Overclocked i7 5960X or Xeon 14-Core E5-2697v3?

[Stephen Hughes]

Hi guys,

 

As the title suggests i'm considering whether to go for the 8-cores of the i7 5960X which can be overclocked to 4.4GHz, or buy the Intel Xeon 14-core E5-2697v3 which cannot be overclocked (i think) and has a speed of 2.6Ghz per core.

 

The price difference when allowing for the mb etc is about £1000. Is there some way of calculating the performance difference for rendering? If i thought the xeon machine would render the same scene (max + vray) in half the time of the i7, purchasing the xeon would be a no brainer, but if it was down to maybe just 25% better I wouldn't be so sure.

 

Can anyone estimate the performance difference between the two?

 

Stephen

[RyderSK]

v3 Xeons can be slightly overclocked when seated in Asus Workstation motherboard, but think 3-5perc. Certain ES(Engineering samples) have unlocked multiplier. And then you have all-core turbo feature. So in the end, even Xeons are rendering at higher frequency than the stated default. How much depends on individual model, but it's 10perc. on average.

 

V3 Xeons and i7 5960X are the same architectural family (Haswell) so you can simply multiply their cores with frequency and roughly get your comparison.

 

8x4,4= 35,2GHz

2x14x3,0 = 84GHz

 

So in ideal conditions, the Xeon configuration is twice as fast in multithreaded performance. I am saying "ideal", because ray-tracers don't scale linearly, although they scale pretty well today, it's not 100perc. scale, but more like 90-98, depending on renderer. Than there is issue with processor groups and 3dsMax architecture because 72 threads is already beyond it's original limit, and it took some time for certain renderers (Vray and Corona namely) to be able to properly use it without tricks (like running distributed on single PC). Vray seems to have solved this in most recent service pack, but I didn't test.

 

Sometimes it's therefore worth not to go for the most cores models, but the ones with less but higher frequency and most of all, highest all-core turbo. Think like 2667v3 for example,etc... but I don't have any table of comparison at hand.

 

User Numerobis created such table :- ) Germans... maybe he can post it here or you could find it in other thread here.

[VisCorbel]

Did you mean getting a workstation with just single xeon cpu?

If so, I think an overclocked 5960x is a more sensible choice.

 

Xeons start to make sense when you go for dual cpu setups, as there is no other way to get all that renderpower in a single pc.

[numerobis]

User Numerobis created such table :- ) Germans... maybe he can post it here or you could find it in other thread here.

 

Of course

 

The basic system consists of case, PSU, mainboard, cooler. No GPU or SSD.

The 4.4GHz on air for the 5960X will not be possible normally, so in this case there should be added maybe 100€ more for a water cooler. And depending on the chip you can gain a few percent more performance from cache overclocking. So in reality the values for all overclocked CPUs should be a bit better if it is done properly.

 

Components:

 

Single CPU:

ASUS X99-A/USB 3.1

Kingston HyperX Fury DIMM Kit 32GB, DDR4-2400, CL15-15-15

Noctua NH-D14 SE2011

Lian Li PC-7HX

Seasonic X-Series X-750 KM3 750W ATX 2.3

 

Dual CPU

ASUS Z10PE-D16 WS

2x Crucial DIMM Kit 32GB, DDR4-2133, CL15-15-15, reg ECC

2x Noctua NH-U14S

Lian Li PC-A76X

Seasonic X-Series X-750 KM3 750W ATX 2.3

 

[/img]

Edited October 25, 2015 by numerobis

[Nikolaos M]

The basic system consists of case, PSU, mainboard, cooler. No GPU or SSD.

Not if the cpu doesn't have an integrated gpu. The 5960X and the Xeons doesn't, so a gpu is essential.

[numerobis]

There is also no CPU that includes a SSD...

