[XMG] M7M CPU mining discussion thread

[joelao95]

This thread is to collect hash rates of CPU mining on the M7M algorithm (XMG, OP: https://bitcointalk.org/index.php?topic=735170.0). This will facilitate us to make a wiki page.

Please post your CPU, hashrate, the miner (which compilation or compiled from source code), and any other means of improving performance you want to share.

M7M miners:

http://coinmagi.org/files/magi-minerd/Magi-miner-v0.1.zip
This includes Spexx and MarcusDe's compilations (generic). Spexx's software package is way of better.

Spexx's latest package: https://bitcointalk.org/index.php?topic=735170.msg9503440#msg9503440

MarcusDe full list: http://graymines.net/miners/magi/v2/

For linux, compile from source code: https://github.com/noncepool/m7magi-cpuminer-v2

Estimate hash rate in kh/s (according to Spexx's research)

Locate your CPU in https://www.cpubenchmark.net/cpu_list.php; divide "Passmark CPU Mark figure" by 1000 and then multiply by 3.6.

[1715461317]

1715461317

[1715461317]

1715461317

[1715461317]

1715461317

[go6ooo1212]

Im reaching around 30kh/s with haswell i7-4790 with 7 treads

[lordmick]

Ubuntu miner from source code

Intel(R) Xeon(R) CPU       E5440  @ 2.83GHz 8 treads +- 32 khash/s
AMD Athlon(tm) II X2 220 Processor 2.8Ghz  2 treads -+ 5 khash/s
Intel(R) Pentium(R) CPU B960 @ 2.20GHz 2 treads -+ 4.5 khash/s

Win_64 with spexx miners

AMD Athlon II X3 450 3.2ghz 3 treads +- 12 khash/s
Processor: Intel(R) Core(TM) i3-2330M CPU @ 2.20GHz , ~2.2GHz 4 treads +- 7 khash/s
Processor: AMD Athlon(tm) 64 X2 Dual Core Processor 5000+ , ~2.6GHz 2 treads +- 6 khash/s

[tubesteak]

2x Intel i5 3570k 4.2ghz - 3 threads: 24 khash avg
3x Intel Pentium G3220 3.0ghz - 2 threads: 12 khash avg
W/spexx

[Atomicat]

Im reaching around 30kh/s with haswell i7-4790 with 7 treads

Whaaaaaat?  And I thought you Intel boys were fast!  $330 bucks for that?  Should get your money back.



Sorry, but you know that oh ok, 'certain' Intel users treat AMD users like 2'nd class citizens.  Couldn't resist.

*sigh*  Oh ok, confession time... Overclocked to 5Ghz.  Spexx miner, 7 threads, Win-7 64 bit.

Lately I've been playing around with some extreme ram clocks and timings (2133's taken to 2650 anyone?) and I'm pretty sure there's some sweet spots.  I've noticed better hashing with looser timings for instance.  So if you find it's sitting there and twiddling, maybe give it some slack.

[djm34]

Im reaching around 30kh/s with haswell i7-4790 with 7 treads

Whaaaaaat?  And I thought you Intel boys were fast!  $330 bucks for that?  Should get your money back.


Sorry, but you know that oh ok, 'certain' Intel users treat AMD users like 2'nd class citizens.  Couldn't resist.

hum, I always thought that AMD were 2nd class processor, no sure what I am thinking about AMD users though 

[go6ooo1212]

Im reaching around 30kh/s with haswell i7-4790 with 7 treads

Whaaaaaat?  And I thought you Intel boys were fast!  $330 bucks for that?  Should get your money back.


Sorry, but you know that oh ok, 'certain' Intel users treat AMD users like 2'nd class citizens.  Couldn't resist.

hum, I always thought that AMD were 2nd class processor, no sure what I am thinking about AMD users though 

I'll always like the AMD cpus and gpus , but they havent any real architecture CPU development since several years...
Yeas I admit that overclocking is significant, but for 1Hz haswell is doing more work than any AMD CPU.

[Atomicat]

Im reaching around 30kh/s with haswell i7-4790 with 7 treads

Whaaaaaat?  And I thought you Intel boys were fast!  $330 bucks for that?  Should get your money back.


Sorry, but you know that oh ok, 'certain' Intel users treat AMD users like 2'nd class citizens.  Couldn't resist.

hum, I always thought that AMD were 2nd class processor, no sure what I am thinking about AMD users though 



Oh we're definitely a bit sleazy, but that's a good thing... right?

I'll always like the AMD cpus and gpus , but they havent any real architecture CPU development since several years...
Yeas I admit that overclocking is significant, but for 1Hz haswell is doing more work than any AMD CPU.

