|
United Kingdom20318 Posts
Heya, i can edit some more stuff into OP later~
I wanted to make a thread to discuss haswell overclocking, performance, temperatures, etc. In particular, the weird really high temperatures with AVX instructions enabled. I spent some time putting some numbers together:
ALWAYS THE SAME:
Gigabyte z87x-ud3h
4.5ghz core and uncore
2000mhz 10-10-10-28 RAM
1.2v uncore
1.22vcore, +0.15 analog io, +0.15 digital io, +0.1 system agent, default llc < volts need some work
~20c ambients open case because exhaust fan broken, xigmatek utgard, 120mm front intake
170mm top exhaust, silver arrow SE two ty-147 fans @1200rpm
6 minutes IBT temp, 4gb RAM:
HT on, avx on - 132 gflops - Max temps 92, 93, 91, 85*****
HT on, avx off - 67.2 gflops - Max temps 73, 74, 72, 66
HT off, avx on - 129 gflops - Max temps 83, 84, 82, 79
HT off, avx off - 67.6 gflops - Max temps 67, 67, 67, 61
x264 Benchmark HD v5.0.1
HT on, avx on - Pass 1; 88.12fps - Pass 2; 19.65fps - Max temps 71, 71, 70, 66
HT on, avx off - Pass 1; 88.08fps - Pass 2; 19.36fps - Max temps 70, 71, 70, 64
HT off, avx on - Pass 1; 70.98fps - Pass 2; 17.30fps - Max temps 63, 63, 62, 58
HT off, avx off - Pass 1; 67.39fps - Pass 2; 17.29fps - Max temps 63, 63, 62, 60 ???*
Starcraft 2: HOTS, maxed physics and effects, reflections on. Med shaders etc. 5 min replay segment, x4 speed
HT on, avx on - 2013-06-13 13:32:37 - SC2
Frames: 4022 - Time: 53290ms - Avg: 75.474 - Min: 35 - Max: 114
HT off, avx off - 2013-06-13 13:58:14 - SC2
Frames: 4032 - Time: 53368ms - Avg: 75.551 - Min: 35 - Max: 113
Note; CPU temps ridiculously low when playing sc2. Three cores dipping below 40c - one around 50-53c with ht AND avx.
No gain from HT and lack of load on other 3 cores = cool as shit fully loaded core
*Ambients varying within reason; no temperature control
*****STOP 0x0101 on third run, possible thermal failure?
Discussion here; It seems pretty ridiculous to me, and shocking that most people are tuning their overclocks around stress tests such as IBT and Prime with AVX enabled. That has been cause for much misinformation and confusion about Haswell's overclocking performance; It actually seems pretty damn good. A bit hotter than Ivy perhaps, but people are definitely doing more on lower vcores. I'll be tuning OC around the most stressful loads i expect to be doing; Not around IBT with AVX, which somehow pushes my OC, which sits in the high 60's with CPU at 100% load from x264 encoding, into the 90's and to the point of crashing system with silver arrow on 1.22v.
The sc2 benchmark i used was this replay: http://drop.sc/342789
TGeuSohel vs SeleCT: Medium shaders, maxed physics and effects, reflections on, SeleCT PoV (after selecting an orbital command) played 15 minute to 20 minute at x4 speed (after playing through a few times to remove stutters) and ingame performance is kinda silly.
|
Pretty good work! I'll update my post with some results (although I don't have all night to test like Cyro  ) but my results seem pretty similar. I'm testing to get 4.4Ghz stable, I'll post my benchmarks there and see if I can go for 4.5Ghz sometime soon! Thanks for your work Cyro o:
Also for anybody building a new PC, we can leave detailed 'guides' on how we tuned ours I guess. Not so much guides as saying what settings we changed, since they're different for each motherboard. I have an asrock, and I think cyro has a gigabyte. Just need an Asus
|
United Kingdom20318 Posts
I'm just getting started on tuning. My uncore was set completely blindly, i didn't experiment with digital/analog io or system agent properly yet, and because of some weird behaivour i have hopes that i can get better clocks on lower vcores with them. My vcore is really excessive right now, i added a ton to make sure i wouldn't fail during benching. I've passed IBT max ram with a lot lower. RAM also; i'l try to push to 2400 cas12, since jinglehell's old blog showed scaling with frequency but not timings.
