Haswell Preview Benchmark - Page 25

Batalov · 2013-09-05, 19:29

Quote:

Originally Posted by ewmayer

That costs a smidge more per-byte than the ddr3 2400 George & I got ... is the issue that the 2400 is not available [or not as cheap per-byte] in a 4GB option?

Ernst, George, or anyone else,

did you have any problems with your 2400 memory or did it work by simply switching XMP profile? Did you have to up the "System agent voltage" (or VCCSA)? (Is this a virtual alias for the CPU-NB?)

I got the 2400MHz CL10 ddr3 yesterday and it is annoyingly unstable (just a pinch, i.e. an error on average once in a few hours, but that's enough to render it useless for our purposes, quite obviously). I am using XMP profile as is (and obviously it set the V to 1.65v, and so on: 10-12-12-31 etc...) and 4200MHz CPU Freq (it is a 3570K). All was rock solid for months with 4200MHz married to a 1866Mhz CL9 memory (also G.Skill's, just other color).

Any advice will be appreciated.

Prime95 · 2013-09-05, 19:36

Quote:

Originally Posted by Batalov

Did you have any problems with your 2400 memory or did it work by simply switching XMP profile? Did you have to up the "System agent voltage" (or VCCSA)? (Is this a virtual alias for the CPU-NB?)

Mine simply worked by switching to the XMP profile. My mobo won't let me play with System agent voltage. I am running at about 1.2V and maybe the mobo boosts VCCSA when the CPU voltage is bumped.

Batalov · 2013-09-05, 20:32

Ok, thanks. Looks like I'll have to do it manually: mine is 1.05v, I think. (the mobo is ASUS LK V, not the worst and not the shiniest. Something from the middle. F7, or maybe on your board, a Ctrl-F1 may unlock advanced settings.)

I do hope that I didn't get a dud, and just a bit of juice will cure it. I found similar 1.2v recommendations upon a cursory run over gu:gl.

The tests on 77000003 did show improvement on 4-thread throughput: ~+9% all things being equal compared to the 1866 memory, as expected (so it is like huffing'n'puffing as if at "4.6GHz" - but for free and with much less heat). That's the good news.
...But stability hit me from behind by morning, after a night of torture and other tests. Without stability, this is not a step forward, this is three steps back.

ldesnogu · 2013-09-05, 22:13

I just switched XMP profile in my BIOS, and CPU-Z confirmed my memory is running at 1200 MHz. I have no stability issues, though I didn't heavily stress the system (except for running LinX and Prime 95 for a few dozen of minutes).

My MB is an ASUS Z87-Pro and my RAM is this Kingston HyperX.

ewmayer · 2013-09-06, 05:24

Quote:

Originally Posted by Prime95

Mine simply worked by switching to the XMP profile. My mobo won't let me play with System agent voltage. I am running at about 1.2V and maybe the mobo boosts VCCSA when the CPU voltage is bumped.

Same for me; have same cpu/mobo/ram as George, only difference is he OCs his CPU and I run at stock.

The MSI mobo defaulted to 1600 Mhz when I first installed the faster RAM [my initial set of tests was Haswell plus the slower dddr3 1333 I had been using with my SB quad]; I got 10% more throughput from my Mlucas code running in parallel on all-4-cores from swapping in the G-SKIL RAM at 1600 MHz and another 17% from upping the ram speed to 2400 via xmp profile. Have not yet tried going above the rated 2400 MHz on the ram, especially now that George reported his poor results for that.

db597 · 2013-09-10, 02:59

How much more voltage do you find is required to keep things stable with 28.1?

Using the previous 27.x build:
4770K was rock solid with 4.2GHz @ 1.25V / Memory 2000MHz @ 1.65V.

After switching to 28.1 build:
Started to get BSODs a few times a day. To stablise, I raised the VCore to 1.265V, and downclocked the memory from 2000MHz to 1866MHz (still at 1.65V).

To get the memory back to 2000MHz, I'm not sure if it's more VCore I need or more voltage for the memory (it's a Corsair XMS3, rated for 2000MHz @ 1.65V).