[RyderSK]

lol

 

think I am guilty of some reading comprehension as well though :- ) As Austris correctly guessed, the OP might have simply been comparing single Xeon from 2xxx series (aka "apple workstation" ...), something I would never consider so the thought didn't even cross my mind.

Although the price difference should have given it off, I simply brushed that away that maybe some really good deal or ES parts.

[numerobis]

I had considered for myself to put a single 10- or 12-core in my workstation. But the problem is the low core frequency and regarding multicore there is almost no difference (if any) between a 12-core 2690v3 and a 5960X@4.5GHz/4.2GHz cache. I had found some ES chips on a chinese website, but i didn't trust them and so i took the easy route with a pretested 5960X...

Looking only at the multicore performance and not the single core speed, a single 14-core E5 would be an option to get more power in one box - but it would be almost 70% more expensive and for this money i can buy another 5820K.

Edited October 15, 2015 by numerobis

[Nikolaos M]

There is also no CPU that includes a SSD...

I don't know what's funny about what I wrote, but I don't agree with the logic behind the charts you posted. You calculate a system's total cost and based on that, you find the €/GHz ratio. But these numbers could be totally different if you had chosen other parts (why an Asus-A mobo, why a Lian Li case, why a d14 as a cooler and why a seasonic 750 as a psu? aren't there cheaper choices that could alter the total cost and consequently the last €/Ghz ratio ?). And also, why aren't a gpu and a ssd/hdd taken into account when you calculate a basic system's total cost? Aren't they essential to run a system?

 

Sorry about this post, but I find some answers ironic and pejorative...

Edited October 16, 2015 by nikolaosm

[numerobis]

It was not meant pejorative and i didn't make fun of it - this was a ;-)

 

First of all... this is all based on a chart that i made FOR MYSELF to find the best solution for my workstations AND render nodes.

The initial single CPU configurations doesn't include cases and PSUs so i added some components to match the other config. Sure, i could have chosen cheaper parts in some cases, but this depends on the usage and personal preferences and i wanted to keep both setups comparable and more workstation oriented because i thought this was what the OP asked for (but i don't know).

I could add a SSD, but in reality i would choose different SSDs for ws and node. The same applies to the GPU. Or maybe someone needs a bigger SSD or HDD in his ws? You can also add an OS and other licenses... The next thing you could say is that there are different prices in different countries and maybe not all parts available everywhere.

I could also simply take some generic numbers of maybe 700€/900€ for single and 1300€ for dual, but would this be more precise? And as i said, in reality i would overclock the cache too, which also gives you different numbers.

 

This chart should be a rough comparison and everyone who needs a more precise calculation should do it for himself with his own parts and prices!

 

I could make different systems for ws and node, but then this chart would be much longer incl. the two RAM options. But maybe i'll do this... (if anyone is interested)

Edited October 16, 2015 by numerobis

[Nikolaos M]

I was mislead by the phrase "basic system". Seamed like a general purpose chart and not one for custom system. It would be far more objective and useful if the €/Ghz ratio was calculated by the cpu cost solely, so that anyone could add the cost of the rest of his personal build by himself and make the calculations accordingly.

[numerobis]

But the mainboard, case and 64GB RAM (and another cooler) makes a big difference when you compare dual CPU to single CPU.

[Nikolaos M]

Ram is the only true systemic difference imo, because it always has to be double in amount in a dual Xeon system for comparison reasons. There are cases that are ssi - eeb compatible and aren't so expensive, like the Phanteks enthoo pro, for instance. As for the coolers, there must be two of them, but not expensive. A 25-30€ air cooler could keep a Xeon cool enough during rendering. And lastly the motherboard. The dual socket boards are expensive alright, but I think their price start from 350-400€ I think. But there are plenty of 1p x99 mobos in this price range, so it's not a difference of a systemic type I think.

Edited October 16, 2015 by nikolaosm

[RyderSK]

Ram is the only true systemic difference imo, because it always has to be double in amount in a dual Xeon system for comparison reasons.

 

And why would this be the case ?

