PDA

View Full Version : Image Editing - Optimization



Gobe
04-09-2016, 02:39 PM
I thought I'd post this as a new thread since there's been some talk on this in a couple places.

Here's the question: Why can I top 270K on image editing when the next nearest score is a solid 18K lower? If anyone has something they want me to try like recording my secondary/tertiary memory timings, I'll try to get to it if it sounds reasonable and post my results.

The system that I'm running is based around an Intel i7 5960x (pre-binned) running currently on an ASUS X99A USB 3.1 motherboard. My storage is dual Samsung 950 Pro NVMe (no raid). Memory is G.Skill Ripjaws V (32 GB Quad) with XMP at 3200/14-14-14-34. I run 1T. Video is dual GTX-960 (SLI). CPU cooling is via a dedicated custom loop.

Edit: Forgot to mention, Windows 7 professional.

Here's the first test I've run:

CPU = 4,900 MHz
Cache = 4,500 MHz
Dram = 3,200 MHz @ 14-14-14-34-1T

Let's see if the storage device makes a difference.

I installed various storage devices and made 5 runs for each of the RB image editing benchmark. Standard tweaks: Diagnostic Mode, 800x600, Explorer closed, and RealBench at realtime priority. NTFS for all. Best score for each device is recorded.

Samsung 950 Pro NVMe (PCIe) - 271,794
Toshiba Q Series Pro SSD SATA - 271,562
WD 7200 RPM SATA - 271,128
Toshiba 7200 RPM SATA via USB 3.0 - 271,345
USB 1.0 Flash - 108,717
USB 2.0 Flash - 250,689
USB 3.0 Flash - 152,198

It ain't the storage. NVMe and SATA whether solid state or spinner all clustered at 271K including a SATA spinner via USB 3.0. The USB 1.0 flash drive introduced a real bottleneck, but the performance hit for the USB 2.0 flash drive was only 21K. The USB 3.0 flash drive (PNY @ 32 GB) was interesting as it didn't perform nearly as well as the USB 2.0 flash drive (DataTraveler 100 G2 @ 32 GB) . Not all flash memory is the same. Apparently my USB 3.0 memory stick sucks.

Next, we'll drop the cache from 4,500 MHz to 3,000 MHz and see what we get.

Gobe
04-09-2016, 02:56 PM
Cache @ 3,000 MHz

For this test, all I did was drop the cache from 4,500 MHz to 3,000 MHz. Everything else is the same, same 5 runs with the best score recorded. I only used the Samsung 950 Pro NVMe here since the previous storage test says the storage isn't the magic bullet.

Cache at 3,000 MHz = 262,809

OK, here's a noticeable drop-off of around 9K dropping the cache from 4,500 MHz to 3,000 MHz. The cache speed does matter, but it doesn't explain my 18K lead on the rest of the field since I'm still 9K ahead of the field with the cache loitering along at 3,000 MHz. The faster the cache, the better, but there probably won't be a dramatic improvement going from say 4,200 MHz to 4,500 MHZ or higher.

Next, we'll return the cache to 4,500 MHz and start farting around with the DRAM.

Gobe
04-09-2016, 03:09 PM
Primary Memory Timings

OK, everything back to the settings from the initial test running with the Samsung 950 Pro NVMe.

CPU = 4,900 MHz
Cache = 4,500 MHz
DRAM = 3,200 MHz @ 14-14-14-34-1T

Let's loosen up those timings.

DRAM = 3,200 MHz @ 18-18-18-42-1T = 271,562

Well, look at that! Primary memory timings don't seem to matter... at all.

Gobe
04-09-2016, 03:21 PM
Memory Speed

Back to initial conditions:

CPU = 4,900 MHz
Cache = 4,500 MHz
DRAM = 3,200 MHz @ 14-14-14-34-1T

That DRAM looks awfully speedy at 3,200 MHz. Let's fix that.

DRAM = 2,133 MHz @ 14-14-14-34-1T = 269,548

So, yea, memory speed matters... barely. Dropping from 3,200 MHz down to 2,133 MHz drops us barely 2K.

Gobe
04-09-2016, 03:37 PM
I guess next I'll sabotage the whole works.

Our initial conditions:

CPU = 4,900 MHz
Cache = 4,500 MHz
DRAM = 3,200 MHz @ 14-14-14-34-1T

We'll keep the CPU speed the same since we know it makes a big difference. Let's bog down the cache speed AND memory timings and speed.