Yeah, but I got more of them!  They're just different chips.  AMD out-performs Intel in some things, not in others.  Just now found out that the trick of running 7 individual threads bound to individual cores doesn't work with my 8350.  At 5 threads they all slow down a wee bit, and at 7 threads there's no real difference.  The most important thing though is that price difference.  Consider, I got chip and board for $400!  You can't even get a motherboard over $300 for AMD, top stuff for Intel is $500+.  There are a hell of a lot of people out there who probably wouldn't even have a system if it weren't for AMD's el-cheapo combos.  Essentially though, they're just different.

[ulfie]

AMD Phenom II X6 1075T
Linux.  The miner was built from  https://github.com/noncepool/m7m-cpuminer-v2.git

Passmark is 5422 so the estimate would be 19.5 khash/s, presumably on all 6 cores.
I have it set for 8 threads and it is taking most of 4 of the cores (~365% CPU).  Sorry, I have a couple threads devoted to other coins, too, plus some wallets.

Still, I am seeing 16.5 khash/s.  That is giving me about 4.4khash/s/core or 2.2khash/s/thread

I'm happy.  I'll never be rich, but I am having a lot of fun.

[SpeedDemon13]

I know I posted this on the XMG thread, but has anyone done a comparison of hash/wattage of cpus? What is the most ideal cost/performance cpu for mining M7M algo?

[Atomicat]

Continuing a discussion on the main thread....

Spexx!    Just wanted to possibly confirm a bit of bad news for us AMD miners.   It seems that when limiting single threads to single cores, we still get a drop in speed.  Basically, once you go past five instances you get that drop no matter what.  7kh/s per thread up to 5 or so, then slowly drops to 5 and a bit per.  Just thought I'd mention before I start digging for ways and means.

Would this be on a machine with 8 cores?

You are probably seeing a near linear increase in total hashrate as you add single-threaded minerd instances up to 5 in number, then see a drop-off in the per-thread hashrate as you add further instances. The reason for this is that the processor load from each instance is spread across 3 cores in any event (given at least 3 cores to run on).

If you run up a single instance of minerd.exe with a single thread on a multi-core machine that is otherwise idle, this can be seen using Task Manager. Once you have 5 instances on an 8 core machine, you will start to get overlap - there is no core that will be 100 percent free for a sixth, seventh or eighth instance and the instances start to compete with each other for CPU time.

While this will cause a drop in the per-thread hashrate as you have observed, the total hashrate obtained is optimal when you have the number of instances/threads equal to the number of processors in the machine and the CPU utilization will be maxed-out 100 percent at this point.

Deliberately restricting an instance of minerd.exe to a single core on a multi-core machine is not the best way to do it. It is very inefficient.  Each instance of minerd.exe should be allocated a subset of 3 processors to run on for best performance. This applies to Intel processors as much as it does to AMD processors from what I have observed. This is why I wrote a shed load of code to handle it

I see... that's somewhat counter-intuitive but maybe because of old-school thinking.  I would have figured the increase came from having task run on "it's own core" so what I was doing was starting different instances with the /affinity command, which I saw you were using in your batch files.  So you're setting them to spread over three cores?  As in 123,234,345,456,567,678 etc?  Guess I'll try that and see what I get.

Y'know... I think I'd gain more by moving my rig to where it can get some sub-zero air.  Then of course there's more problems.  Last winter I found out they don't put antifreeze in closed-loop systems.  There is such a thing as being too chill.

[Spexx]

 So you're setting them to spread over three cores?  As in 123,234,345,456,567,678 etc?

... then 781, 812 to finish-off.

Yes that is precisely what the automine.bat program does. There is a refinement in as much as when the number of cores is a multiple of 4 it will wrap around itself to saturate a group of four cores. E.g on an eight-core machine it would go:-

123, 234, 341, 412 then 567, 678, 785, 856

The only reason that I wrote it in such a way was for mathematical convenience when dealing with hex affinity masks. If you start crunching the numbers yourself you will see what I mean The net result is exactly the same as going 123,234,345,456,567,678,781,812 since each core is used three times in either regime.

It is certainly possible to start multiple instances of minerd without specifying a processor affinity mask at all, thus letting Windows make all those decisions for you at runtime, but it tends to make a pig's ear of it. You end up with a chaotic mess which fails to fully utilize the CPU and generates inconsistent hashrates. Everything seems to run smoother when you take control and do a proper job on it.

[JimmyTaylor]

I think this looks to be the best way to organize convos about that specifically.

[Atomicat]

It is certainly possible to start multiple instances of minerd without specifying a processor affinity mask at all, thus letting Windows make all those decisions for you at runtime, but it tends to make a pig's ear of it. You end up with a chaotic mess which fails to fully utilize the CPU and generates inconsistent hashrates. Everything seems to run smoother when you take control and do a proper job on it.