FPS Shown in Min/Max/Average
2.8Ghz CPU Clock
1066 CL9 FPS: 28/133/61.6--95%
1066 CL6 FPS: 28/134/62.6--96.6%
1333 CL9 FPS: 30/139/64.8--100%
1333 CL6 FPS: 30/139/65.2--100.6%
1600 CL9 FPS: 31/142/67.2--103.7%
1600 CL6 FPS: 31/142/67.3--103.8%
|
Bunch of questions time.
Played with Base Clocks yet? If so any difference at all?
45 x 100?
36 x 125?
27 x 166?
Any testing uncore?
Are you delidded?
Hm can't think of more questions at this time.
|
United Kingdom20318 Posts
Bunch of questions time.
Played with Base Clocks yet? If so any difference at all?
45 x 100?
36 x 125?
27 x 166?
I tried 27x167 quickly, but it didn't seem to give better results initially (maybe voltages have to be tuned differently) and there's some issues with changing bclk straps.. RAM multipliers are weird and i couldn't set it how i wanted IIRC, and you have to cut power and mess around in bios for it to accept the bclk strap setting and boot properly, like it would stick on 167mhz even if you set 100x45 in bios and fail to boot over and over again until i just cut power for a few minutes, and i heard that's a common problem
Any testing uncore?
It didn't seem to make much difference in a few cases, i'l try it with sc2.
Are you delidded?
No, maybe in the future, i bought CLU and i was going to at least do some testing, get down overclock first, but i'm having some doubts now - who wants to hit a $430 CPU with a hammer? I could replace it, but then i'd have to face the silicon lottery again (reshuffling the deck when you had a decent hand) and i wouldnt be able to get a graphics card. I'm pretty damn sure that i wouldn't mess it up, but there's always a chance for something silly to go wrong, and that would really suck.
(:
|
I've tried like 40x 112.5 (something like that, it was 4.5Ghz) and had a bunch of problems, so I just went back to normal. I've heard some people having better success at 38x1.25, but I'm going to start with normal stuff and then go back to BCLK changing whenever I'm done finding it with just multipliers.
It seems like the old stress tests are performing terribly for haswell. It wouldbe interesting if they behaved more like x264. Then again, I'm nit an expert so maybe that's expected.
|
United Kingdom20318 Posts
If anybody has a sandy or ivy bridge CPU (oc'd) and a bit of time to spare, it would be VERY interesting to see temperatures with and without avx enabled for IBT. It's pretty easy to do
And Alryk, that's a terrible bclk to use ;p
Your straps are 100mhz, 125, 166.6. You should stick VERY close to those. I wouldn't expect 112.5 bclk to do anything at all
|
Could you grab a recent build of x264 and redo the benches with encoding some test file (@crf whatever)? Just make sure that you're not getting I/O bottlenecked.
Not that there are huge gains from AVX2 (so far), considering where most of the code time actually is and all of the supporting operations (see links here)
http://forum.doom9.org/showthread.php?t=168037
So that means that likewise, temps shouldn't be all that much higher when using it, I think.
|
On June 13 2013 22:57 Cyro wrote:No, maybe in the future, i bought CLU and i was going to at least do some testing, get down overclock first, but i'm having some doubts now - who wants to hit a $430 CPU with a hammer? I could replace it, but then i'd have to face the silicon lottery again (reshuffling the deck when you had a decent hand) and i wouldnt be able to get a graphics card. I'm pretty damn sure that i wouldn't mess it up, but there's always a chance for something silly to go wrong, and that would really suck. (:
You can use the LiquidUltra you bought instead of your current thermal paste. It should do a few degrees C of improvement. You'll get some stains reportedly, but it's supposedly fixed through metal polish and half an hour of rubbing. You'll also only have to do that polishing if you care about a stain, or perhaps when selling the CPU.
