• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

Ram speed and timings

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

Tbrooks

Member
Joined
Jun 23, 2006
Location
Regina,SK
I just started out test ram speed right now and I will check out timings later.

First off I had a 3930K a while ago and it was a
3930k @ 4.7Ghz
4 x 4GB Ram at 1866Mhz
9-10-9-31 Patriot Viper 3 Ram
It ran
Measured floating point speed 4385.91 million ops/sec
Measured integer speed 15450.43 million ops/sec

I sure miss that rig it was about 9000 - 10000 Rac it had all the 32 core servers beat but couldn't quite keep up to the 64 cores .

New one is a 3930K @4.7Ghz
2 X 8GB Ram at 2400MHZ
10-12-12-31 Gskill Trident
and it runs
Measured floating point speed 4426.42 million ops/sec
Measured integer speed 14851.56 million ops/sec

So here I am. My floating is winning but Integer is falling behind.

I will try out 4 x 4GB at 2600Ghz
10-12-12-31 Gskill Trident
Tomorrow and let You know.

Then I will tighten timiings and see f that helps.


Thanks all for checking this out and maybe not having to spend big bucks on Ram.

Sorry My Sig is so wrong.

That was my first 3930K

This second one is 3930K
X79 Deluxe
2 R9 290
16Gig Gskill 2400 Ram (for now)
1050 watt Seasonic X series
and well still lot of beer but when it's this cold whiskey is good too
 
Last edited by a moderator:
id figure those quad-ram channels do help, faster access to ram, even if ram itself is slower
 
Sorry I tried out that 2600Mhz Ram but the system didn't like it at 4.7Ghz. Then I had some plumbing issues at the house here so I haven't had time to tinker with it.
 
TL;DR - any progress on testing settings for this? Also sad about Adak :(

sorry to come out of the blue to bump this, but I'm curious if you ever had a chance to test some different settings with rosetta.

I'd also not obsess too much over what RAC is reported by bakerlabs. For example if you check here the top 3 machines:

http://boincstats.com/en/stats/14/host/list/

The Credit/Week and Credit/Month are the real figures you should be concerned with. Unfortunately I don't know of a way to "reconstruct" those figures for your old machine.

But in this example, note the machine in 3rd place has a higher RAC than the 1st place - and it's always been that way. But note the weekly/monthly totals - that's the real story. (disclaimer: I am the guy with W3670, which is why this has driven me nuts for the last 2 years. Also how the #$@# do you get a i7-995 but I digress)

So don't live and die by RAC alone is what I'm saying. The RAC calculation is pretty flakey. In fact if you dig around BOINCstats you'll see they have their own "BS-RAC" they calculate specifically because of this.

Another thing - Rosetta is a "mix" of different work, with for lack of a better term "sub-projects" that they send out units for (different research projects/Grad Assistants) all under the umbrella of Rosetta@Home. So some weeks/months (this is key) it seems the work is harder/doesn't credit as much vs. others. So you'll have months were you wonder why you lost 1000 RAC for no reason. Best to see how other known machines are doing.

But sorry I ramble. My necrobump is because it would be interesting to find out what system settings are more important to rosetta. Integer most definitely, since AMD Bulldozer/Piledriver seems to perform very well from the scraped together "research" I can get from trolling stats. Which since AMD archs have the shared FPU between INT 'modules' vs. Intel.

Before I pass out (... RIP Adak, drink another for you) things like chipset timings (like, chipset/uncore internal, not memory), uncore settings/speed (is there a way to crank uncore mhz on socket 2011?) etc I think might matter quite a bit more than people let on.

At least with the socket 1366 procs I've used, the "internet wisdom" is that cranking uncore doesn't really matter that much - however it seems to make a pretty tangible difference to Rosetta for me at least.

Ugh, okay I'm incoherent at this point, and it's taken me an hour to post this. So someone else please carry on this thread lol :facepalm:

edit add: I don't know if Solar is still around, but he had quite a few 2011 rigs - maybe ask him about any pointers for tweaking?

also heh, this guy awesome profile and shows the franken-rig http://www.stonepics.com/folding/ but assuming he still runs the clocks he says (compare int/float numbers to see) maybe you can use it as a reference machine. But only if he's 100% Rosetta - he used to be but i didn't check
 
Last edited:
Sorry I've had a few myself tonight.