Prime95 · 2013-09-10, 03:06

When I switched to 28.1, CPU and memory voltage increases (that I felt comfortable with) did not bring stability. I undid the voltage increases and downclocked from 4.2 to 4.1 GHz (leaving memory at 2400) and all was good again. This is a better option than reducing your memory speed.

TheMawn · 2013-09-10, 03:29

Quote:

Originally Posted by Prime95

When I switched to 28.1, CPU and memory voltage increases (that I felt comfortable with) did not bring stability. I undid the voltage increases and downclocked from 4.2 to 4.1 GHz (leaving memory at 2400) and all was good again. This is a better option than reducing your memory speed.

+1

In my particular system, 4.2 GHz accomplishes roughly the same as 4.6 GHz because of my memory speed limitations. If I didn't hate the software with a passion, I would use ASUS's AI Suite to downclock my CPU a few notches (and the voltage too, while I'm at it) while doing overnight / at school crunching to reduce temperatures, wear and tear, etc.

I still run 4.6 GHz basically for the same reasons anyone would.

If you're comfortable with having different profiles based on your usage of your computer, I would say sacrificing 200 MHz of RAM to get 200 MHz of CPU is probably worth it if you're doing basically anything else than GIMPS work.

VBCurtis · 2013-09-10, 04:19

Quote:

Originally Posted by TheMawn

Conventional wisdom, as far as I knew, was that higher frequency meant better bandwidth and lower timings and lower timings meant better latency, but I think that's a bunch of bogus.

The timings essentially represent how many "ticks" the module has to accomplish a certain part of the task of reading and writing memory and that the frequency defines "ticks per second". Any bandwidth calculation I have found does not seem to take into account timings, however, which has me just as confuzzled.

Bandwidth measures how much information it's possible to get from the memory. Latency does not matter, as we can issue a request every cycle and get information out every cycle- regardless of how many cycles lapse between an individual request and reply. If a program issues a continuous stream of memory requests, it will get a continuous stream of replies at nearly the theoretical bandwidth of the memory- merely delayed by {latency number} ticks.

My understanding of GIMPS is that it fully saturates the bandwidth, so running higher memory clocks with looser timings "should" help. George's results suggest that timings above 2400mhz do weird things with settings and fail to follow my intuition.

As an aside, programs that respond to memory latency will also respond well to hyperthreading, while those which respond to bandwidth will not. HT gives the illusion of cutting latency in half (as half the CPU cycles are spent on each of two programs, while memory runs at regular speed). This is why ECM is useful on HT, but extra copies of Prime95/LLR are not.

db597 · 2013-09-10, 05:00

I suppose it depends on whether the PC exclusively runs LL testing. For a dedicated LL cruncher, the 15% speed boost from FMA3 is worth 600MHz, so even reducing clock speed from 4.2GHz down to 3.6GHz is still worth it.

But if the machine is running other applications during its non-idle time, you might miss the 600MHz for those other times.

Interestingly, by default, Haswell voltage bumps up 0.1V when AVX/FMA3 is used. Practically it is Intel recommending more VCore to get stable operation in AVX/FMA3 mode. If we were happy with 1.25V earlier, it should be logical to use 1.35V now. Heat problems aside that is....

(If anyone finds this useful, my BSOD errors were all 0x101 and 0x124 - which according to several sources means insufficient VCore)

TheMawn · 2013-09-10, 06:01

A 0.1 volt jump is a pretty big one. A rep from Intel actually recommended about a 10% overvolt being a safe limit. Of course, that probably means 20% is fine, too, but it's different from the old "no comment" we're used to hearing.

Anyway, a 0.1 volt bump from stock is already pushing 10%, so I don't know if using your 10% increase and then having an additional 0.1 V increase is good or bad by their standards.

Any BIOS options to disable that?

