***Asus Rampage III Extreme Owners Thread***

[zoson]

9.1GB works well, no voltage increases needed anywhere, and pulling some nice GFLOPS The 77.xx's are due to me doing things, I was working at the same time, when i leave it alone the GFLOPS are very consistent on my system with 12GB installed.

78.3962 GFLOPS at 4005MHz | DDR3-1868MHz 7-8-7-20 1T | 3600MHz uncore

You'll probably get better performance with 183bclk and 1833mhz memory cyrptik.
22x cpu multiplier, 183bclk, 5x mem multiplier, 20x uclk multiplier will give you almost exactly the same speeds, but you'll have a uclk 2x your memory speed.

I'm also kind of curious why you're sitting at 4ghz with a 990x? You can probably do 4.4 without breaking a sweat!

[CryptiK]

Not sure about the high vs low BCLK thing, lower BCLK the system has less inherent latency than with higher BCLK's. Since these silly boards dont sit evenly on a given BCLK say I used 182 BCLK (~182.4 in reality) i'd be on ~4012MHz core, DDR3-1820MHz and 3640MHz uncore. Slightly higher CPU, 48 MHz less on the ram and 36 more on the uncore. I might get a little better performance. Thing is I hate uneven CPU/ram/uncore speeds and I try and make numbers that seem 'nice' and 'balanced'. That's part of the reason why its at 4005MHz. This CPU has no problem doing 4.5GHz @ 1.336v (see submission in 4.5GHz LinX club) but I dont need that much power, I probably wouldnt even notice. When I work it hard, renders/decoding/encoding/simulations etc, it takes so long anyway its a walk away come back in a few hours scenario and if it finished 15mins sooner I wouldnt be waiting for it so I see no need to push it and create more wear and tear on all the components. I also dont like running EIST etc so its full speed 24/7 and I'd rather not push it like that. With the 6 cores and 12GB ram it does everything I want it to at the current settings.

[zoson]

Not sure about the high vs low BCLK thing, lower BCLK the system has less inherent latency than with higher BCLK's. Since these silly boards dont sit evenly on a given BCLK say I used 182 BCLK (~182.4 in reality) i'd be on ~4012MHz core, DDR3-1820MHz and 3640MHz uncore. Slightly higher CPU, 48 MHz less on the ram and 36 more on the uncore. I might get a little better performance. Thing is I hate uneven CPU/ram/uncore speeds and I try and make numbers that seem 'nice' and 'balanced'. That's part of the reason why its at 4005MHz. This CPU has no problem doing 4.5GHz @ 1.336v (see submission in 4.5GHz LinX club) but I dont need that much power, I probably wouldnt even notice. When I work it hard, renders/decoding/encoding/simulations etc, it takes so long anyway its a walk away come back in a few hours scenario and if it finished 15mins sooner I wouldnt be waiting for it so I see no need to push it and create more wear and tear on all the components. I also dont like running EIST etc so its full speed 24/7 and I'd rather not push it like that. With the 6 cores and 12GB ram it does everything I want it to at the current settings.

You can't turn off EIST, as shown in your CPU-Tweaker shot. ASUS mislabels EIST in BIOS and it's really the turbo disable. With R3E, it's impossible to stop your cpu from dropping the multiplier to save power.

The point to going to 183bclk is you wouldn't be bottlenecking your memory. As it stands, sure your ram runs at 1866, but your UCLK is bottlenecking it to 1800mhz worth of performance. So as far as internal 'balance' of your cpu... 183bclk makes a lot more sense. If you jumped to the 23 multiplier and 183bclk(183.5 on my board) you get an even 4220MHz... What makes 4005 any more even/balanced than 4220?

Also, you're an overclocker! We don't do it because we'll notice!

[donmarkoni]

Also, you're an overclocker! We don't do it because we'll notice!

Yeah! We do it because we can!

BTW, there is a temperature drop expected here, so I'm looking forward to some action.
I'll just have to post in a different thread.

[zoson]

Yeah! We do it because we can!

BTW, there is a temperature drop expected here, so I'm looking forward to some action.
I'll just have to post in a different thread.