Interesting, however... Might not be an issue with AMD's bulldozer and might have more to do with hyperthreading.  Dig.  Here's a run, 7 threads given time to settle in.  Shows that drop of course down from 6.9. 



This however, is a run, single threads, affinity set to e0, 70, 38, 1c, 0e, 07, 83.  It looks like it's going to average out to about the same, possibly less, and there's a lot of work restarts, no doubt from all those single threads jammed at the doorway so to speak.



This is the first AMD chip I've used since this 1800+ I had quite a while ago.  An upgrade from a quad-core socket 775 chip/board that went to better pastures (bro's accounting and email).  This scene has re-sparked my curiosity bug for programming, the low-level kind.  Previously I would have probably got into programming PovRay, now that I've got something faster than a 486 to render on, but now I wanna know what's in those cores, and what the hell all these new instruction sets are about (including the gpu ones).  First project... something that will adjust core voltages and frequencies on the fly, dependent on core temperature.  Why?  Because the manufacturer's overclocking software sucks so much ass.     This will be fun, I haven't done asm since I had a 386!

[Atomicat]

AMD Phenom II X6 1075T
Linux.  The miner was built from  https://github.com/noncepool/m7m-cpuminer-v2.git

Passmark is 5422 so the estimate would be 19.5 khash/s, presumably on all 6 cores.
I have it set for 8 threads and it is taking most of 4 of the cores (~365% CPU).  Sorry, I have a couple threads devoted to other coins, too, plus some wallets.

Still, I am seeing 16.5 khash/s.  That is giving me about 4.4khash/s/core or 2.2khash/s/thread

I'm happy.  I'll never be rich, but I am having a lot of fun.

That chip has six cores so that's the max threads you should be running.  I just started up my bulldozer with 12 threads to see what happens.  Collision city, and less hash than with 7 cores.  Just give it five, best for efficiency.  There is that drop-off that Spexx and I were discussing, you might try single threads and setting processor affinity as he described.

Yeah it is fun!  I'm fairly competitive, love to learn, and love to overclock to the moon.
 

[ulfie]

Thanks, mate.  That gave me an extra 6% on the overall hash rate.

[Atomicat]

So I downloaded some ASM materials, reference manuals for op-codes, registers, data types, addressing modes and such, and this is what I find...

VFNMADD132SS/VFNMADD213SS/VFNMADD231SS — Fused Negative Multiply-Add of Scalar Single-Precision Floating-Point Values

VFNMADD213SS: Multiplies the low packed single-precision floating-point value from the second source operand to the low packed single-precision floating-point value in the first source operand, adds the negated infinite precision intermediate result to the low packed single-precision floating-point value in the third source operand, performs rounding and stores the resulting packed single-precision floating-point value to the destination operand (first source operand).

Mommy... I'm scared!  Can we go home now?

Calm blue ocean calm blue ocean calm blue ocean.

Y'know, the 6502 was such a nice chip, such a nice chip, why did it have to grow up and get all mean?

6502... Mult operation?  We don't need no steenkeeng MULT operation!

Revised ambition... Brush the cobwebs off of C+, such a nice clean thing, sparkly.

[JimmyTaylor]

Sorry what are you talking about exactly?

So I downloaded some ASM materials, reference manuals for op-codes, registers, data types, addressing modes and such, and this is what I find...

VFNMADD132SS/VFNMADD213SS/VFNMADD231SS — Fused Negative Multiply-Add of Scalar Single-Precision Floating-Point Values

VFNMADD213SS: Multiplies the low packed single-precision floating-point value from the second source operand to the low packed single-precision floating-point value in the first source operand, adds the negated infinite precision intermediate result to the low packed single-precision floating-point value in the third source operand, performs rounding and stores the resulting packed single-precision floating-point value to the destination operand (first source operand).

Mommy... I'm scared!  Can we go home now?

Calm blue ocean calm blue ocean calm blue ocean.

Y'know, the 6502 was such a nice chip, such a nice chip, why did it have to grow up and get all mean?

6502... Mult operation?  We don't need no steenkeeng MULT operation!

Revised ambition... Brush the cobwebs off of C+, such a nice clean thing, sparkly.

[Atomicat]

Sorry what are you talking about exactly?

Getting back into programming.  I'm trying to find the right niche where I can brush off the cobwebs and exercise old skills.  One thing I was good at back when was optimizing machine code, back when we actually had to write our own .h libraries for things like drawing lines with those fancy COLOR videocards!  Last of that I did was on a 386 though, then it was off to Acad Lisp and other more profitable things.
 

[r3morph3d]

Here is mine Intel B970 2cores-2threads
Win64bit-8GB Ram with spexx miners



and an Intel core2duo E6550 2cores-2threads
Win64bit-4GB Ram with spexx miners



Intel i53210M 2cores-4threads
Win64-6GB Ram with spexx miners