PDA

View Full Version : GTX-960 Boat-anchor?



Gobe
03-01-2016, 05:12 PM
I've been optimizing and benching my system w/Intel i7 5820K @ 4.5 gHz, and I've gotten these numbers:

245,039
232,808
50,817
209,619

For a total score of 174,407

This is with a GTX-960 graphic card.
How might the 50,817 score and the 174,407 total score improve with a GTX-980 or GTX-980 Ti?
[Edit: or am I mistaken in assuming the graphic card is being benched?]

Also, I'm running 32 Gb of DDR4 at 3,000 MHz per the XMP profile via 16x2 Gb in dual channel. While the raw bandwidth of running quad channel is significantly better than dual channel, I don't expect I'd see much real world advantage over the dual channel setup I'm running. Does anyone see any reason I should revisit this assumption?

mirzet1976
03-01-2016, 07:39 PM
I don't now about RAM but GPU score gives little to overall score see this 1xGPU+16bgRAM VS 3xGPU+32gbRAM http://rog.asus.com/realbench/show_comment.php?id=11979&compare=11608

Arne Saknussemm
03-01-2016, 08:27 PM
980 would be 75000-80000...980Ti...90000ish.....

You can run that part of the test on the CPU as well...check tweaks and tips section...increase score a bit...

Gobe
03-07-2016, 12:33 AM
980 would be 75000-80000...980Ti...90000ish.....

You can run that part of the test on the CPU as well...check tweaks and tips section...increase score a bit...

Or move over to my 5960X system... increase score a bit. Lol.

cekim
03-07-2016, 12:51 AM
Or move over to my 5960X system... increase score a bit. Lol.
Yep, if you show all users on the leaderboard, you'll see your 5960x/x99-Pro right there with mine and Menthol's RVE. I have 2 heavily OC'd GTX980's for that run and barely edge out your 960. This benchmark is heavily skewed toward other things.

Interesting that at least with SLI - the openCL is largely bound by memory throughput. I see a 1:1 increase with DDR4 speed there. Given all the bulk copies that go on, that isn't too surprising, but it does confirm what I'd seen in my compute usage that DDR4 speed and x16 PCIe matter where they don't much to games.

My 5960x hits a wall at 4.8GHz. Timing, not temps stops me... 4.81 will crash pretty much no matter what I do. 4.6 is 24/7 stable at 1.275-1.28 4.75 is stable, but at 1.31-1.32 runs a bit warm for my usage (heavy loads for days at a time). 4.8 requires 1.37 and 85C. No go for 24/7 but so far hasn't crashed.

Gobe
03-07-2016, 02:34 AM
Yep, if you show all users on the leaderboard, you'll see your 5960x/x99-Pro right there with mine and Menthol's RVE. I have 2 heavily OC'd GTX980's for that run and barely edge out your 960. This benchmark is heavily skewed toward other things.

Interesting that at least with SLI - the openCL is largely bound by memory throughput. I see a 1:1 increase with DDR4 speed there. Given all the bulk copies that go on, that isn't too surprising, but it does confirm what I'd seen in my compute usage that DDR4 speed and x16 PCIe matter where they don't much to games.

My 5960x hits a wall at 4.8GHz. Timing, not temps stops me... 4.81 will crash pretty much no matter what I do. 4.6 is 24/7 stable at 1.275-1.28 4.75 is stable, but at 1.31-1.32 runs a bit warm for my usage (heavy loads for days at a time). 4.8 requires 1.37 and 85C. No go for 24/7 but so far hasn't crashed.

Is that 4.8 GHz at less than 1.3 volts via a pre-binned CPU or did you get lucky with a winning silicon lottery ticket? I can get 4.5 rock stable at 1.25V but it takes around 1.35 to stabilize 4.6. So I run routine at 4.5. Stabilizing 4.7 was doable, but it took stupid extremes to pull it off.