Maybe we should make a high performance club thread? Like the 4.5ghz club but for low speed 100gflops, highest ht gflops, and highest non ht gflops? break it up by architecture, so there would be three categories for each cpu type?

[CryptiK]

The uncore portion provides it's own performance gains from higher MHz (uncore = everything else except the core) but there is certainly a performance gain to be had going from DDR3-1600MHz to DDR3-1868MHz at the same uncore of 3604MHz. It won't 'bottleneck the ram to DDR3-1800MHz' as you put it, but there is an interaction between the IMC portion of the analogue controller and the ram speed.

Also, my R3E does not drop to low multi at no load. I can watch the C-state and its in 100% speed mode all the time with a lowest calculated multiplier of 29.997x when 30x is set in BIOS.

I'm keen for an efficiency test, to do it right and to standardize for any CPU speed we'd need to calculate the GFLOPS/MHz (we can of course also have brackets of results as you suggested). What could be fun is doing something like the low clock SPI32M thread, where there is categories of say 100GLFOPS, 90GFLOPS, 80FLOPS etc and the lowest speed required to hit as close to but not over that mark, ie: the 70GFLOPS bracket would have an ideal target of 79.99 GFLOPS etc.

In absolute terms, for my 4GHz run I just did, I get 78.3962/4005.1 = 0.01957 GFLOPS/MHz, or to make it more meaningful, 19.57 GFLOPS/GHz.

Looking at the data from the last 2 pages, here's the absolute efficiency rankings:

• mr too short - 92.1319/4610.5 = 0.01998 = 19.98 GFLOPS/GHz

• donmarkoni - 91.9528/4610.8 = 0.01994 = 19.94 GFLOPS/GHz

• zoson - 91.5017/4657.9 = 0.019644 = 19.64 GFLOPS/GHz

• CryptiK - 78.3962/4005.1 = 0.01957 = 19.57 GFLOPS/GHz

• Spoolindsm - 86.6040/4504.9 = 0.019224 = 19.22 GFLOPS/GHz

[zoson]

The IMC is part of the uncore. The IMC bandwidth is directly related to uncore speed.

When Nehalem came out, Intel mandated 2x uncore to memory speed because unless uncore is 2x, there is not enough bandwidth to the IMC to support the full speed of your memory.

Intel relaxed this requirement to 1.5x on gulftown because hardware limitations made it impossible to reach high enough uncore speeds to accommodate the fastest memory out there. To fully utilize your memory without bottlenecking it, you must have 2x uncore. When you go to say 2.1x uncore, the performance gains drop off very quickly when memory is a primary component.

So essentially, if your uncore is less than 2x, you are bottlenecked. Of course you would see a performance gain from 1600mhz to 1866 with a 3.6ghz uncore. It's because you have extra bandwidth above 1600mhz.

That's why i get better performance at 2000mhz memory with 4ghz uncore and 4.6ghz than i do with 2176mhz memory, 3.95ghz uncore and 4.665ghz.

[spoolindsm127]

I think my R3E killed my 980X

[CryptiK]

WTF man. What happened?

Zoson - yeah I guess we are using the term bottlenecked slightly differently but I agree with you. I guess in my case though I'm about at the limit of what I'm comfortable running 24/7 at 1.250 Vtt (load), so 3800MHz uncore isn't going to be an acceptable setting for me as it will require more voltage.

[spoolindsm127]

Dude I'm so PO'd you have no idea. Months and thousands spent on finding this batch and this POS chip just DIES.

You know what's even funnier? I never pushed Vcore past 1.35v. QPI never went above 1.35v also. Except when I had QPI LLC disabled and QPI set to 1.38125v. I turned the system off and switched the jumper to enable QPI LLC. Turned it back on so I guess QPI got set to 1.38v actual. I set it down to 1.35v and restart, and the system keeps restarting after POST. Set Bios to defaults and on restart, Vcore was being set to 1.14375v manually by the Bios??

I'm thinking it was a combination of dieing motherboard that took the cpu with it. man why always me...

[CryptiK]

Tried it in another board? Maybe the boards just acting funny - I hope the CPU's ok mate. If you reset BIOS it could be setting 1.14375v as a lower power state.