First off it is really hard to find out if a hardware change has helped in Your performance on Rosetta as it takes over a month for Your RAC to reflect it.

I will try to answer more of Your questions tomorrow when I sober up.
 
TL;DR - any progress on testing settings for this? Also sad about Adak :(

sorry to come out of the blue to bump this, but I'm curious if you ever had a chance to test some different settings with rosetta.

I'd also not obsess too much over what RAC is reported by bakerlabs. For example if you check here the top 3 machines:

http://boincstats.com/en/stats/14/host/list/

The Credit/Week and Credit/Month are the real figures you should be concerned with. Unfortunately I don't know of a way to "reconstruct" those figures for your old machine.

But in this example, note the machine in 3rd place has a higher RAC than the 1st place - and it's always been that way. But note the weekly/monthly totals - that's the real story. (disclaimer: I am the guy with W3670, which is why this has driven me nuts for the last 2 years. Also how the #$@# do you get a i7-995 but I digress)

So don't live and die by RAC alone is what I'm saying. The RAC calculation is pretty flakey. In fact if you dig around BOINCstats you'll see they have their own "BS-RAC" they calculate specifically because of this.

Another thing - Rosetta is a "mix" of different work, with for lack of a better term "sub-projects" that they send out units for (different research projects/Grad Assistants) all under the umbrella of Rosetta@Home. So some weeks/months (this is key) it seems the work is harder/doesn't credit as much vs. others. So you'll have months were you wonder why you lost 1000 RAC for no reason. Best to see how other known machines are doing.

But sorry I ramble. My necrobump is because it would be interesting to find out what system settings are more important to rosetta. Integer most definitely, since AMD Bulldozer/Piledriver seems to perform very well from the scraped together "research" I can get from trolling stats. Which since AMD archs have the shared FPU between INT 'modules' vs. Intel.

Before I pass out (... RIP Adak, drink another for you) things like chipset timings (like, chipset/uncore internal, not memory), uncore settings/speed (is there a way to crank uncore mhz on socket 2011?) etc I think might matter quite a bit more than people let on.

At least with the socket 1366 procs I've used, the "internet wisdom" is that cranking uncore doesn't really matter that much - however it seems to make a pretty tangible difference to Rosetta for me at least.

Ugh, okay I'm incoherent at this point, and it's taken me an hour to post this. So someone else please carry on this thread lol :facepalm:

edit add: I don't know if Solar is still around, but he had quite a few 2011 rigs - maybe ask him about any pointers for tweaking?

also heh, this guy awesome profile and shows the franken-rig http://www.stonepics.com/folding/ but assuming he still runs the clocks he says (compare int/float numbers to see) maybe you can use it as a reference machine. But only if he's 100% Rosetta - he used to be but i didn't check


Where to start?

Ok. I came to realize that the architecture of the 3930K will not run 2600 ram at that speed so I can't post those results. If only I had a 4930K that should be able to run those speeds I would try.

Dustin has had that 995 OC'd to the nuts for a long time.
His float is 3230.53
Int. 12060.41


And yours is.
Float 3985.26
Int. 12917.79

Both are 32nm and running SSE4.2. That makes no sense. Do you use yours for anything? He might just let his crunch 24\7.

Next.
I think BS-RAC does have their own calculation for RAC. But it includes RAC from all of the other BOINC projects You are running.

Next.
Yes Rosetta does have different projects with different points. That is the reason for the RAC system. So it evens everything out. Kind-of. It is really hard for them to keep everything even and equal. In order for them to do that they would have to have one of every computer out there to set a 'lets say" a benchmark.I know points matter to us but I am here for the cause more than that.

