whatever it is comes in from MS right now. Haven't updated . . .
OS: Win10 Pro x64

Ships:1
Rocks:16000
Draw Calls: 16022
FPS: ~17.5

 

[artins90]

CPU: i7 7820x 4.8 Ghz
GPU: Gtx 970 1557 core / 4000 memory
GPU Driver: 382.33
OS: Win10 Pro x64

Ships:1
Rocks:16000
Draw Calls: 16022
FPS: ~ 29.65

Proof:

 

[MajinCry]

Unfortunately, due to driver shenanigans from NVidia, the seemingly good results are useless as an indicator of draw call performance. It involves a specific optimization that could only ever be used in extremely simplistic, synthetic tests such as the one in this thread; when there is only one object, with no lights, materials, shadow map, parallax maps, etc, so the whole scene consists of duplicate draw calls of a single source, NVidia's driver will appear about twice as fast as AMD's.

But it's not, it only appears to be when we configure the synthetic as outlined in OP.

What I'd like to see, is how Ryzen performs after all the microcode and bios updates.

 

[Mr Evil]

...What I'd like to see, is how Ryzen performs after all the microcode and bios updates.

CPU: Ryzen 1800X (stock)
GPU: Fury
GPU Driver: 17.7.2
OS: Win 10 x64

Ships: 1
Rocks: 16000
Draw Calls: 16022
FPS: 15.66

Motherboard is an ASRock Taichi with the most recent 3.00 BIOS, which uses AGESA 1.0.0.6a. I don't have figures for older BIOS versions, so I don't know if performance has changed.

 

[MajinCry]

CPU: Ryzen 1800X (stock)
GPU: Fury
GPU Driver: 17.7.2
OS: Win 10 x64

Ships: 1
Rocks: 16000
Draw Calls: 16022
FPS: 15.66

Motherboard is an ASRock Taichi with the most recent 3.00 BIOS, which uses AGESA 1.0.0.6a. I don't have figures for older BIOS versions, so I don't know if performance has changed.

Was the program using two CCX's? Performance should jump up if you assign it to the first/last four cores.

 

[Mr Evil]

Ok, I ran it again, then set the affinity to the first 4 cores and it made no difference.

EDIT: And looking at the charts in Task Manager, threads are only being assigned to the first 4 logical cores anyway. I'm using the standard "balanced" power plan.

 

[MajinCry]

Ok, I ran it again, then set the affinity to the first 4 cores and it made no difference.

EDIT: And looking at the charts in Task Manager, threads are only being assigned to the first 4 logical cores anyway. I'm using the standard "balanced" power plan.

Why would you use balanced? Set it to high performance, see if it makes a difference.

You should be getting 15% higher frames. What's your RAM speed?

 

[Mr Evil]

I keep it in balanced mode because the power consumption is lower. I tested again in high performance mode and got 16.12 fps. RAM is running at 2933MHz, CAS 18.

Why do you think it should be higher though? Page 1 shows a stock 1800X and Nano getting 12.8 and 14.69. I would expect to get higher than that, but slower than the 17.72 for an overclocked 1800X with Fury X. 15.66 (or 16.12) thus fits my expectations.

 

[MajinCry]

I keep it in balanced mode because the power consumption is lower. I tested again in high performance mode and got 16.12 fps. RAM is running at 2933MHz, CAS 18.

Why do you think it should be higher though? Page 1 shows a stock 1800X and Nano getting 12.8 and 14.69. I would expect to get higher than that, but slower than the 17.72 for an overclocked 1800X with Fury X. 15.66 (or 16.12) thus fits my expectations.

I assumed that bios updates and such would have improved performance. My mistake.