The tight clustering of the top 5960X scores says that these numbers represent about all that this CPU can do. Seems that some of us get a scoring edge in one aspect and might lag in another. Some of it is the silicon lottery and some of it is know-how. There's also a bit of luck and the good graces of the overclocking gods to be had. I've been OCing on and off for the past 15 years (longer if you count fooling with old i486 CPUs in the mid 90s via motherboard jumper clusters and painfully sparse documentation) and I'm finding the Haswell-E in combination with the X99 motherboards to be the most fun and versatile platform yet.

cekim
03-07-2016, 04:06 AM
Is that 4.8 GHz at less than 1.3 volts via a pre-binned CPU or did you get lucky with a winning silicon lottery ticket? I can get 4.5 rock stable at 1.25V but it takes around 1.35 to stabilize 4.6. So I run routine at 4.5. Stabilizing 4.7 was doable, but it took stupid extremes to pull it off.

Voltages are actually roughly:
4.6 1.275-1.280
4.7 1.30-1.32 (I didn't tune much, temps were uncomfortable for 24/7 100% load)
4.8 1.36-1.37

I put a premium on high cache and DDR as well since my intended use benefits greatly from them. So, all of the above are with a 4.4 cache at 1.27.

Not pre-binned - dumb luck and perhaps a little coin-toss probability. I have 2. The other tops out at 4.4 @1.290. In both cases, I haven't gone past 1.37, maybe there is more, but I'd need to do something more extreme than the EK mono-block and 420 rad. I am not doing ice or chilled water.



The tight clustering of the top 5960X scores says that these numbers represent about all that this CPU can do. Seems that some of us get a scoring edge in one aspect and might lag in another. Some of it is the silicon lottery and some of it is know-how. There's also a bit of luck and the good graces of the overclocking gods to be had. I've been OCing on and off for the past 15 years (longer if you count fooling with old i486 CPUs in the mid 90s via motherboard jumper clusters and painfully sparse documentation) and I'm finding the Haswell-E in combination with the X99 motherboards to be the most fun and versatile platform yet.
Agreed, though it also speaks to how the benchmark is scored.

Arne Saknussemm
03-07-2016, 09:51 AM
Or move over to my 5960X system... increase score a bit. Lol.

Yep! LOL

Your image edit score is quite a bit higher than mine....32GB memeory at 14 14 14 seems to run very well! I've got my eye on one of those Ripjaws V kits....

Though Menthol's score is pretty much same as mine on 32GB even tighter....so not sure if it's just down to memory...

Gobe
03-07-2016, 12:25 PM
Yep! LOL

Your image edit score is quite a bit higher than mine....32GB memeory at 14 14 14 seems to run very well! I've got my eye on one of those Ripjaws V kits....

Though Menthol's score is pretty much same as mine on 32GB even tighter....so not sure if it's just down to memory...

I just posted on this elsewhere, but storage could have a small but real impact here if the image editing benchmark is doing disk read/write. Samsung 950 Pro NVMe is crazy fast. I've got two of them at 512 GB each and my initial thought was to set them up in RAID 0 until I learned this can't be done on the X99-Pro/USB3.1. There are boards that will do this, so if anyone has the notion to boot from NVMe RAID0, do yourself a favor and make sure your motherboard allows this.

Arne Saknussemm
03-07-2016, 03:31 PM
Storage wasn't a factor before....we ran RB on ramdisks in the past and no gain...but maybe with changes?

Running image edit more than once has been a tweak from the beginning...second run scores way higher...

cekim
03-07-2016, 10:50 PM
Storage wasn't a factor before....we ran RB on ramdisks in the past and no gain...but maybe with changes?

Running image edit more than once has been a tweak from the beginning...second run scores way higher...

yep - clearly the file-cache is showing up in the gimp benchmark, which is to be expected. I am running those numbers posted on an 840 SSD, not the 951, so its only ~400MB/s. I expected I might gain a decent amount if that was addressed from what I've seen there.

Back to linux to get some work done and despite the pounding in windows this weekend without issue at 4.6/4.7/4.8, I still had to dial it back to 4.5 this morning after some work in linux (it just shut-down when things got hairy). So, more tuning to do next weekend.

