![]() |
Suggested Rig, please
I'm thinking of buying a new rig for factorising big numbers using yafu. Obviously I'd like it to have several cores, to make it run reasonably fast, but I don't want to spend too many quid on something that may well be over the top.
Can anyone make any decent suggestions? |
How often are you going to be doing the work and how comfortable are you with linux? You might be better off renting cloud services. You could engage a large amount of compute for a short time, finish the job and not have to keep paying off the machine.
|
[QUOTE=Uncwilly;581899]You might be better off renting cloud services.[/QUOTE]
Seconded. Just be sure to configure persistent storage, and then use GCE's "Preemptable" and/or AWS's "Spot" instances, and you'll definitely get more "bang for the buck"! If you haven't already, sign up for Google's free $300 credit trial to get a feel for the environment. AWS's free trial doesn't give you anything worth computing on. Oh, and of course Azure; its free trial is also worth experimenting with, but I've found their compute is the most expensive of all the options. |
If you do prefer physically-owned hardware, I suggest used Xeon workstations, such as HP Z640 or Z840. Something around $1000 should get you twin 10- or 12-core Haswell Xeons and 64GB memory.
Then again, I'm not sure that's faster than a new/current 16-core AMD Ryzen, which is what I would likely choose if I were building my own machine from new parts. For GNFS jobs, CADO on linux is much faster than Yafu/GGNFS; we haven't collected enough data yet on SNFS jobs to compare the two packages. |
Physically for bang-for-buck IMO it's either go cheap with a 6 core Zen2/3, or go bold with a second hand server (but that's rough in the UK with our electricity prices, if better-electrically-blessed people are suggesting cloud then that's saying something).
|
Depending on how comfortable you are with dirtying your hands with strange hardware, and in tone with what Curtis said, try googling your favorite rubbish-selling site (like ebay, amazon, aliexpress, etc) for "lga2011 set". With a bit of luck, some average risk, and lots of work (work includes searching) you can get an 80-to-240-cores machine for as much as 2000-to-4000-bucks (which includes 5-to-10-sets, the power supplies you will have to add separate, local buy, a net switch, a monitor, etc.).
We posted about that in [URL="https://www.mersenneforum.org/showthread.php?p=575626"]Mike's yolo thread[/URL] some time ago. Edit: if you buy from Amazon, maybe you also consider [URL="https://www.mersenneforum.org/showthread.php?p=573636"]this[/URL]. |
[QUOTE=BudgieJane;581895]I'm thinking of buying a new rig for factorising big numbers using yafu. Obviously I'd like it to have several cores, to make it run reasonably fast, but I don't want to spend too many quid on something that may well be over the top.
Can anyone make any decent suggestions?[/QUOTE] Maybe i can show you my rig here i just built. It is 2 x 22 core Xeon e5-2699 v4 ES processors. So ES means: engineering sample. that means: not supported officially - but they work ok in china it seems they produced at least a 100k+ of them, so probably for an old supercomputer they were used. They go for 389 dollar each - when i bought 2 of them a few months ago. You will need a good motherboard for it and psu. I bought a supermicro XDAI motherboard for it. This motherboard is not overly expensive. It eats 360 watt under full load - yet the cpu's are close to each other so under full load you can only use watercooling practical. I can show some photos. I'm not having it in a rackmount case as that is too noisy so it is a 100% custom build. The watercooling i built is custom - not recommended to do unless you know what you are doing. I bought 2 radiators which can mount 12 cm fans and put 2 noctua fans on each radiator. some 3d printed plastic holds the gpu - as i need a reasonably fast gpu for it - the box main purpose is a CAD computer to me. Simply the rmaining system time is for prime numbers and sometimes my chessprogram Diep :) Now i already had a psu. Under full load 360 watt and add the gpu. The gpu i have is idle 22 watt yet in theory could eat 250+ watt under full load. It is very old AMD one that can do a high resolution and is delivering some Tflops single precision though it would be 0.5 Tflops double precision - for graphics only single precision matters obviously and a resolution of 3840x1200 for the CAD spreaded over 2 TFTs. You can get away with a lightweight gpu there for 25 euro newprice - or a second hand one. Now the better psu's you want to give just over 50% load. So you will want for a rig like this at least a 700 watt psu preferably of course 80+. Those are not so extremely cheap. Preferably you get a psu that is modular so that you can have enough cables for the motherboard. It needs 24 pins + 2x 8 pins for this motherboard. then the RAM i use cheap ram from china 8 sticks. killsre or something. ddr4 ecc reg. not sure of current price. the waterblocks i use cheap gpu waterblocks and some ABS 3d printed stuff moving over it and with simple m4 bolts (din912) from which i have thousands in stock of course as i design 3d printers and robotics, i bolt it. So all together if you would build this new and buy everything as i reused PSU and harddrives and bunch of fans it's like 1500-1700 euro depending upon where you live. Cheaper in USA. Ah photo of my rig exceeds the 1MB max size. it's 1.48MB. Doh. And this is linux this box - no good editor. Will do under windows soon in simple paint.exe resize it to smaller. p.s. for ES processors cheap chinese X99 motherboards will not work. |
| All times are UTC. The time now is 16:37. |
Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2023, Jelsoft Enterprises Ltd.