If it died at that voltage without some kind of spike then I'd be RMAing that at the least you can sell whatever they give you back if its not a nice batch.

[CryptiK]

OK this is a bit odd. Thought my setup was at the max for the voltages it was running, but after thinking about what zoson said I thought I'll just do a quick test. Now I played around with different BCLK and multi combos, different ram etc, but thought instead of re-establishing what everything needs for different speeds lets just crank up the uncore to 3733MHz (2 x ram) and leave everything else where it was. Only thing I did different was re-install my corsair ram fan since it was getting hot...........

Result. Appears fully stable at 3733 MHz uncore @ ~1.257 Vtt running 12GB (6x2GB) 1866C7's at DDR3-1868MHz 7-8-7-20 1T..........I wasn't expecting that, as I left Vtt where it for 3600MHz uncore and the same ram settings. I obviously had a ram temp issue affecting things so I'm going to re-investigate ram timings/vdimm and Vtt requirements now the ram is properly cooled. Bandwidth now has increased a few hundred MB/s and latency is a little lower as expected.

Everest at new settings:



LinX 5 run @ 9100MB at new settings (sorry guys I'm the new king of efficiency :P) :




Highest GFLOPS = 80.3288

80.3288/4005 = 0.02005 GFLOPS/MHz = 20.06 GFLOPS/GHz

[spoolindsm127]

Nice clocks there Cryptik . Interesting to see the role of Uncore in efficiency. Congrats though now you've made yourself the target lol

So we're going with only 5 runs of LinX to determine efficiency? You guys know right how sometimes if the OC is unstable you'll get HIGHER GFlops in LinX than you would when it's actually stable?

I love the affects of thermal dynamics though. So basically you're ram was just getting too hot? I'm sure better fans on the AirFlow module would definitely help. Reminds me of back when I was on air cooling and I'd wait 20-30 minutes before each run to let the mem cool off ha. That was the only reason I went with the EK Dominator block. Although it's restrictive for the system's overall flow rate, it does always keep the DIMMs cool. I took mine off because I didn't want to restrict flow. But now I'm thinking maybe I should put it back in.

About the 980X though, God knows what happened. I tried it in the OC as well, no go. I'm pretty sure that thing is dead from a voltage spike or something similar. But like you said at least I can RMA it and get my money back. Mann I wanted to get in on this GFlops competition . Regardless of what batch I get though, I'm gonna overclock the life out of it

[zoson]

Yeah, I think it should be similar rules to the 4.5ghz club... 20 loops... and we need to standardize ram size because higher problem size = higher gflops.

Cryptik, try running with 5120, and see what kind of performance you get. Not everyone is going to have 12gb mem unfortunately.

[CryptiK]

Yeah I was thinking about how many loops we'd need to standardize it to, but since this isn't really a stability club (is it?) less should be sufficient so long as it's consistent. You can see in my run it's got about 0.0357 GFLOPS variance between the highest and lowest loops. Also 20 loops with 12GB of ram will take me until next year to complete, perhaps that's your grand plan! I dunno, say 10 loops? That should be more than enough to establish efficiency and basic stability, although I must say I've always got lower GFLOPS when slightly unstable.

I will try running with less as a test, but I shouldn't be excluded for running 12GB We can always have installed ram as a delimiter and have 6/12GB categories.

Sorry to hear about the 980X mate, hope you get a good one in return. Intel are very prompt in their turn around time for a claim and this is one of those situations where it cannot be determined exactly what happened but it certainly wasn't abuse on your behalf.

Edit - I actually only run the fans on the airflow module at 60% duty cycle they're way too loud at full speed, and the ram actually doesnt need much airflow to stay cool. Its just warm to the touch now under full load and the fans are basically silent.

[zoson]

Yeah I was thinking about how many loops we'd need to standardize it to, but since this isn't really a stability club (is it?) less should be sufficient so long as it's consistent. You can see in my run it's got about 0.0357 GFLOPS variance between the highest and lowest loops. Also 20 loops with 12GB of ram will take me until next year to complete, perhaps that's your grand plan! I dunno, say 10 loops? That should be more than enough to establish efficiency and basic stability, although I must say I've always got lower GFLOPS when slightly unstable.

