Welcome Back SpeedRaider!

[guru]

http://cpuboss.com/cpus/Intel-Core-i...-Core-i5-2500K

You make a mistake, you are only considering the investment, you also need to consider the production per watt consumed. Core-i5-4690K is the king for prime search (LLR).

For me power is cheap! Also I would go with the 4770. Yea you don't need the HT but it has 8MB L3 cache vs the 6MB L3 of the 4690K. That cache should help with the larger FFT's

Also don't get misled by typical power usage. Prime95 is anything but typical. You need to be looking more at max power. Haswell has much better idle power than the previous two generations but max power hasn't changed much. The speed settings of released Intel processors has more to do with fitting the max power into a specific envelope than what the processor is capable of running at. One reason why they overclock so well. If you really want to compare system power per watt I can connect a power meter to do some measurements. Two systems will use more power than one but it is about the power per workload unit.

[guru]

i5-4690K
6 MB cache
1.5MB cache per core

i7-2700K
8MB cache
2MB cache per core

Xeon X5690
12 MB cache
2MB cache per core

E5-2697V2
30 MB cache
2.5MB cache per core

[pinhodecarlos]

In case for a domestic processor the CPU I mention is the best solution. Server is Intel Xeon E5-2697 V3.
If I had £7000 I would invest on a dual socket Intel Xeon E5-2697 V3.

[pinhodecarlos]

Guru, forget the Xeon X5690 processors and its family, it doesn't have avx on it so they produce half or less than an E5-2697V2.

For me energy is not cheap and you need to think that common people do pay electricity.
Please do the comparison you are saying, with an energy meter read the consumption and determine the production per day.

[guru]

In case for a domestic processor the CPU I mention is the best solution. Server is Intel Xeon E5-2697 V3.
If I had £7000 I would invest on a dual socket Intel Xeon E5-2697 V3.

Yea the dual socket systems are not worth the price for crunching but if you happen to have them sitting around they will put out some good numbers. I included them in the benchmarks just because I could get some data from them. I have more systems with data to release once I get some time.

[guru]

Guru, forget the Xeon X5690 processors and its family, it doesn't have avx on it so they produce half or less than an E5-2697V2.

For me energy is not cheap and you need to think that common people do pay electricity.
Please do the comparison you are saying, with an energy meter read the consumption and determine the production per day.

The systems I have setup now were geared towards dnetc. Now that I am investing in SOB the systems will change. Newer systems don't have any advantage for dnetc which is why I have so many older systems. Not having good work statistics in SOB makes it a challenge to know which systems are worth running. Right now I am going off benchmark results from specific systems.

It will be very interesting getting the running power data for the newer systems on prime95. I won't bother with power ratings for the dual socket servers.

[guru]

pinhodecarlos, what is your SOB user name?

[pinhodecarlos]

pinhodecarlos, what is your SOB user name?

I don't run SOB, I run another private prime search at mersenneforum.org called RPS.

[guru]

Doh! Found this in the logs from my i7-3960X

[Tue Feb 10 16:36:23 2015]
Iteration: 6934673/15990667, Possible error: round off (0.4375) > 0.40
Continuing from last save file.
[Tue Feb 10 17:00:31 2015]
Disregard last error. Result is reproducible and thus not a hardware problem.
For added safety, redoing iteration using a slower, more reliable method.
Continuing from last save file.

I still say there is something wonky with the AVX code. I only have crashes on the AVX enabled systems and all of them are running stock speeds. I could accept one or two systems but 4 of 7 systems is a lot. I have non AVX systems that are overclocked and no issues at all. Only one of the systems affected is running linux. The others are running windows 7 x64.

[pinhodecarlos]

I already had those errors when I was using Prime95 on dual core server (linux mprime).
When I shared at mersenneforum that error Prime95 told me not to worry. Few weeks after when someone found a bug on the software I decided to retest all candidates in a another CPU and OS and I got different residues. I shared this on mersenneforum but didn't get any reply.

[guru]

I added ~ 25 AVX systems yesterday and three have already crashed.

[tim]

Guru, were they overclocked much? The only stability issues I've had with prime95 client were hardware related. But I don't overclock either (don't slap me too hard).

Thanks for posting the comparisons of cpus. I noticed i5-4690K didn't have any data, did it not appear properly in the post? My rookie interpretation of the posted data makes me think on a per core basis the i7-4770K and the E5-2697V2 are roughly the same in production. Is that how you read it?

I appreciate you taking the time to hash this out. I hope it's not affecting your day job.

[guru]

Nope these systems are in 1U racks and used only as computing systems. 24/7 stability vs high performance. Many of my AVX systems have the same motherboards and very similar memory. It is possible it is not related directly to the CPU. I have noticed turning off Turbo helps. I will do more testing with a few specific trouble systems and try to isolate the issue.

I don't have a 4690K for testing so I just put the specs to compare it to what i have. If someone else has one that can provide the data that would be good. I have more results to share and I may just put it into a spreadsheet and upload it.

