Haswell Preview Benchmark - Page 24

Prime95 · 2013-09-02, 04:47

Quote:

Originally Posted by kracker

On 28.1 with one thread, times went from 10 ms to 8 ms.

Sadly, memory bandwidth prevents us from enjoying this gain when running 4 workers.

On another note, I successfully downloaded LinX and it drives temperatures a good 10 degrees higher than prime95's small FFT torture test.

LaurV · 2013-09-02, 10:27

Wow! This means that is still room from improvement in P95!
About (75/65-1)*100=15.38%

Waiting for the improved version...

TheJudger · 2013-09-03, 16:18

Quote:

Originally Posted by LaurV

Wow! This means that is still room from improvement in P95!
About (75/65-1)*100=15.38%

Waiting for the improved version...

75 = ?
65 = ?

Oliver

ewmayer · 2013-09-03, 21:31

Quote:

Originally Posted by Prime95

Sadly, memory bandwidth prevents us from enjoying this gain when running 4 workers.

Around the same time I e-mailed George thusly:

Quote:

I've been meaning to ask: Are you OCing both cpu and ram, or just the former? If also the ram, do you use a preset OC config or a custom? [e.g. do you need to up the ram voltages and such].

I'm thinking those big alu heatfins on the ddr3 2400 ram we bought are just screaming "overclock me!", and that is something I might be able to do even with cpu at stock.

I've been quite busy the past month writing code aimed at boosting || performance of Mlucas and have not wanted to interrupt my workflow for mem-OC experiments, but if I'm not mistaken George may soon have some good news on that front.

Prime95 · 2013-09-03, 22:01

Quote:

Originally Posted by ewmayer

if I'm not mistaken George may soon have some good news on that front.

My good news was premature. Last night I went into the BIOS and selected DDR3-2600 and DDR3-2800 and the machine booted and torture tested OK. Sadly, this morning I benchmarked prime95 and it was slower!!

Digging deeper, the mobo chooses a 266Mhz memory clock for DDR3-2400, but a 200MHz memory clock for DDR3-2600 and DDR3-2800. With timings for 2400 a nice tight 10-12-12-31. Timings for 2600 were a slow 11-13-13-33 (I assume this is clock counts for the slower 200 MHz memory bus).

Next I tried DDR3-2666. While it isn't stable, it does use the 266 MHz memory clock and timings of 11-13-13-34. Surely this will be faster in prime95, right? Wrong. This too is 15% slower than DDR3-2400.

Clearly I don't understand memory overclocking. Can anyone explain to me what might be going on?

ewmayer · 2013-09-03, 22:19

Ah, I inferred - incorrectly - from your e-mail yesterday that your preliminary timing/torture tests had indicated a performance gain, which you wanted to make sure still left the system running stably on a 24/7 basis. My new-system experiments in ratcheting up the RAM clocking from 1600 to 2400 also made it sound like the only issue in pushing past the rated 2400 would be stability under load. Sorry about the overoptimism, but it really sounded to me like we were gonna get a nice little freebie here.

(And perhaps we still shall, once we get help from the OC experts).

LaurV · 2013-09-04, 03:00

[offtopic]

Quote:

Originally Posted by TheJudger

75 = ?
65 = ?

that was just a joke: he said that the cpu is 10 degree hotter, and I assumed that the temperature goes from 65C to 75C, something which is "in his favour", but you can use any numbers, inlcuding 3 and 13

, which would mean other tester uses the cpu "better", as is squeezes more heat from it. The conclusion is that p95 can be improved... for the case when his initial temperature was 3C, and it is now 13C, well... p95 is very inefficient program, and it has to be improved (13/3-1)*100=333%

[/offtopic]

kladner · 2013-09-04, 03:02

FWIW- this chart gives some idea of the trade-offs between frequency and latency.
http://en.wikipedia.org/wiki/DDR3_SDRAM#Modules

I'm not sure it really addresses your issues, though.

axn · 2013-09-04, 03:06

Quote:

Originally Posted by LaurV

[offtopic]

that was just a joke: he said that the cpu is 10 degree hotter, and I assumed that the temperature goes from 65C to 75C, something which is "in his favour", but you can use any numbers, inlcuding 3 and 13

, which would mean other tester uses the cpu "better", as is squeezes more heat from it. The conclusion is that p95 can be improved... for the case when his initial temperature was 3C, and it is now 13C, well... p95 is very inefficient program, and it has to be improved (13/3-1)*100=333%

[/offtopic]

Physics fail! Everyone knows you must use Kelvins when taking ratios.

LaurV · 2013-09-04, 03:18

Quote:

Originally Posted by axn

Physics fail! Everyone knows you must use Kelvins when taking ratios.

Where have you been? I was missing you ..

TheMawn · 2013-09-04, 03:44

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.

I've been looking a good bit into this and sadly I seem to know less and less the more I read. I was hoping to just add my knowledge to the pool here but if I keep going I won't have anything useful to say.

2013-09-02, 10:27	#255
LaurV Romulan Interpreter Jun 2011 Thailand 2⁶×151 Posts	Wow! This means that is still room from improvement in P95! About (75/65-1)100=15.38% Waiting for the improved version... Last fiddled with by LaurV on 2013-09-02 at 10:28*

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

2013-09-03, 22:19	#259
ewmayer ∂²ω=0 Sep 2002 República de California 19·613 Posts	Ah, I inferred - incorrectly - from your e-mail yesterday that your preliminary timing/torture tests had indicated a performance gain, which you wanted to make sure still left the system running stably on a 24/7 basis. My new-system experiments in ratcheting up the RAM clocking from 1600 to 2400 also made it sound like the only issue in pushing past the rated 2400 would be stability under load. Sorry about the overoptimism, but it really sounded to me like we were gonna get a nice little freebie here. (And perhaps we still shall, once we get help from the OC experts).

2013-09-04, 03:02	#261
kladner "Kieren" Jul 2011 In My Own Galaxy! 2×3×1,693 Posts	FWIW- this chart gives some idea of the trade-offs between frequency and latency. http://en.wikipedia.org/wiki/DDR3_SDRAM#Modules I'm not sure it really addresses your issues, though.

2013-09-04, 03:44	#264
TheMawn May 2013 East. Always East. 6BF₁₆ Posts	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. I've been looking a good bit into this and sadly I seem to know less and less the more I read. I was hoping to just add my knowledge to the pool here but if I keep going I won't have anything useful to say.