I will try running with less as a test, but I shouldn't be excluded for running 12GB We can always have installed ram as a delimiter and have 6/12GB categories.

Sorry to hear about the 980X mate, hope you get a good one in return. Intel are very prompt in their turn around time for a claim and this is one of those situations where it cannot be determined exactly what happened but it certainly wasn't abuse on your behalf.

I've seen both higher and lower GFlops when not stable... If you're stable at 9216, all you gotta do is set ram size to 5120 and do 20 loops. It'll take less than an hour!

Yeah Spool... That's some bad luck... Sorry for your loss.
I've run all my chips with qpi llc on and 1.35v set in bios for 24/7. It usually sits at 1.33v-1.34v in hwmon. The 980x I'm running now was apparently from Intel rma. It's not as good as the one you had, but still very nice. Hopefully you'll get one at least as good.

[CryptiK]

Re-ran it with 5000 MB/s selected. 20 passes. Max GFLOPS = 79.6098, so that gives me 79.6098/4005 = 19.88 GFLOPS/GHz. Not bad, good enough for 3rd behind Don.

[spoolindsm127]

Zoson you got a good point about standardizing the test size. Does your guys' OS have enough free memory to test 5120MB? I'm a weirdo and like to test 'whole' gigabytes and I believe 5120MB is exactly 5GB. And how about 10 loops instead of 20? Maybe 10 loops for 12GB runs since it takes so much longer and 20 loops for 6GB?

And I'm with Cryptik on having separate categories for 6GB and 12GB. It's the fair way to do it. And I somehow accumulated 30GB of Hypers so I'm down to do some 12GB testing myself ha.

Thanks for the kind words and wishful thinking fellas. What's bothering me the most is the fact I never pushed the chip outside it's limits, and it still died. Feels like a waste. I really really hope they send me a good batch. But judging from my past experience, I'm not trying to get my hopes up. I feel Intel purposely sends THE worst batch they have on file when you do an RMA.

P.S. Cryptik - I just saw your edit about the AirFlow Module. So at 60% it keeps 12GB warm to the touch? And that's while benching right? If so it seems like you got good airflow in your case, so no need to watercool them.

And for your run I saw you had memory at 7-8-7-20. Have you tried 7-7-7-21? What's the difference in GFLOPS?

[zoson]

Zoson you got a good point about standardizing the test size. Does your guys' OS have enough free memory to test 5120MB? I'm a weirdo and like to test 'whole' gigabytes and I believe 5120MB is exactly 5GB. And how about 10 loops instead of 20? Maybe 10 loops for 12GB runs since it takes so much longer and 20 loops for 6GB?

And I'm with Cryptik on having separate categories for 6GB and 12GB. It's the fair way to do it. And I somehow accumulated 30GB of Hypers so I'm down to do some 12GB testing myself ha.

Thanks for the kind words and wishful thinking fellas. What's bothering me the most is the fact I never pushed the chip outside it's limits, and it still died. Feels like a waste. I really really hope they send me a good batch. But judging from my past experience, I'm not trying to get my hopes up. I feel Intel purposely sends THE worst batch they have on file when you do an RMA.

P.S. Cryptik - I just saw your edit about the AirFlow Module. So at 60% it keeps 12GB warm to the touch? And that's while benching right? If so it seems like you got good airflow in your case, so no need to watercool them.

And for your run I saw you had memory at 7-8-7-20. Have you tried 7-7-7-21? What's the difference in GFLOPS?

The issue with 12GB memory is you have to set 9216 or less to get full GFlops with HT on. So we can never really tell what kind of performance we could really extract out of 12GB, unless HT is disabled. Also... People with 12GB memory can easily just select 5120 and run loops... We want to get an easy to relate performance measurement against each other... Not a bunch of different categories that make it difficult to relate each other's performance!

I think it's also important to set a high speed requirement, to keep everyone in the same scaling area. Since performance scales worse at higher clockspeeds. Sorry Cryptik. It's also so I won't have to maintain a TON of entries... 4GHz is super easy to hit nowdays.