2013-09-10, 02:59	#270
db597 Jan 2003 203₁₀ Posts	More voltage? How much more voltage do you find is required to keep things stable with 28.1? Using the previous 27.x build: 4770K was rock solid with 4.2GHz @ 1.25V / Memory 2000MHz @ 1.65V. After switching to 28.1 build: Started to get BSODs a few times a day. To stablise, I raised the VCore to 1.265V, and downclocked the memory from 2000MHz to 1866MHz (still at 1.65V). To get the memory back to 2000MHz, I'm not sure if it's more VCore I need or more voltage for the memory (it's a Corsair XMS3, rated for 2000MHz @ 1.65V).

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Haswell-E Prelim. Benchmark	sdbardwick	Hardware	37	2015-02-10 18:49
Prime95 and Haswell	Pleco	Information & Answers	22	2014-07-13 16:03
Haswell Rig	Mini-Geek	Hardware	64	2014-05-27 13:22
Prime95 version 27.1 early preview, not-even-close-to-beta release	Prime95	Software	126	2012-02-09 16:17
Missing mouse-over preview text	retina	Forum Feedback	1	2011-09-12 15:32

2013-09-05, 20:32	#267
Batalov "Serge" Mar 2008 Phi(4,2^7658614+1)/2 9497₁₀ Posts	Ok, thanks. Looks like I'll have to do it manually: mine is 1.05v, I think. (the mobo is ASUS LK V, not the worst and not the shiniest. Something from the middle. F7, or maybe on your board, a Ctrl-F1 may unlock advanced settings.) I do hope that I didn't get a dud, and just a bit of juice will cure it. I found similar 1.2v recommendations upon a cursory run over gu:gl. The tests on 77000003 did show improvement on 4-thread throughput: ~+9% all things being equal compared to the 1866 memory, as expected (so it is like huffing'n'puffing as if at "4.6GHz" - but for free and with much less heat). That's the good news. ...But stability hit me from behind by morning, after a night of torture and other tests. Without stability, this is not a step forward, this is three steps back.

2013-09-05, 22:13	#268
ldesnogu Jan 2008 France 2·5²·11 Posts	I just switched XMP profile in my BIOS, and CPU-Z confirmed my memory is running at 1200 MHz. I have no stability issues, though I didn't heavily stress the system (except for running LinX and Prime 95 for a few dozen of minutes). My MB is an ASUS Z87-Pro and my RAM is this Kingston HyperX.

2013-09-10, 03:06	#271
Prime95 P90 years forever! Aug 2002 Yeehaw, FL 1D77₁₆ Posts	When I switched to 28.1, CPU and memory voltage increases (that I felt comfortable with) did not bring stability. I undid the voltage increases and downclocked from 4.2 to 4.1 GHz (leaving memory at 2400) and all was good again. This is a better option than reducing your memory speed.

2013-09-10, 05:00	#274
db597 Jan 2003 7·29 Posts	I suppose it depends on whether the PC exclusively runs LL testing. For a dedicated LL cruncher, the 15% speed boost from FMA3 is worth 600MHz, so even reducing clock speed from 4.2GHz down to 3.6GHz is still worth it. But if the machine is running other applications during its non-idle time, you might miss the 600MHz for those other times. Interestingly, by default, Haswell voltage bumps up 0.1V when AVX/FMA3 is used. Practically it is Intel recommending more VCore to get stable operation in AVX/FMA3 mode. If we were happy with 1.25V earlier, it should be logical to use 1.35V now. Heat problems aside that is.... (If anyone finds this useful, my BSOD errors were all 0x101 and 0x124 - which according to several sources means insufficient VCore)

2013-09-10, 06:01	#275
TheMawn May 2013 East. Always East. 11·157 Posts	A 0.1 volt jump is a pretty big one. A rep from Intel actually recommended about a 10% overvolt being a safe limit. Of course, that probably means 20% is fine, too, but it's different from the old "no comment" we're used to hearing. Anyway, a 0.1 volt bump from stock is already pushing 10%, so I don't know if using your 10% increase and then having an additional 0.1 V increase is good or bad by their standards. Any BIOS options to disable that?