5960 or 6900

[adom22]

Hey guys

I'm currently planning a new build around an 8 core x99 setup. I'm really torn between the 5960 and the 6900 with the 5960 generally being a better overclocker. I will mainly be using it for fluid simulations, vfx simulations, rendering and compositing. Main thing is to get plenty of SSD storage in there and fast memory for the fluid sims.

Obviously the 6900 is newer and overclocks to a modest 4.2 (on average) looking at the reviews. Wondering if anyone has any input on this? Will the PCI lanes help with SSD's or is it mainly a help when having multiple GPU's and SLI etc.

Thanks for any help!

Adam

[EndlessWaves]

HWBot reports are 4.4Ghz average on air:
http://hwbot.org/hardware/processor/core_i7_6900k/

Have you looked at a dual-CPU Xeon system? You may find that delivers more memory bandwidth and more performance at the same price if your workload scales well to lots of cores.

[adom22]

HWBot reports are 4.4Ghz average on air:
http://hwbot.org/hardware/processor/core_i7_6900k/

Have you looked at a dual-CPU Xeon system? You may find that delivers more memory bandwidth and more performance at the same price if your workload scales well to lots of cores.

Morning, thanks for that Well I was using a 3930k overclocked system and now had to switch to a dual E5-2620 machine but they tend to be slow in fluid simulations in 3D packages, which is what I do a lot of.

The i7 setup trumps it even with rendering so I don't have much confidence in the Dual Xeon setup for the same price range. I think it would work well if more money was thrown at it but £3k is my budget for a complete system (minus the monitor, keyboard etc).

I think a high frequency and fast memory with lots of SSD storage will be the way to go.

[scaryjim]

... I was using a 3930k overclocked system and now had to switch to a dual E5-2620 machine but they tend to be slow in fluid simulations in 3D packages ...

Intriguing - clearly shows your bottleneck isn't core count, so perhaps pushing for the greater core number isn't the way to go? You've either got a single thread bottleneck (most likely scenario, IMNSHO), a memory bandwidth problem (on a quad channel system? Scary if that's true), an IO issue (did the i7 and Xeon systems have the same storage subsystem?), or (least likely) the code uses an instruction that's accelerated in Ivy Bridge but not in Sandy Bridge.

OK, that's a lot of waffle, but what it boils down to is needing to work out why the Xeon machine is slower - if it's a single thread bottleneck then you should really go for the fastest stock clock speed available (i7 6850k); if it's bandwidth then you need the bandwidth of DDR4 (but might not need the extra cores/clockspeed), if it's IO then you probably want the largest memory capacity available (which should be a fully populated dual Xeon system), and if it's instructions then anything newer than Sandy Bridge should be better.

The other thing to consider is whether the work can be distributed - if the fluid simulations can be run separately on another machine you could get a cheaper consumer-grade rig which will have much higher clockspeeds, just for running the fluid simulations.

Also - is this work, side projects, just for fun, potential future career, university/training/education? Each situation would make me look at the problem differently...

[adom22]

Intriguing - clearly shows your bottleneck isn't core count, so perhaps pushing for the greater core number isn't the way to go? You've either got a single thread bottleneck (most likely scenario, IMNSHO), a memory bandwidth problem (on a quad channel system? Scary if that's true), an IO issue (did the i7 and Xeon systems have the same storage subsystem?), or (least likely) the code uses an instruction that's accelerated in Ivy Bridge but not in Sandy Bridge.

OK, that's a lot of waffle, but what it boils down to is needing to work out why the Xeon machine is slower - if it's a single thread bottleneck then you should really go for the fastest stock clock speed available (i7 6850k); if it's bandwidth then you need the bandwidth of DDR4 (but might not need the extra cores/clockspeed), if it's IO then you probably want the largest memory capacity available (which should be a fully populated dual Xeon system), and if it's instructions then anything newer than Sandy Bridge should be better.

The other thing to consider is whether the work can be distributed - if the fluid simulations can be run separately on another machine you could get a cheaper consumer-grade rig which will have much higher clockspeeds, just for running the fluid simulations.