I've kind of envisioned this club to be everything the 4.5ghz club wasn't...

This is what I had in mind:
The High Performance Club
Welcome to the club for overclockers that maintain high performance along with high frequencies!


How it works:
HT ON and HT OFF results are ranked by Peak GFlops.
Low clock 100 GFlop results are ranked as a function of GFlops/GHz(Rating).
Only the highest performance entry per category will be kept per user.

General Rules that apply to all categories:
1. You need LinX to submit an entry.
2. You must be using the latest Linpack libraries.
3. You must have at least 6GB memory installed in the system.
4. All memory channels must be in use.
5a. Bloomfield/Gulftown/Lynnfield must be above 4500MHz and 85GFlops.
5b. Sandy Bridge/Ivy Bridge must be above 5000MHz and 110GFlops.
6. CPUz must be visible. If your voltage does not show, list your vBIOS in your post.
7. CPU-Tweaker must be open showing sub timings.

Additional HT ON Rules:
1. Linx must be running/calculating the 21st or higher loop.
2a. Bloomfield/Gulftown memory size must be set to 5120, regardless of your overall memory capacity.
2b. Lynnfield/Sandy Bridge/Ivy Bridge memory size must be set to 7168, regardless of your overall memory capacity.

Additional HT OFF Rules:
1. Memory size must be set to max capacity - 1024. (So if you have 12GB, 11264MB)
2a. Linx must be running/calculating the 21st or higher loop for less than 8192MB memory.
2b. Linx must be running/calculating the 11th or higher loop for more than 8192MB memory.

Low clock 100GFlop rules:
1. HT OFF only.
2. Results must be less than 101GFlops (100.9999 is fine).
3. No more than half of your loops may drop below 100GFlops.
4. Linx must be running/calculating the 21st or higher loop.
5a. Bloomfield/Gulftown memory size must be set to 5120, regardless of your overall memory capacity.
5b. Lynnfield/Sandy Bridge/Ivy Bridge memory size must be set to 7168, regardless of your overall memory capacity.

---

Bloomfield/Gulftown/Lynnfield

---

Bloomfield/Gulftown/Lynnfield - HT ON

Rank	User Name	CPU Type	Rating	Peak GFlops	Speed(MHz)
1	MrTOOSHORT	i7 990x	19.98	92.1319	4610.5
2	donmarkoni	i7 980x	19.94	91.9528	4610.8
3	zoson	i7 980x	19.64	91.5017	4657.9
4	spoolindsm127	i7 980x	19.22	86.6040	4504.9

---

Bloomfield/Gulftown/Lynnfield - HT OFF

Rank	User Name	CPU Type	Rating	Peak GFlops	Speed(MHz)
1	zoson	i7 970	21.74	100.2570	4611.0
2	donmarkoni	i7 980x	21.98	100.2397	4560.0
3	name	CPU Type	Rating	GFlops	Speed
4	name	CPU Type	Rating	GFlops	Speed

---

Bloomfield/Gulftown/Lynnfield - Low clock 100 GFlop

Rank	User Name	CPU Type	Rating	Peak GFlops	Speed(MHz)
1	donmarkoni	i7 980x	21.98	100.2397	4560.0
2	zoson	i7 970	21.74	100.2570	4611.0
3	name	CPU Type	Rating	GFlops	Speed
4	name	CPU Type	Rating	GFlops	Speed

---

Sandy Bridge/Ivy Bridge

---

Sandy Bridge/Ivy Bridge - HT ON

Rank	User Name	CPU Type	Rating	Peak GFlops	Speed(MHz)
1	name	CPU Type	Rating	GFlops	Speed
2	name	CPU Type	Rating	GFlops	Speed
3	name	CPU Type	Rating	GFlops	Speed
4	name	CPU Type	Rating	GFlops	Speed

---

Sandy Bridge/Ivy Bridge - HT OFF

Rank	User Name	CPU Type	Rating	Peak GFlops	Speed(MHz)
1	name	CPU Type	Rating	GFlops	Speed
2	name	CPU Type	Rating	GFlops	Speed
3	name	CPU Type	Rating	GFlops	Speed
4	name	CPU Type	Rating	GFlops	Speed