My rookie interpretation of the posted data makes me think on a per core basis the i7-4770K and the E5-2697V2 are roughly the same in production.

It depends on the FFT size. The 4770K is faster at 1024K FFT size but most of what I have seen running are 2560K which puts the 2697V2 about 45% faster than the 4770K

This doesn't directly relate to my day job but everything I learn from it does apply. Knowledge like this doesn't come in a book or a lecture.

[pinhodecarlos]

Can I suggest running SRBASE project instead of SOB? With such power on SOB with such errors it will cost SOB more doublecheck effort in the future.
http://srbase.myfirewall.org/sr5/index.php

The invitation code for creating new accounts is pillepalle

[guru]

Can I suggest running SRBASE project instead of SOB? With such power on SOB with such errors it will cost SOB more doublecheck effort in the future.
http://srbase.myfirewall.org/sr5/index.php

The invitation code for creating new accounts is pillepalle

Does it support proxy, Linux, and command line?

[guru]

Here are the power number comparing the 2700K to the 4770K. Both are very similar and basic systems.

2700K
850W power supply
Intel SSD
2 x 4GB 1600 Memory
DVD
Intel DZ68BC motherboard
OpenSuse 13.1
HT off
4 Cores
Turbo disabled

4770K
610W power supply
Intel SSD
2 x 4 1600 Memory
DVD
Asus Z87-Expert motherboard
Windows 8
HT off
4 Cores
Turbo disabled

             Idle   SOB   TT small      TT blend       TT large      OGR   RC5
2700k        42     129   124           122            129           104   104
4770K        66     142   142           129            140           97     95

Take away: 4770K with avx2 is faster but requires more power to do it(nothing for free). I suspect the extra power is the stress on the memory system due to the increased throughput. FFM3 instructions use more power than AVX. Makes sense given the larger registers.

99% sure the difference in idle power has to do with the c-states in the bios. I believe they are disabled on the 4770K(something we normally do for testing) but enabled on the 2700K after the last bios update. This would not affect the power usage under load.

[tim]

Thanks for clarifying that. Electricity isn't as big a concern for me, we have plentiful hydroelectric power in the Pacific Northwest. But the dollars remain one.

Looking at http://www.newegg.com/Product/Produc...Size=30&Page=5 just daydreaming. I'm thinking I'm better off sticking with multiple quad core boxes, since I'm clueless about server maintenance, costs and requirements. A quick google search on "servers vs quad core distributed computing" and nothing looked pertinent. There was a thread from a guy in England considering getting some for his flat; seems like he found more problems than he could take on. I've built home machines, but it looks like servers are a whole different thing. Is it hard to jump from building my own boxes to mounting, powering and maintaining servers?

[guru]

You must not be far from where I am. I get 80% of my power from the Columbia River. Desktops are much cheaper than a server. The servers advantage is power density and large memory. I use the waste heat from my systems to keep my shop warm during the winter. HT doesn't help with SOB but the i7's usually have 8MB cache vs the i5 with 6MB. I haven't done a comparison to see how much it affects performance of SOB. I suspect as the FFT size goes up there is a cutoff where the 6MB cache performance drops off. Having high speed memory becomes more important at that point.

[tim]

I'm in eastern WA near Moses Lake. I heat my basement too, but less so than years ago. Back in the days of SETI classic, I used to have stacks of mobos booted from linux off floppies. I learned some about scripting and the command line. It was fun. When it went to boinc is when I changed to SOB. So I'll keep my eyes open for i5 or i7 with large cache and pair it with fast, low latency memory.

[pinhodecarlos]

Does it support proxy, Linux, and command line?

I don't know if BOINC supports proxy server as dnet but it runs under linux, GUI or command line.

About your calculations, you need to determine the specific energy consumption using the same batch of candidates, you can't compare watts to watts.

Specific energy consumption=(Watt per hour) / (candidates tested per day)

[guru]

Watts are watts! What you are talking about is watts/work unit or efficiency. Yes the 4770K is more efficient than the 2700K but not by a lot.

Getting back to the two 2700K vs single 4770K output. That is 258 watts vs 142 watts. With power at $.064 a KW here that comes to $63.91 a year extra in power but it also outputs more units. The added bonus is the two 2700K's will crush the 4770k in other work. SOB is one of the few workloads that has a big advantage on the 4th generation core processors.

[pinhodecarlos]

Watts are watts! What you are talking about is watts/work unit or efficiency. Yes the 4770K is more efficient than the 2700K but not by a lot.

Efficiency is one thing, specific energy consumption is another thing.
When you are on a site you look at a efficiency of a boiler but for a production line you look at the specific energy consumption. The latter is what you need to be use to compare your computers. Your computer is a production line where you feed with energy (electricity) and it produces X candidates per day. The efficiency is measured on the power supply (consumed/feed)_energy.

Also that unit KW doesn't exist because the K (capitalize letter) is Kelvin, for absolute temperature, what you mean is kW, where k is a SI prefix and it is named kilo. Second you pay energy (kWh) not power (kW). Third you need to consider the overall investment for each case, you can't only compare the energy you pay, you need to determine the ROI (return of the investment). This is what you do in the real world.