Also - is this work, side projects, just for fun, potential future career, university/training/education? Each situation would make me look at the problem differently...

It is part of work To be honest the dual Xeon machine was just purchased as a generalist machine so it was never bought with simulations in mind, both machines were just writing to standard sata drives. It is the main reason why I am looking to build a new one with faster storage and memory.

It is tricky with full core utilisation. Generally the largest bottleneck of fluid simulations in 3D apps is the IO. More cores do help of course If I went the dual Xeon route then I would want to have a lot higher base clock of 2.1 for example but the costs to get Xeons of 8 or more cores at a high clock speed is a bit crazy for my budget.

Unfortunately the apps I use don't support distributed sims (it is on the developers to do list). I know they work in 3D apps such as Houdini which I have also used but we don't use it in the pipeline at work.

Here is the report from CPUID for the dual Xeons, I suspect it is more an IO issue with the memory/storage speed to be honest.

https://www.dropbox.com/s/g1c9ke3i1k...CPUID.txt?dl=0

The i7 machine was writing to a raptor drive (very old!)

[scaryjim]

Hmmm, a 10k raptor could well make the difference for IO purposes. Did the i7 have 32GB of RAM too? Be interesting to know the physical memory usage - I encountered an issue many years ago working with some data analysts where a simple sort of a large dataset (in SPSS) took minutes to complete. One of the senior analysts got a new machine (funded by a research project they were working on) which had ... hmm, might have been as much as 16GB of physical RAM, and the same sort started completing in a few seconds. He turned on system monitor, and found he was hitting over 12GB of physical RAM usage - which explained why the other machines (4GB) were so slow - most of the sort was occurring in swap file and hitting the OI hard.

So I kind of find these bottleneck problems intriguing to diagnose: sometimes you can end up with some surprising answers and unexpectedly large gains in performance.

Anyway, the reason I asked work is that I'd be very cautious about overclocking a work machine to the ragged edge...

[adom22]

Good point, I wont be doing any extreme tweaking with the voltages Yes the i7 had 32gb DDR3, cant remember the speed of it. I'm obsessed with keeping things well within memory limits as causes hell on the render farm! Some machines on there got 12gig memory still, nightmare.

Anyway I could get faster drives and memory for this dual xeon along with a couple of new Xeons maybe... it's definitely worth thinking about. Depends what this motherboard is like I suppose.

[EndlessWaves]

if it's a single thread bottleneck then you should really go for the fastest stock clock speed available (i7 6850k)

When you're overclocking I would suspect the 6900k with two cores disabled would prove faster as the binning generally improves as you go up the range, and the extra cache wouldn't hurt (or does that get disabled with the cores? Not sure on that one).

Although of course given the big price hike it may not make sense.

or (least likely) the code uses an instruction that's accelerated in Ivy Bridge but not in Sandy Bridge.

The 3930k is Sandy-E, Intel bump up the numbers by a thousand for the HEDT platform.

If I went the dual Xeon route then I would want to have a lot higher base clock of 2.1 for example but the costs to get Xeons of 8 or more cores at a high clock speed is a bit crazy for my budget.

Wikipedia has a handy list:
https://en.wikipedia.org/wiki/List_o...icroprocessors

Something like the Xeon E5-2630 v4 may be of interest. 10 cores, 2.2Ghz base, 2.4Ghz full core turbo, around £650 each.

[adom22]

Thanks for that, I really am considering purchasing a new dual board, Xeons, memory and SSD storage.. maybe a pair of 2640's. I think they will be great in the multi-threaded apps but I use 3DS Max a lot and most of it is still single threaded so that will be fun waiting for it to think! Hopefully could re-use the case, gpu, psu from the current dual setup.

Do the Broadwell-EP Xeons fir into the Haswell server boards (it does list them as accepting 2600 v4's).

Anyway thanks for the info guys, making me reconsider the i7 setup!

[adom22]

Decided I'm going to go for a whole new i7 build. Speaking with the devs of the software I use for fluid simulations and they don't recommend the dual setup as it doesn't scale very well. Most likely going to go for the 6900k so I then have the option to upgrade to 128gig memory in the future.