---

Sandy Bridge/Ivy Bridge - Low clock 100 GFlop

Rank	User Name	CPU Type	Rating	Peak GFlops	Speed(MHz)
1	name	CPU Type	Rating	GFlops	Speed
2	name	CPU Type	Rating	GFlops	Speed
3	name	CPU Type	Rating	GFlops	Speed
4	name	CPU Type	Rating	GFlops	Speed

[CryptiK]

You want me to burn my chip don't you

Format looks good, quite hardcore, but good. I would still lilke a 12GB category, its not that uncommon to have 12GB or more ram these days and I'd like to be able to take advantage of the improved efficiency it offers. It's an efficiency club afterall and with the proposed rules, the most efficient setups cannot enter!

[zoson]

The HT OFF category is for everyone with >9GB memory. Since you can do full memory testing with HT OFF. I think it makes the most sense to allow large memory sizes only when the cpu can fully stress it correctly. People with large memory sizes should DOMINATE this category.

It's a club for high performance with high frequencies! It's to separate the men from the boys. There are so many fakes on the 4.5ghz club...

In order to correctly rank, the playing field has to be level for everyone involved. That means there must be a set speed limit and a set memory size, or we will have irregularities that make the results no longer apples to apples.

It's also not feasible to lower the frequency requirement down to 4ghz... If I did... every noob would submit an entry!!! It's so easy now days to run 4ghz that it's no longer a 'high frequency'. Entrance into this club should be VERY difficult. At a minimum, more difficult to get into than the 4.5ghz club.

[zoson]

I think a big issue is that I would like AMD to be able to compete in this club. But I don't know crap about AMD performance. Can current AMD cpu's even hit 100GFlops with LinX?

[CryptiK]

Should be 2 HT on categories for 6 and 12GB its only 2 groups not like its 10 or something so not much harder to administrate. Should make it also full mem capacity -1024MB.

spoolin - yeah now running ram at DDR3-1868MHz | 7-7-7-20 1T, tRFC 60 and slightly tighter subs @ 1.594v. Still have some fine tuning to do but its a bit better than previous settings, which I was limited to due to running with no fan over the ram.

[zoson]

Should be 2 HT on categories for 6 and 12GB its only 2 groups not like its 10 or something so not much harder to administrate. Should make it also full mem capacity -1024MB.

spoolin - yeah now running ram at DDR3-1868MHz | 7-7-7-20 1T, tRFC 60 and slightly tighter subs @ 1.594v. Still have some fine tuning to do but its a bit better than previous settings, which I was limited to due to running with no fan over the ram.

It makes no sense to have a 12GB ht on category at all, since you can't ever stress your full memory amount, it is never known if your memory overclock is fully stable. The way bank interleaving works... You'll never fully engage the ddr channels. It's a flawed test, so it will not be added. This is the main problem. It doesn't actually show or prove anything.

If intel ever fixes the bug in linpack that makes the test not actually stress your machine if you use over 9216mb ram, i'll re-evaluate this position at that time.

It's also the difference between having to maintain 8 groups, instead of 6... which is a 33% increase in work. That is significant. Especially since I want to have AMD included as well... Which will make at least another four categories once BD is out.

[spoolindsm127]

Should be 2 HT on categories for 6 and 12GB its only 2 groups not like its 10 or something so not much harder to administrate. Should make it also full mem capacity -1024MB.

spoolin - yeah now running ram at DDR3-1868MHz | 7-7-7-20 1T, tRFC 60 and slightly tighter subs @ 1.594v. Still have some fine tuning to do but its a bit better than previous settings, which I was limited to due to running with no fan over the ram.

Nice. Your CMG kit is definitely very strong. Did you see any latency improvements in AIDA64? Also wondering what the difference is between 7-7-7-20 and 7-7-7-21.


Zoson - I think what Cryptik is trying to say is more like two categories of <6GB and >6GB. Although I never experienced it myself, I know the bug you are referring to in LinX.