Next.
I am trying to find out what system settings do to help in Rosetta. I started with Ram and was going by Rosetta's measuered #'s. But as You have shown may not prove to be accurate in the PPD a machine puts out.I am open to any other way to test hardware out. Any other suggestions? I'm all ears.
AMD definitely kicks butt on the server side. They can run 64 cores. It seems from my findings that the more cores and the faster those cores run the better PPD you get. But on the home PC side Intel is the way to go minus the price. Ram speed or trimmings don't seem to help. I have 2 4770K's. Both at 4.5Mhz with way different Ram but one is on Linux. I will switch both to Windows today and let them run for a month just to see if 1866 or 2600 helps. Just for You bud( na I want to see also). I didn't want to screw with that linux rig cause it just plain runs happy with no problems at all.

Next.
I won't even try only running one core. As I said before the more cores and the faster they are,core on seems to be the best way to go. Is there a way to run only one core on Lga 2011?


Next.
Yes I would also like to hear from Solar. He has not been around in a while.

Last.
You mentioned Adak a couple of times. Yes I miss him also. I folded for him for almost 2 weeks, He is missed by me also.

I think that is all. Any questions Plz ask.
 
I'm also wondering why you don't have a team with a kick *** rig like that? I was thinking you would make a good team mate on overclockers.com
 
I'm also wondering why you don't have a team with a kick *** rig like that? I was thinking you would make a good team mate on overclockers.com

Quite true. Not only do we appreciate the extra power, but also those who seem to have a deeper knowledge of the project and its inner workings. It's always very interesting to read these types of threads, and it never ceases to amaze me how much I can learn.
 
Where to start?

Ok. I came to realize that the architecture of the 3930K will not run 2600 ram at that speed so I can't post those results. If only I had a 4930K that should be able to run those speeds I would try.

I think that is all. Any questions Plz ask.

I think for SB-E the max memory speed is ~2400MHz and from what I've seen or heard... Not all chip samples can do that.

You should try running your 4x4Gb 2600 @2400 10-12-12-31 and compare it to your 2x8Gb kit?... If your 2600 kit has double sided ic you should be able to show better performance vs. a 2x8GB or even a 4x4GB kit with single sided ic.
 
There is no 2600 ratio in s2011. Next step is 2666 but it won't work at x1.00 strap so you have to set 125bclk/1.25 strap. Also many Samsungs ( generally most of IC in 2133-2400 TridentX ) are working fine @2600 ~1.65-1.75V but have issues with booting or stability @2666.
I just finished testing Patriot 4x8GB 2400 @2666-3100 on 4930K and I can only say that 2400 or 2600 performance isn't much different while anything above 2600 will probably give worse results due to much more relaxed sub timings.
I think that optimal will be ~2400 9-11-11 /10-11-11 CR1 and tight sub timings.
 
argh! sorry for the lack of response. I swear I set the forum settings to send me email notices of replies. I need to check my spam filter stuff.

I've got to run now, but I did want to respond quick so I'm not just a one shot wonder on a forum, plus I'm extremely worried in re-reading this in retrospect that I made some sort of e-peen post, which I'm very much not.

I haven't had a chance to fully digest the posts yet. I just checked it and was horrified I didn't reply since the 14th.

quick glance version Q and A.

About joining overclockers.com.
I always feel bad posting cause yea I'm not on overclockers.com Rosetta Team. Long story short - I'm the team founder of a small rickshaw team named Tenebrosol. Don't join it, it is the scattered crumbs of me and 3 friends after the main team we joined long ago basically didn't care about Rosetta at all. If I had known about overclockers.com team back when I was cut free, I would have totally joined here. Now I'm in the sucky limbo - my buddies probably don't really care if I jumped, but since I'm the team founder... yea that probably would come off bad. They'd probably just quit. Which sucks cause since I really believe in the science of Rosetta. In a dreamland BOINC would have sub-teams, and i'd totally ally with overclockers.com now. No joke. Again I wish I had found overclockers.com first before founding a new team. I only don't cause... team founder jumping is kinda a jerk move =/

Is the machine listed a dedicated cruncher?
No no NO, this is my daily beater rig. Steam gaming, iddling around internet, email etc. It's totally not dedicated. The case has an old file cabinet handle drilled into it for LAN parties (rare as those are these days) so yes it's the daily driver. 4393.10 Mhz, NB freq 3820.1, 764 mem 8-8-8-24-86-1T Gigabyte X58A-UD3R v 1.1 (? - it's not the 2.0) bck wall (?) 191. I only put specs cause this ain't some super rig, or hand groomed anything. It was what was decent/cheap but overclockable rep in April 2010? (started with i7-930, went and got ebay xeon years later to try and get a few more miles - well after i about fried the poor 970 you can read in these forums).

