Dual E5-2679v4 Workstation Build

Joined
Aug 2, 2016
Messages
25
Hey Forum readers,

Since people here seem to be more knowledgable about hardware than most enthusiast forums, as opposed to the typical slew of posts from 15 year olds too lazy to use a search function having difficulty overclocking unlocked CPUs on a "leet pr0 gaming" motherboards, I figured I'd ask this question here.

I'm currently trying to build a workstation whose primary purpose will be doing calculations and maybe some data analysis/machine learning in the future. The programs used generally scale pretty well out to 48 threads or so, but benefits a lot from higher clock speed. As a reference, the current system is a dual X5660 setup on an EVGA SR-2 @ 4.21GHz with 80GB of RAM (setup with much help from the guides on this forum). Unfortunately, this is now too slow for my needs, and time on this machine is now also split with other computing needs as well.

Most potential replacements out there don't seem to offer enough improved performance to justify the cost, such as the 2687w v4, and most of the newer Xeons seem to offer many cores but low clockspeed. The 6950x is exorbitantly priced for what it offers (only 10 cores and RAM limitations), and if reports of degradation turn out to be common, not feasible for long term use. I've even considered getting a used sandy bridge-ep setup for cheap as a stopgap to at least give me some dedicated computing power.

As a long term replacement, I've been thinking about getting the Xeon E5-2679v4 in a dual setup on a workstation board like the supermicro X10DAX or the asus Z10PE-D16 WS. I'd like to use at least 256 GB of ram and have the PCI-E slots for potential compute acceleration in the future. SSD interfaces are not as important - SATA SSD will be used as the scratch directory and a regular HD used for storage. The E5-2679v4 has 20 cores and 3.3 GHz all core boost, which would be ideal for my uses. It also can be found for reasonable prices (albeit not new), as low as $1800. Unfortunately doesn't seem to be all that common (not sold at retail, seems to be designed for eBay custom use), so I haven't found any reports of it used in a dual socket configuration, although there are some people using it in X99 enthusiast boards. I am worried a little about the 200w TDP, although more from the power delivery standpoint (especially since enthusiast boards are usually designed for greater power consumption than design, while the workstation boards may not be) since heat can be ameliorated with water cooling or a good air cooler. Does anyone here have experience with this chip, particularly in a dual cpu setup? If not, do you have any other recommendations for processors?

Many thanks in advance for your advice and help.
 
If power consumption isn't an issue I'd look at using multiple dual E5-2670v1 servers for your data analysis setup. Any CUDA accelerated machine learning you do wouldn't be very CPU dependent and a separate workstation could be used for that purpose.
 
If power consumption isn't an issue I'd look at using multiple dual E5-2670v1 servers for your data analysis setup. Any CUDA accelerated machine learning you do wouldn't be very CPU dependent and a separate workstation could be used for that purpose.

Power consumption isn't a major issue but of course I'd like to avoid being blatantly wasteful.

Unfortunately, the scientific calculations I'm looking to do (which are done on several existing software packages) don't scale very well across multiple nodes. I suppose could just run multiple calculations across the 2670 servers (which admittedly are currently very cheap), but there are some longer calculations (that take 2+ days on the SR-2, which itself is probably about 80% as fast as a dual 2670 build) that could benefit from having even more CPU speed.
 
Last edited:
I'd highly recommend looking into a CUDA build for machine learning. Just slapped together a 4 gtx 1080 build for work and we are training models in ~3 minutes vs 3 days on the 5930k in there...
 
I'd highly recommend looking into a CUDA build for machine learning. Just slapped together a 4 gtx 1080 build for work and we are training models in ~3 minutes vs 3 days on the 5930k in there...

Which machine learning packages are you running on the GPUs, or are you writing your own code?

The machine learning is a secondary purpose to the scientific calculations, which are running on programs that demand CPU power first and foremost.

The first 2679v4 arrived and is working well in a X99 board with some non-ECC DDR4. An air cooler (the cheap Swiftech Polaris) handled the 200w TDP processor for a long Linx run with AVX, so it's not worth the risk of water. Its telling that systems needing custom loops must really be dissipating huge amounts of power. Waiting on the X10DAX to arrive.
 
