The (CPU) compute monster.

[norritt]

Hey everyone,

 

currently I'm snooping a bit around the net to create a build that delivers the best (heavily multithread) compute power/buck. So in general the following rules apply:

• more core's are better (probably even if the clock speed is lower)
• the cpu doesn't necessarily need to be x86 compatible
• the system should be able to manage a lot of memory 128GB+ is preferrable
• the system should have large / fast storage (thinking of an SSD 500 - 1000GB as primary and a HDD RAID >= 10TB as secondary storage). The storage throughput is not very crucial (no pun intended) tho. So good consumer grade products like a Samsung 850 Pro will be good enough. Most of the reads/writes will happen within the systems main memory. Spending more on a higher grade CPU/more RAM will be more valuable to the build than lets say investing into a PCIe SSD like the Intel SSD750.
• the components should be very power efficient since the system will run on 100% load most of the time when utilized.
• OpenCL/CUDA is not planned from the beginning. It would be nice but not mandatory to have it as upgrade option for later.
• peripherals, software, (dedicated) GPU are not required
• The build should include
  • Case + PSU (no rack cases, towers (sth. like fractal design R series) preferred)
  • Appropriate cooling (I'm not very sensitive about noise (GTX 980Ti ACX 2 cooler running on 100% doesn't qualify as particularly noisy for me but it would still be nice if the noise could stay below "hair-blower-level"))
  • Mainboard
  • CPU(s)
  • RAM
  • Storage (SSD / HDD)

Budget shouldn't exceed 2000$ . But if you have an idea for a baller 2500$ build that delivers 50% more performance than the best 2000$ build you can come up with hit me with your suggestion anyways. The performance/dollar (also considering power consumption) is more important than a hard budget limit.

[DXMember]

the best heavily multithered (or you might say massively paralelized) computing is on GPUs

CPUs on the other hand are very good at serial workloads

 

but if you're certaint that you want to do it on CPU then something like 5960X or one of the cheaper 2011 Xeons should do the trick

[Morgan MLGman]

What about something like that (I'm not sure if I understood your requirements correctly so just tell me if something's wrong):

PCPartPicker part list / Price breakdown by merchant

CPU: Intel Core i7-5820K 3.3GHz 6-Core Processor  ($374.99 @ SuperBiiz)
CPU Cooler: be quiet! Dark Rock Pro 3 67.8 CFM Fluid Dynamic Bearing CPU Cooler  ($74.90 @ Newegg)
Motherboard: MSI X99A SLI PLUS ATX LGA2011-3 Motherboard  ($183.98 @ Newegg)
Memory: G.Skill Ripjaws V Series 64GB (4 x 16GB) DDR4-2666 Memory  ($399.99 @ Newegg)
Storage: Samsung 850 EVO-Series 1TB 2.5" Solid State Drive  ($289.99 @ Amazon)
Storage: Western Digital Red Pro 6TB 3.5" 7200RPM Internal Hard Drive  ($279.99 @ Amazon)
Storage: Western Digital Red Pro 6TB 3.5" 7200RPM Internal Hard Drive  ($279.99 @ Amazon)
Video Card: EVGA GeForce GTX 970 4GB SSC ACX 2.0+ Video Card  ($309.99 @ NCIX US)
Case: be quiet! SILENT BASE 600 (Orange) ATX Mid Tower Case  ($84.99 @ NCIX US)
Power Supply: EVGA SuperNOVA P2 750W 80+ Platinum Certified Fully-Modular ATX Power Supply  ($119.99 @ Amazon)
Total: $2398.80
Prices include shipping, taxes, and discounts when available
Generated by PCPartPicker 2016-01-24 09:26 EST-0500

[norritt]

@DXMember: I know but the code is not (yet) written for CUDA/OpenCL. Parts of it may be rewritten hence the mentioned upgrade option.

