Those cheap Xeons... dual itch building up

[porina]

The above thread is starting to sink into my mind...

 

Based on the article in there one of the Asrock EP2C602 series new would seem a low cost low risk option. There doesn't seem to be any cheaper dual socket used mobos I can find, and they're mostly weird form factor and overall higher risk. The Asrock ball park price is £250, and I need to work out what the differences between the variants are for the cost.

 

A pair of E5-2670 CPUs can be picked up for £170 from ebay, less than I paid for either of i5-5675c or i5-6600k. 

 

Two of quad channel ram could be cheaply filled by 4 of 2x4GB Kingston HyperX 1866 ram kits, although I will also look at ECC ram as I think this would also double nicely as a file server.

 

PSU requirements will likely have me getting a similar model to that in my main rig, a Corsair HX750i, after I double check it can supply TWO 8 pin mobo connectors. This should be enough to power both CPUs, the ram and a single meaty GPU. It wont be a gaming box, but I have an old R9 280X I can recycle with enough power for most things.

 

OS storage will be some cheap SSD, and stack it up with my existing spare HDs for bulk centralised storage. I know running power might suck a bit for a storage box, but I'll likely load the CPUs 24/7 anyway.

 

Will need to find a case big enough to fit both the mobo and a ton of HDs.

 

I haven't looked at coolers at all yet, but will most likely get some mid range air cooling.

 

I still need to do some further calculation into the CPU potential performance vs. newer generations. For Prime95 like tasks I find not having FMA instructions to be about 33% raw performance penalty so in theory two of these CPUs might only perform comparable to a single 5820k overclocked to 4.6 GHz. Of course, for non-FMA loads the Xeons should have much more potential.

[Samfisher]

Use case?  The  E5-2670 are pretty damn slow clockspeed wise.

[porina]

12 minutes ago, Samfisher said:

Use case?  The  E5-2670 are pretty damn slow clockspeed wise.

Heater, file storage, distributed computing, astrophotography processing, because I can.

 

Yes they are slow, but so many cores... in Prime95 (which can use FMA instructions this doesn't have) I predict performance is comparable to a single 5820k overclocked to 4.6 GHz. For non-FMA applications it would be 50% faster.

[porina]

Ok, I've done some calculations for performance per watt, and bang for buck. This does rely on many assumptions and I am basing these on what's more important to me.

 

I'm going to compare 3 hypothetical systems.

 

1, i7-6700k at stock (4 GHz, 4 cores) £290

2, i7-5820k at stock (3.3 GHz, 6 cores) £330

3, E5-2670 x2 (2.6 GHz, 16 cores) £170 for two

 

Clocks given are base clocks, as I'm assuming under full load that is what they'll run at. I've built a hypothetical system around the above, picking a low cost mobo for each (for the Skylake system, I went Z170A, although non-OC boards would be cheaper).

 

Ram for #1 is 2x8GB DDR4-3200 for £100, 2 4x4GB DDR4-3200 for £100, for #3 8x4GB DDR3-1866 non-ECC for £120. I'm assuming the latter runs at 1600 but it costs no more to get the higher speed. Note I went for 8 sticks, to allow both CPUs to run at their full potential with quad channel memory. Running less will cripple it and negate the point of considering such a system.

 

Cooler wise I'm assuming high end air at £50 each, so double for the two Xeons.

 

Common components will be a Corsair RM series, probably 750W at around £100, allowing same again for a random case, and £40 for a basic 120GB SSD to get an OS on. I'm not including OS cost or GPU, the assumption being these systems will be used for compute heavy tasks only that do not depend on GPU availability.

 

Metric 1, performance per watt. The Xeons only support AVX. I will assuming all CPUs here are equal in IPC, which is probably not true, but I don't have personal data in this respect. Further assuming it isn't ram speed limited, you can simply multiply clock and cores. No surprises, the dual Xeons win by miles here, at over double the combined clock of the Haswell-E, which in turn is a bit higher than the Skylake system. Factor in the CPU TDP (ignoring rest of system power), it isn't so clear cut. By my scoring system, 6700k scores 176, dual Xeons 181, Haswell-E 141. Although the Xeons have a lot of performance, they also use a lot of power, which balances out. Haswell-E loses out here.

 

To complicate matters, what about applications supporting FMA, which the Xeons lack? In previous testing, Haswell has 1.5x IPC compared to Sandy Bridge in Prime95 like tasks due to FMA, when not ram limited. This swings the ball towards the newer i7 systems. Here Skylake still scores 176, Haswell drops to 124 due to its slightly lower IPC I measured at 14% in this application, and the dual Xeons are down to 106.

 

This still isn't the whole story. Note I said "when not ram limited". They are. Again in the past I've done testing on the performance drop due to ram bandwidth starvation. This is why I went for 3200 ram for both i7 systems, instead of the agonisingly slow 2133 standard. Anyway, figuring in efficiency loss from ram bandwidth starvation, the scores are then 150, 122, 98 respectively. Haswell-E and Xeons didn't drop much, but they started so much lower that even the bigger drop of Skylake didn't lower it to their level. 

 

Power efficiency conclusion: for running costs Skylake wins at FMA supporting applications. For AVX, the Xeons are still competitive. Note these results may shift once diluted by system power usage. I can't model that in any useful way.

 

Metric 2: bang for buck. I can total up the previous values and reach a system cost. The Haswell-E and dual Xeon setup come within a fraction % of each other at £885 and £883 respectively. The Skylake system is a bit cheaper at £770. I can divide these values into the previously calculated effective CPU performance to get a bang for buck rating. In AVX applications, for the 6700k we score 21, for 5820k 22, and dual xeons come in at a whopping 47! Over double the other two. What about FMA? Back to Prime95, in peak scenario not limited by ram, we get 21, 20, 28 respectively. So the Xeons still take it here. And what about when ram limited? 18, 19, 26.

 

Bang for buck conclusion: Dual Xeon wins outright here with the i7's fighting over 2nd place.

 

So, my overall conclusion would be, if you don't care about power costs, just get the dual Xeons for their compute potential. If running costs worry you more than up front, then the Xeons are still a good choice for non-FMA applications. With FMA support, go Skylake. Chances are for most, the running costs are much less significant than up front costs.

 

If I still want to get a dual Xeon system to play with, I'm looking at the best part for £900 for the parts above. I'm not sure I want it that much! This is a classic case of "I could, doesn't mean I should!"

 

Oh, remember I said about using ECC ram possibly? That isn't so cheap. I haven't scoured for the lowest cost, but they're starting £24 a stick on Crucial's website so that would add an extra £72 onto the Xeon build if you wanted it.

 

I should finally add, the value of the i7 systems could be increased by overclocking them which you can't do with the Xeon. For my uses I'd get better value from the i5-6600k and overclocking it instead of the i7-6700k with or without overclock.