Actually, this new instability seems to have come with the new BIOS - the August X99 bios was more stable at these numbers ... hmmm... Maybe I lost some secret sauce when I upgraded the bios, but I sure thought I wrote down everything.

Gobe
03-07-2016, 11:37 PM
Has anyone bothered to enable only one core at a time on a 5960X and find the top frequency per core? I tried this today and aside from being a pain in the ass, it's gives a bit of insight into the silicon lottery.

First, with only one core running, the heat just doesn't build up, so you're getting a result that represents the individual core at the resulting much lower temperature.

I tested by setting 1.35 volts for Vcore and 1.95 volts for the CPU input. Individual cores ran around 57 - 62C. The test was simply to see if a core could complete 10 runs of the Real Bench H.264 video encoding benchmark and 10 runs of the Real Bench heavy multitasking benchmark without crapping the bed. The image editing and OpenCL benchmarks need not apply for this job.

Results: 2 cores were good at 4.9 GHz, 3 cores were good at 4.8 GHz, and 3 cores were good at 4.7 GHz.

0 = 4.7
1 = 4.7
2 = 4.8
3 = 4.9
4 = 4.8
5 = 4.9
6 = 4.7
7 = 4.8

Lest you think that you can just set up the multipliers using "per core" and get the best out of each individual core, it doesn't work that way (or at least I sure don't know how to make it work that way!) As a test of these results, I enabled the 5 cores that were stable at 4.8 and 4.9 GHz (with the 3 lazy 4.7 GHz cores left disabled). The collective heat of 5 cores did not allow me to run stable at 4.8 GHz, but the 5 cores were easily stable at 4.7 GHz with nothing beyond my every day cooling.

Collectively, I can stabilize all 8 cores around 4.6 GHz, but I won't routinely run above 4.5 GHz since the power and heroic cooling requirements for stability at 4.6 GHz with all 8 cores pumping just aren't worth the extreme aggravation. With the collective heat of all 8 cores, "practical" stability was achieved not at the worst single core frequency of 4.7 GHz, but 0.2 GHz lower at 4.5 GHz since 8 cores together run a lot hotter than a single core. Of course, since the single cores ran around 60C, it might stand to reason that more extreme cooling could stabilize a processor at the frequency of the worst core(s).

I was a bit surprised by all of this. Going in, I had expected the 8 cores to cluster around two frequency bins, not spread out across three! With the 2 cores that could run at 4.9 GHz, I could even get partly through the encoding benchmark at 5.0 GHz but no amount of fooling around would stabilize the cores at 5.0 GHz.

If I had to guess, I would say that collectively, a CPU can be stabilized with reasonable voltage (1.35V) at around 0.1 GHz below the least stable individual core if you've got great cooling. For average cooling, take off 0.2 GHz. This suggests that 5960X CPUs that run stable at 4.7 GHz likely has all 4.9 GHz cores. That rarity that runs stable at 4.8 GHz probably also has individual 4.9 GHz cores and your cooling is outrageous. Takes a fine chunk of silicon top to bottom to pull that off!

I could have and probably should have (but didn't) check the stability of the 4.8 and 4.9 GHz cores below 1.35 volts.

Damn those three lazy cores!

cekim
03-08-2016, 12:04 AM
Not too surprising. I haven't played with that much - its a struggle just to mess around and get things stable again before needing to get some work done. 9am Monday morning black screens are not amusing ask me how I know today? ;-)

I upgraded to the latest bios this weekend and clearly I lost one of my 4.6 stable settings, because it isn't (but only in linux when I get an hour into something important). 4.5 to the rescue!

Gobe
03-08-2016, 12:08 AM
yea, I was supposed to fly to the west coast today, but my manager flaked and I ended up sitting around home today getting paid to do nothing. Left me with free time and a need to find a way to waste it.

cekim
03-08-2016, 12:12 AM
yea, I was supposed to fly to the west coast today, but my manager flaked and I ended up sitting around home today getting paid to do nothing. Left me with free time and a need to find a way to waste it.
"stand back... I'm a professional"... ;-)