[comefightme]

http://pcpartpicker.com/p/LXWVMp

 

PCPartPicker part list / Price breakdown by merchant

 

CPU: Intel Core i7-5960X 3.0GHz 8-Core Processor  ($999.99 @ SuperBiiz) 

CPU Cooler: Corsair H100i GTX 70.7 CFM Liquid CPU Cooler  ($104.20 @ Amazon) 

Motherboard: Gigabyte GA-X99-SLI ATX LGA2011-3 Motherboard  ($179.99 @ Micro Center) 

Memory: G.Skill Ripjaws V Series 16GB (2 x 8GB) DDR4-2400 Memory  ($74.99 @ Newegg) 

Storage: Samsung 850 EVO-Series 500GB 2.5" Solid State Drive  ($150.88 @ OutletPC) 

Storage: Western Digital Caviar Green 3TB 3.5" 5400RPM Internal Hard Drive  ($104.99 @ Newegg) 

Storage: Western Digital Caviar Green 3TB 3.5" 5400RPM Internal Hard Drive  ($104.99 @ Newegg) 

Storage: Western Digital Caviar Green 3TB 3.5" 5400RPM Internal Hard Drive  ($104.99 @ Newegg) 

Case: Cooler Master MasterCase 5 ATX Mid Tower Case  ($98.99 @ SuperBiiz) 

Power Supply: EVGA 850W 80+ Gold Certified Fully-Modular ATX Power Supply  ($117.99 @ NCIX US) 

Other: u wot m8 

Total: $2042.00

Prices include shipping, taxes, and discounts when available

Generated by PCPartPicker 2016-01-24 09:30 EST-0500

[comefightme]

Check out the one in my reply, it's got everything you asked for

[norritt]

@Morgan MLGman @comefightme: Thanks for your suggestions they look both quite nice and go into the direction I was aiming for. My favourite so far is comefightme's proposal since it provides more cores. It does have way to few RAM though I'm already running low with the 32GB in my current workstation. At that point Morgans build is way better. Calculating that in both build's should hit about the same price tag with Morgan's having more/faster storage and and comefightme's havin more CPU horse power (which has priority).

 

I actually kinda was hoping to get into the Xeon area (like 10-12 cores, i.e. E5-2650 V3) but damn those thingys are expensive.

[DXMember]

@DXMember: I know but the code is not (yet) written for CUDA/OpenCL. Parts of it may be rewritten hence the mentioned upgrade option.

well you'd have to test out your code on different platforms look maybe if AMD Opertron architecture benefits you or if indeed you need more cores rather than higher clock.

And if you are certain to go with 128GB RAM then you pretty much have to stick to servergrade hardware with ECC memory, Xeons or Opertrons which are quite hard to fit into $2000 budget

the budged is plenty to get a wicked gaming multi-GPU setup, but if you want to step into the servergrade stuff then it's nothing special really

 

so basically run some tests on different hardware configurations to figure out what setup benefits your code or maybe consider doing micro-architecture specific optimisations

 

out of curiosity, can you share what exactly your code will be doing? Because those requirements for huge core count on CPU and 100GB+++ RAM seem total nuts for a single application to use, only thing I could imagine it's doing is some scientific calculations or some meteorology simulations but those are generaly done on GPUs as well. And from the sound of it, it's not a host for VMs either

[DXMember]

I actually kinda was hoping to get into the Xeon area (like 10-12 cores, i.e. E5-2650 V3) but damn those thingys are expensive.

When looking at Xeons the first number in the 4 number sequence E5-[2]650 V3 - in this case the "2" indicates that they can also run in multi CPU setups so you are expected to get a dual socket motherboard and run two of these in one system. something like E5-4650 V3 would indicate that it can run 4 of these CPUs in one system. The other three numbers indicate general performance which takes core count and clock speed

 

and yeah, one Xeon pretty much eats up all your budget alone...

Maybe looks at AMD Opertron pricings, might be relevant - but I wouldn't  expect them to be a lot cheaper either

btw recently AMD released some Opertrons with ARM based cores

[Morgan MLGman]

@Morgan MLGman @comefightme: Thanks for your suggestions they look both quite nice and go into the direction I was aiming for. My favourite so far is comefightme's proposal since it provides more cores. It does have way to few RAM though I'm already running low with the 32GB in my current workstation. At that point Morgans build is way better. Calculating that in both build's should hit about the same price tag with Morgan's having more/faster storage and and comefightme's havin more CPU horse power (which has priority).

 

I actually kinda was hoping to get into the Xeon area (like 10-12 cores, i.e. E5-2650 V3) but damn those thingys are expensive.

To be honest, it's like this: If you want more cores, you shouldn't get the 5960X. You should wait a few months, because Broadwell-E releases this year and there will be two 6 cores, one 8 core and one 10 cored CPU with 20 threads for X99 platform. If you need the PC now, get the 5820k and sell it when broadwell-e releases this year to get one of those because 5960X will be a waste of money.

