Hi. I'm looking to build a workstation that supports at least 128GB of RAM. I would like to build sooner rather than later, and I am sensitive to the price. I'm not a big fan of Intel, but given my restrictions, I don't see there being much choice otherwise. I came up with this very rough build with dual Xeon E%-2670s:
The prices might be slightly off, as I just went into eBay and searched for compatible parts. I'm not even sure if the Xeon's are used or not, but the price seems about right for a used part. I can probably get this price down by looking for more used parts. I have the RX 580 in there for some potential crypto currency mining and because with this mother board, I will need to dedicated graphics to run a 4k external display. I have an 850 pro 512 GB SSD that I can use for the interim as well, that's why there is no SSD in the build. This build does seem somewhat expensive however if I compare it to what I could be getting if I can wait for x399. I guess x399 would add about another 1000 EUR to the price of this build though. Same as if I built x99, granted I don't want to use a CPU with less than 8 cores.
Dual 2670s have been a great option for some time, I built two of those systems. But Threadripper is around the corner and even a Ryzen 1700 at 3.8 to 4GHz comes very close to that level of performance.
By mentioning an AM4 CPU, you are ignoring the restriction I have for at least 128GB RAM. If I didn't have that restriction, I would agree that it would be a silly idea. Threadripper doesn't have a precise release date, just in the next three months, and it is also more expensive by about 60-70%. I should clarify that I need this computer to fulfill a workload I currently have. The way I am managing at the moment is by using Google Cloud Compute, however I realised that their instances are actually not that high performance. The Xeon I was assigned with 16 threads was running at 2.2 Ghz, and so it actually takes quite a bit of time to complete tasks. Also the server is not that convenient to use (Transferring large files between the cloud takes time and bandwith; Sometimes the machine runs for longer than it needs to, because I can't monitor it while I'm asleep and I'm charged for it; I need time to figure out the intricacies of Google Cloud itself etc.).
Without knowing what the requirement for the hardware is we can't offer a good balanced system for your needs, but based upon price/performance and that you need it now I would get the dual Xeon workstation, it likely wont be as fast as Threadripper but again we don't know the specs of the new CPUs.
Additionally as far as performance metrics go do you know if your application is optimized in a specific way?, for example takes advantage of higher clock speeds/less cores like photoshop, or does it require multiple slower cores, something like VMWare may require.
Should be the deciding factor, some applications benefit from more cores / threads, others it's all about clock speed and cycles on a given core, of course ideally a well balanced system would have lots of cores / threads and a high clock speed, since you are considering dual Xeons I'd assume the clock speed is of secondary importance for your application.
I'm kinda' wrestling with the same problem (although I don't have the memory requirements your looking for) I've decided to wait till the end of summer / fall to see how things shake out with Ryzen and threadripper before committing my cash on either platform.
The software benefits from higher clock speeds, but also from more cores. The author said to a customer, who said his 64-core machine with 416 GB RAM was not significantly faster than the 10 core machine he was using previously, that he thinks a 10-core machine with faster cores could beat a 64-core machine with slower cores. I took this to mean roughly that a 6950x for example would beat whatever Xeons they are typically using in cloud servers for machines with say 64 vCPUs.
I've asked the author of the software if he could provide some more detailed information on benchmarks. E5 2690 looks interesting. Shame the v2 10-core part is so expensive. Do you know if it overclocks well? Also, do you have a motherboard recommendation for x79?
E5 1xxx do very well depending on the cpu. They are normaly equivalent to the i7s just for servers.
People overlook that apparently and thus prices are higher for the HEDT i7 then for the E5 16xx Zeons. Sadly those i7s / 16xx v1 / v2s just go to 6cores.
X79 motherboards are genuin pretty expensive compared with X99 considering their age. You get x99 used for around 120€ here in germany. I'll have alook at prices again, because i allways forget.
I myself buy "broken" and then i have fun fixing them. Or no fun because they work ore are completely dead ... it depends.
You may have just answered your own question there.
If the author says a 10 core could win the 64 core CPU in their application, then it prefers core clock speed not core count, so by that I would focus on a faster CPU with a decent core count.
Additionally is the application written for dual socket CPU's? while it may run on a dual socket machine it may not run well, I believe it was George Nevill-Neil who mentioned you would normally need to re-write code to correctly take advantage of multi socket systems, additionally I believe you would then need a Windows Server licence instead of something like Pro (Don't quote me on that) unless you are using Linux.
I've asked the author whether he would hypothetically expect the application to run well on a dual-socket configuration vs a single-socket configuration with the same spec. I will see what he says.
I am running the software under Linux.
I've been thinking that a config with 512GB RAM might be very expensive to own (even used). One solution that was proposed to me was to build some sort of HEDT system that can cover 80-90% of the workload (using say 128GB RAM) and then outsource anything that requires more RAM than that to the cloud. That might be the most cost efficient solution and also give me good performance, because honestly it's hard to get a server with 128GB RAM that get's close to the performance of something like an overclocked x99, x299 or x399 system for the software I'm running (I have no proof of this, but based on the evidence at hand, it seems plausible.). I really wish I could build a Ryzen 1800x system, but the 64 GB RAM cap is just too low sadly.
If you can afford the cost I would use thread ripper if possible, I have a feeling Intel is basically bricking it at the moment and thread ripper may just steal the show this year.
But yes the best way will be to wait until the developers respond to you and then we can help you finalise the build.
Yea, I am contemplating waiting for Threadripper. I typed out a whole post defending Intel, but then as I reached my conclusion, I sort of agreed that Intel is indeed bricking it a bit, simply because of what is going on with their release date for their 12-core and higher parts.
FYI, the software I am using is not optimised for dual-socket configurations. It is unclear how much worse it would perform compared to a single-socket equivalent spec CPU though. Could be quite a bit potentially.