I just noticed this thread, OP if you see this any chance you could list the turbo bins you observe? I know allcore is 3.3 which is impressive for 20 cores but really curious what the max 1/2/4 core bins are.

Also this is kind of out there, but any chance you tried turbo boost 3.0 (w/ the intel utility) on a windows image? Kind of a novelty but at least you will find out which cores are your "best" ones :)
 
Unless you really need V4, last gen Xeons can had dirt cheap on ebay.
 
the 2679V4 is special because it has extremely high all-core turbo, which can be locked on consumer boards like the R5E
you get > 2 6950X's worth of computing power on all threads with sufficient 4-thread power for gaming, which is really rare these days.
 
the 2679V4 is special because it has extremely high all-core turbo, which can be locked on consumer boards like the R5E
you get > 2 6950X's worth of computing power on all threads with sufficient 4-thread power for gaming, which is really rare these days.

The 2679V4 has a nice all core turbo, but really crappy single threaded turbo...maxing out at 3.3GHz. This is only a fraction above the maximum all core turbo of 3.2GHz. The chip is a nice example of what Intel could do if they wanted to...they could have four cores at 4.0GHz without too much hassle, with the rest at 3.2GHz. Dell didn't require it, though. 3.3GHz is insufficient for gaming, IMHO...

What I'm interested in is a 260W Socket 3647 chip that doesn't have a crappy, self defeating FIVR to worry about...the clocks should be quite healthy on such a chip.
 
The 2679V4 has a nice all core turbo, but really crappy single threaded turbo...maxing out at 3.3GHz. This is only a fraction above the maximum all core turbo of 3.2GHz..

Can confirm this. The 2679v4 can only go up to 3.3 ghz single core, you can maybe squeak out another 100mhz (so around 3.4 ghz) through BCLK overclocking - my preliminary efforts show you can go up to 104.5 mhz (a 4.5% increase from stock). Not particularly interested in overclocking though, because of the instability it'll introduce and the minimal gains achievable with a limited BCLK and a locked multiplier.

I'm not gaming on this system though - I need the multithreaded performance. The all core turbo goes up to 3.2 GHZ as lutjens pointed out.
 
The 2679V4 has a nice all core turbo, but really crappy single threaded turbo...maxing out at 3.3GHz. This is only a fraction above the maximum all core turbo of 3.2GHz. The chip is a nice example of what Intel could do if they wanted to...they could have four cores at 4.0GHz without too much hassle, with the rest at 3.2GHz. Dell didn't require it, though. 3.3GHz is insufficient for gaming, IMHO...

What I'm interested in is a 260W Socket 3647 chip that doesn't have a crappy, self defeating FIVR to worry about...the clocks should be quite healthy on such a chip.

3.3GHz is...all right, but I agree, it is pretty pathetic in a day when even lowly laptop processors turbo up to 3.6.

So fortunately Skylake nominally decouples the PCIe clock from the base clock; in theory this means we could finally have overclocking on dual socket boards again. Whether Intel will let it happen is a different question (they evidently didn't like Asrock's non-K OC feature and were able to threaten them enough to make them remove it). The situation is even worse on Xeons because they are heavily segmented by performance; I could easily see a 12c part being overclocked to outrun the fastest 28c part, which has a stupidly low clock speed to be able to stay under 165W.
 
3.3GHz is...all right, but I agree, it is pretty pathetic in a day when even lowly laptop processors turbo up to 3.6.

So fortunately Skylake nominally decouples the PCIe clock from the base clock; in theory this means we could finally have overclocking on dual socket boards again. Whether Intel will let it happen is a different question (they evidently didn't like Asrock's non-K OC feature and were able to threaten them enough to make them remove it). The situation is even worse on Xeons because they are heavily segmented by performance; I could easily see a 12c part being overclocked to outrun the fastest 28c part, which has a stupidly low clock speed to be able to stay under 165W.

The problem is that all the motherboard makers came out with BCLK overclocking way too early and made it pop up front and center on Intel's radar. It also won't surprise me to have the 1S WS parts locked out from the enthisiast X chipsets, exactly like what happened with the E3 V5 chips.

A crying shame...as the FIVR-less HCC Skylake-E chips would have been decent overclockers.
 
Back
Top