CPU = 4,900 MHz
Cache = 3,000 MHz
DRAM = 1,866 MHz @ 20-20-20-48-1T

259,883

We've choked the memory all the way back to 1,866 MHz from the original 3,200 MHz and we've played Hell with the timings loosening them from 14-14-14-34 to 20-20-20-48. We've dropped the cache from 4,500 MHz to 3,000 MHz. I also replaced the water in the cooling loop with maple syrup.

Well, I didn't do that last thing.

We're down 12K from the original test... still ahead of the field by 6K.

Gobe
04-09-2016, 03:47 PM
OK, that's all I've got.

I tested various storage devices and I threw wrenches into my memory speed and timings and cache speed both individually and in combination. And here's the damned thing, my image editing scores are still out of this world.

I'm beginning to wonder if there's something to the way I've got my OS (Win7 Pro) configured. I've got no idea what it could be since I just do the same stuff every time I do a new OS installation.

I know, I know, I can hear Arne screaming, "What are your damned secondary/tertiary memory timings?" I'll look at those this weekend and post them.

I'm also thinking about doing a Win7 Pro install to an SSD without doing all the usual stuff I do and see if that makes a difference.

Arne Saknussemm
04-09-2016, 04:21 PM
LOL...not screaming...just talking loudly...I have a kit of 32GB Ripjaws V incoming so will find out....:o

I'm on Win 10...if I get a chance I will try a win 8.1 run....you strip win7?

I clean installed W10 and just had RB there and got 257000 image....but no ice to do a full run...

Menthol
04-09-2016, 06:23 PM
What vcore does it take you guys to complete RB at 4.9ghz?

cekim
04-09-2016, 06:38 PM
What vcore does it take you guys to complete RB at 4.9ghz?
I haven't finished it at 4.9 yet - only 4875 at 1.37.

Diagnostic mode? Real-time? 800x600? It's like you guys are benchmarking? ;-) I had lowered the resolution, but not all of that stuff.

Arne Saknussemm
04-09-2016, 06:50 PM
What vcore does it take you guys to complete RB at 4.9ghz?

I have limited myself to 1.45 max and I don't think 4.9 will clear for me at that...4845 was my run...

Gobe
04-09-2016, 08:11 PM
4.9 is very problematic. I've done it for the full test, but it's more trouble than it's worth. Hours of frustration for maybe one complete test.

For the image editing, I'm taking advantage of the fact that image editing really doesn't push the CPU very hard.

Now I come back to the rig and I find that one of the graphic cards is lights out and the other is acting all OpenCL error.

Gobe
04-10-2016, 07:30 AM
Yep. Moved the cards to another system and one of them is lights out (problem LED on motherboard remains on) and the other still works but OpenCL apps crash, browser crashes, etc. Have 2 days left for Amazon to accept a return for refund but the other one has to go back to EVGA.

Guess I'll shop me some new cards, maybe upgrade, and sell the replacement once the EVGA RMA is finished.

I wonder what the heck happened?!

Arne Saknussemm
04-10-2016, 09:14 AM
Jeepers! You leave the system plugged in while away? Thunderstorm?

Were you using latest Nvidia driver?

Different PSU on both systems right?

cekim
04-10-2016, 09:18 AM
Ok, there has to be something else and/or "diagnostic mode" is not what I think it is and matters...

So, I ran at 5.0GHz - as Gobe points out Image editing doesn't really stress the CPU (which begs the question of why CPU clock speed matters so much???). I can run this at 5.0GHz 1.35v.

w/ speed-step enabled - 5.0GHz:
253832
w/o speed-step enabled - 5.0GHz:
256608

When you say "diagnostic mode" what exactly do you mean? Not "safe mode" I gather as that seems to peg my clock at 3.0GHz. Both of the above runs were 800x600 and real-time priority. Both were run 2x to get a "warm" cache speed-up.