Another thing is, that if you said your build is going to run @100% 24/7, an AIO is not the best option since it has more ways to break than the usual aircooler. Another thing is that you should choose a platinum efficiency PSU like the P2 I picked (10 years warranty) cause it wastes less power.

I also picked a very silent CPU cooler and silent case so you won't be bothered by the noise of the PC running @100%

BTW the Green drivers are power efficient, but also slow. If you want good, reliable drives go Red or Black series.

[norritt]

out of curiosity, can you share what exactly your code will be doing? Because those requirements for huge core count on CPU and 100GB+++ RAM seem total nuts for a single application to use, only thing I could imagine it's doing is some scientific calculations or some meteorology simulations but those are generaly done on GPUs as well. And from the sound of it, it's not a host for VMs either

 

I will use the system for scientific calculations. There will be "workpackages" that have to be processed. They can be processed in parallel so it pretty much doesn't matter how many cores you throw at me I'll be able to throw a work package back that can be computed in parallel. You might compare it workloadwise to some of the projects distributed by that "Boinc" client.

[DXMember]

I will use the system for scientific calculations. There will be "workpackages" that have to be processed. They can be processed in parallel so it pretty much doesn't matter how many cores you throw at me I'll be able to throw a work package back that can be computed in parallel. You might compare it workloadwise to some of the projects distributed by that "Boinc" client.

well then deffinately explore the server market - Xeons, Opertrons.

Maybe Xeon Phi???

 

but be prepared to drop your wallet on the counter

Maybe you can do away with some Buldozer stuff from AMD FX series - you could really save up Money on the cores by getting a lot for cheap  and put the rest of the budget into high amounts of ECC memory, maybe you can build two or three boxes, like setting up two separate FX8320 chips each with 64GB ECC memory could be cheaper than some xeon stuff

[norritt]

and yeah, one Xeon pretty much eats up all your budget alone...

Maybe looks at AMD Opertron pricings, might be relevant - but I wouldn't  expect them to be a lot cheaper either

Yeah well the Xeon is expensive but not *that* expensive seems you can get those for around $1000. Still the problem is that i.e. the HDDs even though considered realy cheap meanwhile quickly add up if you surpass the magic 3TB border. (BTW refurbished workstations/servers/parts *with warranty* are an option too).

 

 

To be honest, it's like this: If you want more cores, you shouldn't get the 5960X. You should wait a few months, because Broadwell-E releases this year and there will be two 6 cores, one 8 core and one 10 cored CPU with 20 threads for X99 platform. If you need the PC now, get the 5820k and sell it when broadwell-e releases this year to get one of those because 5960X will be a waste of money.

Thanks for the hint that Broadwell-E actually looks *very* interesting to me. Funfact I just discovered looking into the Broadwell-E: Apparently Intel plans on releasing a Xeon with super high clock speed (5.1GHz). While this is not interesting for my scenario I assume this thing will deliver impressive single-thread performance for gaming.

 

Its also not that urgent to get the system right now. I just wish to get the compute jobs of my workstation to a dedicated machine so they're not limiting productivity.

[norritt]

well then deffinately explore the server market - Xeons, Opertrons.

Yes I'm definitly doing that. One of the reasons I posted this topic is to get more ideas what to look for (all the stuff I didn't think of).

 

 

well then deffinately explore the server market - Xeons, Opertrons.

Maybe Xeon Phi???

Those Phi's look damn amazing price/per core is really good. But apparently there is some special intel software/compiler envolved which most likely will cause some overhead, what might hinder the ability for rapid prototyping (also one of the reasons I'm not concentrating on CUDA/OpenCL only).

 

 

Maybe you can do away with some Buldozer stuff from AMD FX series - you could really save up Money on the cores by getting a lot for cheap  and put the rest of the budget into high amounts of ECC memory, maybe you can build two or three boxes, like setting up two separate FX8320 chips each with 64GB ECC memory could be cheaper than some xeon stuff

Yes I also thought about that (just using a stack of Banana PI's instead). But this approach has the major downside that it involves tinkering with Open MPI for parallelization (meaning again a lot of overhead), also the additional bottleneck and latency introduced to through the network communication between the boxes is to be considered.

 

While the Bulldozer idea sounds tempting at first I'm afraid the savings will be eaten up by the power consuption pretty fast (due to the full load scenario). Also the additional heat produced by these might add up to a problematic level especially during summer, since that box(es) stand under my desk rather than inside an air conditioned server room.