 

While it's true that for multi-socket CPU system you need double the amount of memory modules to have equal channel setup, it's not even needed in reality to properly feed the threads.

If you feel quad-channel makes difference to your work tasks (it doesn't), but only have budget for 32gb of ram for example, you would simply spec it as 8x 4gb modules, as opposed to 4x 8gb modules in single-socket quad-channel setup. Result would be the same.

 

But dual-socket systems definitely don't need double amount of memory for any comparison, this is completely false statement. Unless this is simply confusing wording on your part.

 

 

There are cases that are ssi - eeb compatible

 

SSI-EEB is identical enough in layout and physical size to simple E-ATX, with exception of single bolt position when it comes to mounting and that is freely to be ignored ;- ). Every server board therefore fits into regular XL Tower (today even some mid towers which forfeited optical bays).

 

I would have known since I built few into regular Fractal XL ;- ) Just like thousands others have done.

 

Nonetheless, currently all the boards intended to workstation are E-ATX (or even ATX now). So no need to bother with this, since workstation boards (like Asus Z10PE-D8 /D16, etc..) are superior for WS purpose compared to server boards (like Supermicro series, though even those now offer atx layout directly).

 

 

A 25-30€ air cooler could keep a Xeon cool enough

 

Cooling two up to 165W top models costing up to 2500 euros each using 20 euro cooler is possible and employed in hall cooled racks all the time. But for home/office workstation maybe not the best idea. At this level of workstations, no matter the CPU setup, you want the same reliable and silent operation.

Doesn't mean it needs NH-D15/ custom water loop to run well, but haven't seen anyone go for less than AH-14 or similar.

 

I think your over-analyze the chart from single direction only. I think it was obvious Numerobis's chart was simply comparison made for himself from which others can draw, but don't necessarily fit to all questions.

 

Storage and GPU needs can vary wildly (from simple node to full-scale workstation, it can be difference up to few thousands of euros), but the stated parts in his "basic setup" are essential, and can only vary in quality or brand, and don't command much price difference.

 

Yes GPU is essential, but is it passive 750 GTX, or M6000 ? Would depend on needs and purpose right ? But 32GB of memory, quality /Motherboard/PSU/Case/Coolers would be there always and tie much closer to the core of system than the overall build itself. Which is why they are essential in broader comparison than full systems would have been.

 

That doesn't negate the usefulness of full system comparison, which could simply be added imho to his chart for completness sake. But it's neither illogical the way it is now.

Edited October 16, 2015 by RyderSK

[Nikolaos M]

Storage and GPU needs can vary wildly (from simple node to full-scale workstation, it can be difference up to few thousands of euros), but the stated parts in his "basic setup" are essential, and can only vary in quality or brand, and don't command much price difference.

 

Now, "this is completely false statement. Unless this is simply confusing wording on your part", if you allow me to use your words.

 

Prices in motherboards, pc cases, psu's, cpu coolers and RAM can vary widely and could result in a price difference margin that could reach 2000€. And don't say that I exaggerate, because this example is what I call exaggeration:

Yes GPU is essential, but is it passive 750 GTX, or M6000.

If a Quadro M6000 is a common choice (so common, that we could bring it as an example here), then a 500$ pc case, a 600-700$ custom cooling loop and a 700$ motherboard are not so rare too...

 

As for the RAM part and what you said about the chassis compatibility, I respect your experience (I've read some reports about the opposite, in the case of a Corsair 750D) but what do you thing about nomerobis's statement: But the mainboard, case and 64GB RAM (and another cooler) makes a big difference when you compare dual CPU to single CPU.

Is this a false statement too? Or just confusing words?

[RyderSK]

The difference is you always need decent PSU. The fact that you can get decent PSU in gold quality or double the price in platinum is already stretching the argument because neither serves different purpose and in both case it's essential to the build. Numerobis chose some benchmark model. It was only example.