Arne Saknussemm
04-10-2016, 09:22 AM
Run msconfig and select diagnostic start-up and click ok not apply....it will reboot

56788

Menthol
04-10-2016, 02:13 PM
I am not sure why Gobe is getting a higher score, I played a little yesterday at 4.9 and was getting 255 to 265k, for some reason his system is operating a little more efficient, but he's doing something right

Menthol
04-10-2016, 06:26 PM
Are you ready? Who's the King of all now, suck it Xeon

56802

Wait, I wasn't satisfied with that, tried 5ghz but RB wasn't having anything to do with that on chilled water, maybe SS
Anyway I broke out of the 220K barrier with this, oh and 270K image editing, for beginners

56803

Well there went half a day taking my system out of the case, setting up the water chiller, guess I'll turn it of until someone tops this
Arne give me a day to bask in glory please it's Sunday (Arne had his chip blessed by pope Francis and runs holy water in his system)

cekim
04-10-2016, 06:39 PM
Suck it Xeon.Chuckle.... Worth it! ;-)

Arne Saknussemm
04-10-2016, 07:39 PM
had his chip blessed by pope Francis and runs holy water in his system)

LOL

Nice Menthol!...cracked 230 I won't be beating that I'm sure.....Win 7?

:cool:

Menthol
04-10-2016, 08:38 PM
Win 7 yes, I wonder if that stripped Win 8 would be better, but I'm satisfied for now, now to go back to doing nothing until 10 core CPU's and Pascal are released

Gobe
04-10-2016, 10:57 PM
That should read, "Suck it DUAL Xeon".

Seriously, nice score there. Glad to know somebody else has the image editing sussed out now.

I can push up to 5 GHz and put in some terrifying image editing scores, but there's not a prayer that I could run the more CPU intensive benchmarks without freezing things up. Until some smart-ass puts in a DICE or LN2 score, the 270's probably represents balls to the wall on the 5960X.

I'm still doing the post-mortem on what happened to my system yesterday. Both GTX-960s are bad. The one that's lights out won't fire up the monitor and the other errors out and crashes apps. One is on it's way back to Amazon for refund and the other is on it's way to EVGA for replacement.

I've popped in a GT 730 I had lying around to test out the system in the meantime. Can't find any problems with the system, memory, etc. But here's the damned thing... my Firefox browser keeps crashing upon opening, even when I reload the OS from a backup. The DRAM is in quad channel, but when I drop down to 2 sticks in dual channel, the browser stops crashing. I've run Memtest86 on all sticks in all positions and they're all coming out OK.

Grrrr.

Gobe
04-10-2016, 11:18 PM
Image Editing:
275235

/

270667



Video Encoding:
327436

/
329733




OpenCL:
218148

/
106594



Heavy Multi-tasking:
281382

/
272734




Score:
231920

/
223613





Menthol's best run on the left versus my best run on the right. Some minor setting differences but mainly 4.95 vs 4.90 GHz (1% difference)

Setting aside OpenCL since we know that dual 980 ti is going to trounce dual 960, the rest of the scores are awfully tight and give or take a small amount, probably represent as far as the 5960X can go on water.

It looks like every 20,000 points added to the OpenCL score adds about 1,000 to the total score, and every 1,000 points added to any of the other scores adds about 250 points to the total score.

cekim
04-10-2016, 11:18 PM
That should read, "Suck it DUAL Xeon".

Seriously, nice score there. Glad to know somebody else has the image editing sussed out now.

I can push up to 5 GHz and put in some terrifying image editing scores, but there's not a prayer that I could run the more CPU intensive benchmarks without freezing things up. Until some smart-ass puts in a DICE or LN2 score, the 270's probably represents balls to the wall on the 5960X.

I'm still doing the post-mortem on what happened to my system yesterday. Both GTX-960s are bad. The one that's lights out won't fire up the monitor and the other errors out and crashes apps. One is on it's way back to Amazon for refund and the other is on it's way to EVGA for replacement.

I've popped in a GT 730 I had lying around to test out the system in the meantime. Can't find any problems with the system, memory, etc. But here's the damned thing... my Firefox browser keeps crashing upon opening, even when I reload the OS from a backup. The DRAM is in quad channel, but when I drop down to 2 sticks in dual channel, the browser stops crashing. I've run Memtest86 on all sticks in all positions and they're all coming out OK.

Grrrr.
Yuck, that blows - sounds like a power surge of some sort.

Gobe
04-10-2016, 11:25 PM
Maybe it was. At any rate, I've got some work ahead of me.

My previous comment about how scoring works was meant to be an approximation, but I'll be damned if it doesn't give EXACT total scores.

Score = ((Editing + H.264 + Multi-Tasking)/4) + (OpenCL/20)

Has anyone figured this out before or do I get a prize? :cool:

cekim
04-10-2016, 11:28 PM
On the 5GHz, I am not sure I am ready to provide the voltage required beyond 1.35 to do it until some Broadwell-E stuff comes out and looks worthy... So, I am in the same boat. I can run the easy stuff at 5GHz, but it will surely crash on the encoding.