THIS IS NOT A BRAG, like... I'm old school overclocker. New Intel platforms with intel "blessed OC K" platform make me angry at heart, if only cause.. I'm a dinosaur. Yea i get it, overclocking as a hobby is going away with mobile fetishists. I condemn, but I do so knowing I'm backward. I don't even have a mobile phone so yea, my opinions mean nothing bla bla. sorry I ramble a lot. On that also why my big interest in AMD (see my avatar pic, that K6-III yea still runs) because bulldozer arch seems to do well in rosetta. sorry :blah:

I've got to run. Super summary version: when my 1366 rig blows up finally, do I replace it or quit? I'm not the moneyhouse computer guy, I just like tweaking around with old stuff (now) and making it the "most it can be."

also, yea socket 2011 rigs should curb stomp 1366. I say that with jest but the old days when you went from Socket A to say, 939 Toledo damnit, that was exciting fun. I guess yea I'm old school.

gotta run, sorry that's incoherent. I'll make a more specific response when I'm not 10min late from blabbering typing this lol.

But yea in the end, my rambling incoherent goal for posting here is, say what makes Rosetta fly on ANY machine? What "hot buttons" does Rosie like? Random example say it likes integer a lot - how do we pimp integer performance on various rigs, other normal benchmarks probably would even suffer?

Real world application: like, if the rumors of the haswell Pentium K sku are true, is there any chance to make a cheap rosie rig out of it? Or dieing AM3+ socket, while generally panned in general perforance, maybe 8-9XXX series FX are sleeper rosetta monster rigs?

And of course, simply joining overclockers.com rosetta team will increase your RAC by 300 AUTOMATICALLY :bday: (shhhhhh, don't tell em!)

must run
:salute:
 
Last edited:
Where to start?

Ok. I came to realize that the architecture of the 3930K will not run 2600 ram at that speed so I can't post those results. If only I had a 4930K that should be able to run those speeds I would try.
Sorry I wasn't meaning explicitly more testing with memory speeds. Of course my ham-fisted post was... rambling approaching pathetic - though that's probably because I... hadn't even thought of the... ordeal? needed to start a regimented Rosetta testing study either. And thinking of that, it definitely would take awhile. 2 months at least per setting change while. Wowzers.

Which BTW thanks for being awesome and kicking this as a discussion thread instead of branding me as an idiot (which re-reading my post, I'd fully deserve!). The end game is to try and figure out how Rosie ticks... which yea since it's multi-subproject nature makes that a rather fuzzy goal, but so far there do seem to be trends of things that Rosetta likes "in general," which maybe can be optimized for.

But the thing that got me to unlurk and post. Okay, assuming you are disappointed about your quad-channel results, and you "only" can run "3930k @ 4.7Ghz 4 x 4GB Ram at 1866Mhz 9-10-9-31" that should... no IS kicking the doors in on my old, weak-sauce, 2 gen older, underclocked tri-channel 1526Mhz 8-8-8-24-86-1T.

Like... that shouldn't, that can't, wha? Looser timings aside, you have way more bandwidth. I can't even dream of 4.7 CPU as a daily number, and your memory speeds are totally out of my league w/o frying my uncore. For all intents and purposes a socket 1366 rig should not even be close.

So I can only guess that maybe bandwidth isn't a bottleneck after a certain point for Rosetta? But that doesn't explain the over 300Mhz increase in core speed, let alone those cores being what? 20%? more IPC per clock? (IvyBridge-E is 2 gens newer than gulftown?) Something is fishy. Maybe Rosetta doesn't optimize for those improvements, but still even if not... 300 Mhz! Plus I got the impression software didn't have to "optimize toward" the majority of the improvements? So no Haswell AVX2 boon for example.