Don't use the sponge you get in the CLU package for anything! It's a metal sponge and will scratch stuff easily. It will destroy the printing on the CPU and the mirror finish of the heat-sink's base. I think it's only in the package because it's what's best to clean old LiquidPro.
|
United Kingdom20318 Posts
I bought it to use under the IHS, not above, i have mx-4 and didn't really want to deal it above IHS unless gains were big. I probably wouldn't do it unless maybe if i delidded, for extra temp drop, and because CPU is "messed up" anyway
Thanks for the info
|
I also was pretty interested in what happens on Ivy Bridge with the temperatures and AVX and IBT, so I checked what happens with my i5.
I'm sitting at an i5-3570k at 4.6 GHz, offset VCore set to result in about 1.255 V, LLC "High" which is a tiny drop, C1E and EIST both on and Windows power profile is "balanced", RAM at 2000 MHz.
The numbers I see for temperature with AVX on or off are not very different compared to what you've shown for HyperThreading off.
3570k @ 4.6 GHz:
AVX on: 125 GFlops -- max temp 75, 84, 82, 80 C
AVX off: 67 GFlops -- max temp 67, 73, 71, 71 C
Here's what you showed in the OP:
On June 13 2013 22:16 Cyro wrote:
6 minutes IBT temp, 4gb RAM:
HT off, avx on - 129 gflops - Max temps 83, 84, 82, 79
HT off, avx off - 67.6 gflops - Max temps 67, 67, 67, 61
I turned AVX off and on with "bcdedit /set xsavedisable 1" and "bcdedit /deletevalue xsavedisable" (and rebooting).
EDIT: screenshots:
AVX on: http://i.imgur.com/kqyZnIi.png
AVX off: http://i.imgur.com/o2M1KFz.png
|
On June 13 2013 23:08 Cyro wrote:
If anybody has a sandy or ivy bridge CPU (oc'd) and a bit of time to spare, it would be VERY interesting to see temperatures with and without avx enabled for IBT. It's pretty easy to do
And Alryk, that's a terrible bclk to use ;p
Your straps are 100mhz, 125, 166.6. You should stick VERY close to those. I wouldn't expect 112.5 bclk to do anything at all
Blame the OCN post xD but that explains it. At least I don't pretend to be an expert 
|
United Kingdom20318 Posts
On June 14 2013 00:42 Ropid wrote:I also was pretty interested in what happens on Ivy Bridge with the temperatures and AVX and IBT, so I checked what happens with my i5. I'm sitting at an i5-3570k at 4.6 GHz, offset VCore set to result in about 1.255 V, LLC "High" which is a tiny drop, C1E and EIST both on and Windows power profile is "balanced", RAM at 2000 MHz. The numbers I see for temperature with AVX on or off are not very different compared to what you've shown for HyperThreading off. 3570k @ 4.6 GHz: AVX on: 125 GFlops -- max temp 75, 84, 82, 80 C AVX off: 67 GFlops -- max temp 67, 73, 71, 71 C Here's what you showed in the OP: Show nested quote +On June 13 2013 22:16 Cyro wrote:
6 minutes IBT temp, 4gb RAM:
HT off, avx on - 129 gflops - Max temps 83, 84, 82, 79
HT off, avx off - 67.6 gflops - Max temps 67, 67, 67, 61
I turned AVX off and on with "bcdedit /set xsavedisable 1" and "bcdedit /deletevalue xsavedisable" (and rebooting). EDIT: screenshots: AVX on: http://i.imgur.com/kqyZnIi.pngAVX off: http://i.imgur.com/o2M1KFz.png
Thanks a ton. Wanted to make sure i wasn't losing my mind on avx temperature gap but it seems like there's some funny buisness going on with Haswell*
*Edit: Maybe this is wrong to say? 17c instead of 11c gap - but the complete lack of being able to reproduce this gap (which is 20c with ht on) makes it a bad idea to tune around it i think
|
On June 13 2013 22:43 Cyro wrote:
2.8Ghz CPU Clock
1333 CL9 FPS: 30/139/64.8--100%
1333 CL6 FPS: 30/139/65.2--100.6%
1600 CL9 FPS: 31/142/67.2--103.7%
1600 CL6 FPS: 31/142/67.3--103.8%
The difference between 1600 at CL9 and CL6 really is small. I have no idea: how often does the CPU query RAM during SC 2? The “total latency” of 1600 at CL9 is higher by a factor of 1.5, when compared to 1600 at CL6. Shouldn’t this add up more clearly?