Gobe
04-10-2016, 11:41 PM
On the 5GHz, I am not sure I am ready to provide the voltage required beyond 1.35 to do it until some Broadwell-E stuff comes out and looks worthy... So, I am in the same boat. I can run the easy stuff at 5GHz, but it will surely crash on the encoding.

Hmmm... Doesn't Intel XTU allow you to run specific apps at higher or lower clocks on the fly depending on the app? If so, unless the Intel XTU overhead is too high, there might be something there. Speed up the CPU for Image Editing and drop it back down for H.264 and Multi-Tasking.

Menthol
04-11-2016, 12:16 AM
There is another way to do that, OC Panel and a laptop

Qwinn
04-11-2016, 01:20 AM
Image Editing:
275235

/

270667



Video Encoding:
327436

/
329733




OpenCL:
218148

/
106594



Heavy Multi-tasking:
281382

/
272734




Score:
231920

/
223613





Menthol's best run on the left versus my best run on the right. Some minor setting differences but mainly 4.95 vs 4.90 GHz (1% difference)

Setting aside OpenCL since we know that dual 980 ti is going to trounce dual 960, the rest of the scores are awfully tight and give or take a small amount, probably represent as far as the 5960X can go on water.

It looks like every 20,000 points added to the OpenCL score adds about 1,000 to the total score, and every 1,000 points added to any of the other scores adds about 250 points to the total score.


And here's my best scores:

Image editing: 211629
Encoding: 232520
OpenCL: 178776
Multi: 210015
Total: 172479


So, as usual, for some strange reason I continue to get crappy scores (relative to my equipment) in Realbench for no reason I can figure out. I also have dual 980Ti's, O/C'd to 1455Mhz core 8000Mhz memory in SLI. Why is mine a full 40k lower than Menthol's? That's like a 20% difference, can't see how that can be explained by a few bins higher core clock OC on the GPU's.

This is running at 4.5 core 3.8 cache on the CPU btw (5930k). Oddly, the total score only improves by about 4,000 over my score at 4.4Ghz, where I can get 168,204.

I did also run in diagnostic mode, set realtime priority on the two Realbench processes, ran it at 800x600. Windows power plan High Performance. I miss anything?

Arne Saknussemm
04-11-2016, 07:21 AM
You install Open CL on CPU? ;) https://software.intel.com/en-us/articles/opencl-drivers

Menthol
04-11-2016, 12:02 PM
With the open cl on CPU drivers that Arne linked it will also use the CPU so then the CPU frequency and extra 2 cores will make a big difference

cekim
04-11-2016, 08:19 PM
With the open cl on CPU drivers that Arne linked it will also use the CPU so then the CPU frequency and extra 2 cores will make a big difference

So, clearly intel needs to give us 32 wchore unlocked processors with a 500W TDP so we can clock them to 4-4.5GHz and run as much water as required over them to pull that heat out. ;-)

Gobe
04-12-2016, 12:59 PM
And they need to sell them de-lidded out of the box. Who else is tired of seeing a 15 degree spread on the core temps?


So, clearly intel needs to give us 32 wchore unlocked processors with a 500W TDP so we can clock them to 4-4.5GHz and run as much water as required over them to pull that heat out. ;-)

Arne Saknussemm
04-12-2016, 09:01 PM
Yeah....delided with no guarantee for cracked silicon simple...

Very lucky with the last few chips I've had...very even temp spread....I remember a 3930K that had 11 degrees difference....was a low voltage special in the main but that one core really held it back....

Here really P'd off today....got my Ripjaws V kit...one DIMM is dead! Waiting now for replacement...the three sticks that work, work great at low vccsa and low vdimm.....looking forward to fully functioning kit! Should tune even better!

Amazon let me down for once...packet looked kind of like it had been opened before...I think this was a return sent out as new! Complained and they sent out replacement already...and free return....how customer service should be.

Gobe
04-13-2016, 09:31 AM
950 Pro... Ripjaws V...

Arne's system is slowly turning into my system. Though now that I'm on RVE, I guess my system is turning into his system!

You got the XMP 3200/14-14-14-34 kit?