My gut reaction is uncore, but like, any uncore after 1366 is way more badass right? I mean SandyBridge could pull more raw performance from "just" dual channel than BloomField ever could with tri-channel.

Plus didn't SandyBridge end the whole "uncore ratio" stuff? Isn't uncore 1:1 with CPU afterward? That includes cache too right?

But, footnote here for later... I wonder in order to do that, if "IPC" for the uncore, or latency or something changed that afflicted Rosetta speed somehow.

I'd love to see a 5Ghz 24/7 stable SandyBridge box, to get a reference. What would be really interesting is say okay, If 5Ghz Sandy, 4.7 Ivy, 4.5 Haswell if despite the IPC improvements, showed a decrease trend for Rosetta.

But that's all just conjecture. I wish I had more "newer" rigs to test this.

Dustin has had that 995 OC'd to the nuts for a long time.
His float is 3230.53
Int. 12060.41

And yours is.
Float 3985.26
Int. 12917.79

The only thing I can think of here is I kept the ambiguous Gigabyte named "Performance Enhance" memory setting at "Extreme," which maybe gooses mysterious uncore (former chipset NB) timings? The other thing maybe my uncore speed is higher than what I generally see people using for 24/7 (though that sample size is kinda... small)? Gulftown is evil in that uncore scales beautifully with voltage, but sings you the sweet siren song of volts=Mhz, laughing at you with a linear scale until it suddenly smokes. My gulftown uncore is 3820 with an uncomfortable to me "1.415v (cpu 1.45625v - also overspec). Everything I've seen online warns over 1.35v VTT is tempting fate on Gulftown, and over 1.45v is degradation to death for 24/7.

I guess the joke/truth is things always work until they fail.

I'd love to know what speeds/setting's he/she is using on the 995X though. If only cause the 995X is a mythical chip to begin with =) (unlocked, everything, Intel employee only? "last of the socket" awesome?) but the reverse is true, if they still run the rig, maybe they join the team with some uncore tweaks =)

Both are 32nm and running SSE4.2. That makes no sense. Do you use yours for anything? He might just let his crunch 24\7.

No, this is just my daily beater rig. Play games, surf, my main rig. It's not dedicated - it's actually what I'm typing this reply with right now.

Next.
I think BS-RAC does have their own calculation for RAC. But it includes RAC from all of the other BOINC projects You are running.

Yea I know they have their own calculation. But since I'm a one project guy I dunno about the multi-project blend influence for a given user's stats on a single project? I never realized that. I do know they (and many other) stats sites do have a big emphasis on multiproject stats, which... is good if your multiproject and about the points (no judgement).

Heh, ironically... I know nobody ever see/reads the team descriptions, but yea. "The stats are fun, but the science is what keeps you in for the long haul."

Next.
Yes Rosetta does have different projects with different points. That is the reason for the RAC system. So it evens everything out. Kind-of. It is really hard for them to keep everything even and equal. In order for them to do that they would have to have one of every computer out there to set a 'lets say" a benchmark.I know points matter to us but I am here for the cause more than that.
Heh. "official RAC" is... an interesting thing. There was a time (pre 2006-2007) with the whole "optimized BOINC clients" era, where basically recompiling the BOINC executable with various optimization flags would drastically effect the benchmark of the client. This in itself, okay - but various projects would calculate credit with a metric that included this number. To give a very real example, this resulted in the Conroe Core 2 Duo E6600 I had at the time suddenly giving me 5000 RAC for a good few months.

In a vacuum this was harmless, (well aside from the optimized client arms race) but cross-project it made a mess. And since Baker Labs didn't give a response to this for nearly a year after many other projects "fixed" their RAC calculations (very many users directly asked Baker Labs if this was legit/bad) it became the de-facto way to run Rosetta.