[tim]

Sounds like you two guys know a lot more about this than me. So, if I'm following, I'd need to look at how much work gets done per kilowatt/hour used. Or kW/h if I have the nomenclature right. And for me the initial cost of the system does have a bearing. Looking at cpuboss, the 2700k is behind the 4770k, but not by a lot. 2700k vs. 4770k: AVX vs AVX2, DDR3-1333 vs DDR3-1600, $83.22 vs $73.58 annual cost. Clock speed and turbo speed the same, memory cache amounts the same. $400 through amazon vs. $345 at newegg. If I understand right, for SOB I'd want to look for a 4770k?

[guru]

If you are buying new the 4770k or 4690K are good choices. The 4960K should be cheaper with very similar performance. If you shop used and can get a good deal(say half of the price of 4770k) on the 2500k, 2600k or 2700k they will give good results.

I'll be out crabbing tomorrow so no new updates until Sunday.

[pinhodecarlos]

Tim,

There is memory bottleneck issues which means you need to have faster memory and bigger cache. Also you won't take advantage by having HT on the processor so you can rule out i7 and stick with an i5.
Because SOB is running candidates with exponent n=29M, you will see a big degradation in speed when you keep adding instances, this means that you need to understand on your processor if it is better to have one or two or three or four instances running. Example, I have two laptops, one Haswell and one Ivy Bridge, and I only can run on them two instances of prime95 (or LLR) but I am running RPS project with k=5 and n=4.6M. Memory is DDR3 1600 MHz. At this exponent by adding a third instance I can notice memory issues, decrease on overall output.

What people do on Prime95 project (GIMPS) is to only run 2 LL tests (these tests take like 30 days or more, not sure for now) and 2 factoring because it is more efficient to do this on a quad core machine. In guru machines I would run half of the cores with SOB and the other half on dnet...something like that....

At the end if you don't care about the best efficient way to run SOB just run one thread of Prime95 per psychical core. If you care for now the best overall option, for home users that pay electricity, is the core-i5-4690k with DDR3 at 2400Mhz, taking into consideration that fact that it is needed to make a test on how many cores you should run (for SOB, high exponents to test).

Carlos

[guru]

Tim,

There is memory bottleneck issues which means you need to have faster memory and bigger cache. Also you won't take advantage by having HT on the processor so you can rule out i7 and stick with an i5.

Carlos

Yes you don't need the HT from the i7 but the i7 does have 8mb cache vs the 6mb cache of the i5. That 2mb should make a difference in SOB.

[tim]

Thanks for the advice. I'm not finding refurb i5 4690k. I'll have to look at building. Or get back into it, it's been years. The last one I built was a Athlon xp 3000.
Too bad nobody makes a LGA-1150 dual socket mobo. Or an AM3+ dual. 16 cores on one home machine would be nice.

[tim]

Even as dual socket AM3+, it would be faster than i5 4690k not by much. From looking at mersenne.org's cpu comparison of iteration times, it looks like 2 amd 8 core processors would do about 1/4 more work than i5 4690k in the same amount of time. At a cost of more electricity. I'd still be interested if a dual mobo existed. Or better, dual lga-1150 socket.

[pinhodecarlos]

Even as dual socket AM3+, it would be faster than i5 4690k not by much. From looking at mersenne.org's cpu comparison of iteration times, it looks like 2 amd 8 core processors would do about 1/4 more work than i5 4690k in the same amount of time. At a cost of more electricity. I'd still be interested if a dual mobo existed. Or better, dual lga-1150 socket.

Don't put AMD's to run LL tests, it's a waste of time due to the fact that you can't take advantage of the avx instructions.

[tim]

AMD's . . . can't take advantage of the avx instructions.

I read that AMD FX-8350 does have avx instruction set, not avx 2.0. Has the client not been updated to take advantage of it? I have also noticed that amd chips don't do near as well as intel at sob.

[pinhodecarlos]

I read that AMD FX-8350 does have avx instruction set, not avx 2.0. Has the client not been updated to take advantage of it? I have also noticed that amd chips don't do near as well as intel at sob.

I think you need to add the line "CpuSupportsAVX=1" to local.ini to take advantage of avx on AMD but problems may arise. Just test and see what happens.

[tim]

I don't have an AMD chip that supports avx. My referring to AMD FX-8350 was just me lining up prospective cpus. Wishing there were a dual socket AM3+, to have processing power without so much cost. But I'm leaning toward building an i5 4690k, like you suggested. I'll keep looking as time goes on.

[guru]

AVX2 isn't really what helps with SOB. AVX2 processors have FM3 which is what is helping speed up the processing over AVX. I'm not sure which AMD processors(Not my cup of tea) but many of the newer ones have FM4 which should help but to what degree I don't know off the top of my head. The AMD processors are cheap on the purchase side but it will end up costing more in the long run due to power costs. Even cheap power can't make up that difference.

The problem with many dual socket MB is memory bandwidth. It is better to have two desktops vs a single dual socket system.