He chose completely well specced, but reasonably priced standard components that would well represent all type of builds. But you can't quite do that to the same level for GPU and storage options (although to lesser extent, with external storage being common choice today even for single freelancers, it applies more for the GPU), so he just omitted.

 

Personally, I wouldn't build the chart the way he did ( I would never build a chart because I am hyperactive unfocused person), I would compare full builds the way you would like and simply add some common choice like GTX 970 and 256-512GB SSD. But that doesn't make his chart unreasonable and worth nit-picking imho :- )

 

I just don't see why you chose that battle and find it so odd. Not gonna comment further on the "750 gtx vs M6000" because you kind of lost me there.

 

Regarding the memory I might agree with you too. I looked in the chart again and I see it specced 64 for dual-build but only 32 for single build. While it's common place to do so, no, it doesn't make sense to me.

Edited October 16, 2015 by RyderSK

[Amen]

I think the list is very helpful. In my opinion it is good choice to leave out gpu and ssd. One could ignore the case too, but the 100 € price difference in the calculation should not amount to a big change in €/Ghz.

 

I think it is realistic to assume twice the ram for a dual core machine and the cooler is high end but because it is more or less the same just two of the same for a dual system it calculates correctly for the chart. Using the same power supply helps the dual system but this again is marginal.

 

Now, if you add two items that strongly depend on personal choice and need, it would only confuse and destroy the simplicity of the chart, which is very helpful.

 

@Numerobis: Thank you very much.

 

@Nikolaos: I can see your point, but even though the chart could be more complex, the general tendencies the €/Ghz numbers are displaying are correct and a good starting point to choose the direction one wants to go in building a sytem.

[numerobis]

Ok... so here is a BIGGER version

 

Configurations:

 

Workstation 1 CPU 32GB/64GB

ASRock X99 Extreme6/3.1, Noctua NH-D14 SE2011, 1x/2x Kingston HyperX Fury 32GB DDR4-2400 CL15-15-15, Samsung SSD 950 Pro 256GB M.2 x4, Gigabyte GeForce GTX 970 WindForce 3X 4GB, Seasonic X-Series X-750 KM3 750W, Lian Li PC-6B

 

Workstation 2 CPU 32GB/64GB

ASUS Z10PE-D16 WS, 2x Noctua NH-U14S, 1x/2x Crucial 32GB DDR4-2133 CL15-15-15 reg ECC, Samsung SSD 950 Pro 256GB M.2 x4, Gigabyte GeForce GTX 970 WindForce 3X 4GB, Seasonic X-Series X-750 KM3 750W, Lian Li PC-A76X

 

Render Node 1 CPU 32GB/64GB

ASRock X99 Extreme4/3.1, Noctua NH-D14 SE2011, 1x/2x Kingston HyperX Fury 32GB DDR4-2400 CL15-15-15, Samsung SSD 850 Evo 120GB, AMD Radeon HD 5450, Seasonic G-Series G-550 550W ATX 2.3, Cooler Master Elite 370

 

Render Node 2 CPU 32GB/64GB

ASUS Z10PA-D8, 2x Scythe Ashura, 1x/2x Crucial 32GB DDR4-2133 CL15-15-15 reg ECC, Samsung SSD 850 Evo 120GB, AMD Radeon HD 5450, Seasonic Seasonic G-Series G-650 650W ATX 2.3, Nanoxia Deep Silence 2

 

 

 

As i said before, cache overclocking can give some more improvement for Haswell-E. For example the performance of a 5960X@4.2GHz with a cache of 4.1GHz is increased by 6,3% compared to 4.2GHz/3.0GHz in my benchmarks (Maxwell Render). This would be comparable to 45,0€/Ghz for the render node or roughly 200MHz more.

But you need a board with OC-socket to go higher than 3.5GHz, which would be only ASUS or the second revision of ASRock boards (/3.1) atm as far as i know.

Edited October 25, 2015 by numerobis