There's more to the story, but for this post the take away is... RAC is um... a relative thing. With the phase of the moon. On good days =)

Next.
I am trying to find out what system settings do to help in Rosetta. I started with Ram and was going by Rosetta's measuered #'s. But as You have shown may not prove to be accurate in the PPD a machine puts out.I am open to any other way to test hardware out. Any other suggestions? I'm all ears.
AMD definitely kicks butt on the server side. They can run 64 cores. It seems from my findings that the more cores and the faster those cores run the better PPD you get. But on the home PC side Intel is the way to go minus the price. Ram speed or trimmings don't seem to help. I have 2 4770K's. Both at 4.5Mhz with way different Ram but one is on Linux. I will switch both to Windows today and let them run for a month just to see if 1866 or 2600 helps. Just for You bud( na I want to see also). I didn't want to screw with that linux rig cause it just plain runs happy with no problems at all.
Haha we all are! But seriously you hit the nail on the head here. I'm a Rosetta freak, and finding info about tweaking rosetta is... difficult. It's not as popular as Folding, and if you're "into the science" many default to WCG which is cool. But in the end it makes it hard to find "rosetta tweaking" advice!

Yes, as I beat around the bush until it's caught fire from my cyclic action, lets do it! (says me, who yea, when I start testing combinations that suck means I lose RAC for 2 months but damnit, it's for the better good - I'll suck it up if it means more science gets done in the end).

Also AMD. I'll be honest. I just want any excuse to build an AMD rig lol. When I even saw a hint that Bulldozer/Piledriver/whatever had unexpectedly good Rosetta performance, I latched onto that and ignored everything else =) Fanboi yes. But it also points to a very interesting thing - AMD is "strong/competitive" in integer and apparently Rosetta rewards them for it.

Why?... we must find out.

Next.
I won't even try only running one core. As I said before the more cores and the faster they are,core on seems to be the best way to go. Is there a way to run only one core on Lga 2011?
I... didn't mean to imply this. I don't know what I said to imply this. My bad. Yea single core... on the Socket 754 rigs I had, maybe I could get 350 RAC but that's old school numbers. MOAR CORES =D
Next.
Yes I would also like to hear from Solar. He has not been around in a while.
Him and Pixie.
Last.
You mentioned Adak a couple of times. Yes I miss him also. I folded for him for almost 2 weeks, He is missed by me also.
Sorry yea, I said this like 4 times because... beverage. Re-reading, ugh.
I think that is all. Any questions Plz ask.
Sorry to all for huge posting, twice in a row even. D'oh.
 
Just give me a day or two to re-read those two posts a few times and I will try to respond.
 
Just give me a day or two to re-read those two posts a few times and I will try to respond.

just popping in to say, NO pressure. My own reply rate is... glacial to say the least.

10 days or 10 months it's rosetta forum... it's all good...

;)
 
So basically after making a fool of myself, and 4 months later I've discovered there's nothing really special about my socket 1366 setup ;p

so what I got:

Intel(R) Xeon(R) CPU W3670 @ 3.20GHz (actual 4393 NB 3820 Mem 764)
day 3,456
week 53,751
month 245,468

Then I get the currently common surplus xeon for $100 with low mult:
Intel(R) Xeon(R) CPU X5650 @ 2.67GHz (actual 4202 NB 3629 Mem 764)
day 3,301
week 51,500
month 234,181

So while that might be 10,000 + more credit if you pound your gulftown northbridge to the edge of overvoltage frying it, it doesn't really seem worth it since the "now cheap and cheery x5650 - nobody wants because at stock it sucks" proc is nipping at it's heels.

Lastly since socket 1366 boards seems to be stupid expensive used (WTH???) I think any research into this now is pretty much pointless. Dumping $200+ into a mobo for a dead platform is a fool's errand.

Anyone have an i7-4790K @ 4.8Ghz stable for 24/7 rosetta yet? Judging from one of my buddy's 3570K pulling 4500 RAC a day with 4 cores (dunno his clock speed - doubt it's over 4.7 though) 8K should be within reach for the +1gen haswell + HT.

Which begs the question again - what is with socket 2011 chips? Maybe people cross-project those rigs more, so they always look weak in Rosetta? Or maybe simply I need to dig more in rosetta-stats land to find one that isn't putting out lame numbers.

But anyways, thanks for reading. And sorry for the mess... :D
 
Back