|
United Kingdom20318 Posts
I dunno, im a RAM newbie (:
I remembered that blog from years ago though
|
Those latencies are just to kick off the data transfer, I think. The transfer itself might be what's taking most time overall.
|
Will you make a list of FPS numbers at different memory settings, Cyro? That would be neat. Haswell might be very different compared to the CPU used in that old blog.
|
On June 14 2013 01:35 Ropid wrote:
Those latencies are just to kick off the data transfer, I think. The transfer itself might be what's taking most time overall.
Sounds plausible.
|
On June 14 2013 01:21 blueslobster wrote:Show nested quote +On June 13 2013 22:43 Cyro wrote:
2.8Ghz CPU Clock
1333 CL9 FPS: 30/139/64.8--100%
1333 CL6 FPS: 30/139/65.2--100.6%
1600 CL9 FPS: 31/142/67.2--103.7%
1600 CL6 FPS: 31/142/67.3--103.8% The difference between 1600 at CL9 and CL6 really is small. I have no idea: how often does the CPU query RAM during SC 2? The “total latency” of 1600 at CL9 is higher by a factor of 1.5, when compared to 1600 at CL6. Shouldn’t this add up more clearly?
It depends on the task. In gaming, you usually don't need a short latency, as the bottlenecks are nearly always on the GPU and CPU. What may limit is data loading, which is usually kept to 4K+ chunks, usually higher when talking about texture loading. Keep in mind that approximately 98% of the RAM read commands given to the CPU end up reaching the cache, usually L1 and L2, and that the memory controller aggressively pulls bigger pages from the RAM into the cache to avoid this anyways, the latency is less critical for anything that doesn't require random access to big data sets (> 8MB). Good game engine design makes sure that at level loading, the most heavily used data will be cached on the CPU anyways, as this is predictable information. Beyond that there are some very clever hardware tricks in modern CPUs to semi-cache information (TLB, predictive memory controller, etc), which can help keep the latency down to nano-seconds even with CL12.
There is the "memory wall", a phenomena caused by a slower rate of increase in memory speeds vs. CPU speeds over approximately three decades, so modern CPU design does a lot to work around that, and specifically, latency for most common application usage.
tl;dr: for gaming, memory latency doesn't really matter.
|
On June 14 2013 00:42 Ropid wrote:I also was pretty interested in what happens on Ivy Bridge with the temperatures and AVX and IBT, so I checked what happens with my i5. I'm sitting at an i5-3570k at 4.6 GHz, offset VCore set to result in about 1.255 V, LLC "High" which is a tiny drop, C1E and EIST both on and Windows power profile is "balanced", RAM at 2000 MHz. The numbers I see for temperature with AVX on or off are not very different compared to what you've shown for HyperThreading off. 3570k @ 4.6 GHz: AVX on: 125 GFlops -- max temp 75, 84, 82, 80 C AVX off: 67 GFlops -- max temp 67, 73, 71, 71 C Here's what you showed in the OP: Show nested quote +On June 13 2013 22:16 Cyro wrote:
6 minutes IBT temp, 4gb RAM:
HT off, avx on - 129 gflops - Max temps 83, 84, 82, 79
HT off, avx off - 67.6 gflops - Max temps 67, 67, 67, 61
I turned AVX off and on with "bcdedit /set xsavedisable 1" and "bcdedit /deletevalue xsavedisable" (and rebooting). EDIT: screenshots: AVX on: http://i.imgur.com/kqyZnIi.pngAVX off: http://i.imgur.com/o2M1KFz.png
I'm getting similar temps and I'm running 4.7ghz on a 3570k with the same voltage I figure until I get my phase change cooling its the best I'll do with a h60 in push/pull config T_T
|
|
|
|
|
|
EPT