Arne Saknussemm
04-13-2016, 09:40 AM
LOL

Yeah...well, I will have when the replacement arrives....runs 3200 14-14-14-28-1T at 1.35v and 0.928vccsa....will work on the rest when it runs quad....:(

By the way..WTF is up with XMP profiles these days?...setting all sorts of s**t they have no right to....taking over multiplier and cache?!?! setting vccsa over 0.3v higher than necessary! Manual only way to go...

drop4205
04-13-2016, 11:33 AM
How is the ripjaw 5 ram as its z170 it says compared to the ripjaw 4 that you previously had. BTW you message box is full it says lol

Arne Saknussemm
04-13-2016, 11:57 AM
Well, it looks promising but until I have 4 channels working can't say for sure...will post up when I have it running properly and can say if it's worth it...

Inbox sorted...thanks for heads up!:)

cekim
04-13-2016, 02:57 PM
How is the ripjaw 5 ram as its z170 it says compared to the ripjaw 4 that you previously had. BTW you message box is full it says lol
Only tangentially related, but my trident 3400 32G Z170 kit worked great at 2800 i x99 like every other kit I have... ;-)

I really must be missing a knob somewhere, I can run my 59xx at 37.5GHz with nothing more than my breath blown lightly over the lid of the chip to cool it, but 2801MHz, is just too much to ask of my ram. ;-)

Gobe
04-13-2016, 07:04 PM
My Ripjaws V 3200 kit poops out around 3400 MHz. No amount of volts or loose timings helps to get it faster (but I'll readily admit that I'm a novice at DDR4 overclocking).

Gobe
04-19-2016, 10:55 PM
I'm doing the same stuff on my recently added RVE that I was doing on my X99 series boards, and the best I can do on image editing at 4.9 GHz is 256K. I was getting over 270K on the X99 series boards at 4.9 GHz.

Aside from switching to dual GTX-970 from dual GTX-960 and the RVE, everything is pretty much the same. Weird. So, I pick up 2,700 points on my total score from the graphic cards but I lose 3,500 points on image editing... because, "!@#$ you Gobe..."

Maybe I'll slap a GTX-960 back in to see what happens because you never know.

Arne Saknussemm
04-20-2016, 04:48 PM
Interesting! :eek:

Gobe
04-20-2016, 07:14 PM
You say interesting, I say frustrating. I've dragged image editing up from 256K to 263K, but I'm giving up 7K - 8K versus my old number. That's 1,750 points off my total, or the difference between chasing that Xeon or thumbing my nose at it. I'll get it figured out and put that Xeon in it's place, but I really don't think I'm going to catch Menthol. That CPU of his is just too bad-ass.

cekim
04-20-2016, 10:28 PM
You say interesting, I say frustrating. I've dragged image editing up from 256K to 263K, but I'm giving up 7K - 8K versus my old number. That's 1,750 points off my total, or the difference between chasing that Xeon or thumbing my nose at it. I'll get it figured out and put that Xeon in it's place, but I really don't think I'm going to catch Menthol. That CPU of his is just too bad-ass.
You just need about $6-10K to build a dual 10/12/14/18/22 core v3 or v4 board. Can't beat 'em join 'em :D

Gobe
04-20-2016, 11:36 PM
I think I just recently spent about that much building three systems worth of X99 goodness and blowing up about one system worth of components. Because, "!@#$ you, hardware!"

BTW... note my custom title. I have you to thank for that.

cekim
04-21-2016, 03:38 AM
I think I just recently spent about that much building three systems worth of X99 goodness and blowing up about one system worth of components. Because, "!@#$ you, hardware!"

BTW... note my custom title. I have you to thank for that.
I noticed that and chuckled...

drop4205
04-23-2016, 12:59 AM
When running realbench i have just been starting windows normally and running it. So i need to get open cl and run in diagnostic mode for a better score?

Menthol
04-23-2016, 01:35 AM
drop4204,
Yes start in diagnostic mode, lower resolution to the lowest possible, run image editing once then run all, save score then reboot in normal mode to submit your score

cekim
05-02-2016, 03:09 PM
Diag mode, warm cache, 3200 DDR and I am at ~266K for 4.8GHz. I think the last little bit is nvme and raw GHz?

Curiously, my 5930K and 5960x ended up with nearly identical scores on this setup. Same win7 SSD (not m.2) in both, same video cards. Slightly different ddr, 5930k was actually slower - 3200/CAS16/1T vs 3200/CAS15/1T.

5930k 266248
5960x 265594

Wasn't willing to push the voltage I need for 4.9. I stopped at 1.399, it couldn't finish encoding.