wimver wrote:When buying a new render station, would it be rather 1 big box with a powerful 16 core processor or 2 or 3 smaller ones?
AMD or Intel R Xeon?
server motherboard or normal one?
multi threading processor or not?
how much ram per core?
on the motherboard, should I consider an option for a good video card (should NL change to GPU rendering)?
what is the best option for power supply and cooling?
does it render faster with a SSD?
should I consider other options?
tnx
wim
Regarding AMD vs Intel, on the benchwell V3 list there is a 16-core (16 threads) Opteron 6378 machine rendering the scene in 10m 46s, with a benchmark of 432.01, which is exactly the same result as a 4-core (8 threads) i7-4771 machine of mine, just next to it in the list. The 6378 was an $860 USD processor though, where the 4771 is only $315 USD. And if you look, that is the fastest single-CPU AMD machine on the list, with the only faster ones all being 2- or 4-CPU machines, using 6-, 8-, 12-, or 16-core CPUs, to arrive at 24, 32, 48, or 64 threads.
As far as multi-CPU vs single-CPU, judging by benchwell V3, it appears that with AMD you have to use multiple CPUs, if you want to get a sub-11 minute benchmark. The next few AMDs slower than that 6378 are single 8-core (8 thread) FX8350 machines, all overclocked, running in the 11-13 minute range. Single Intel CPUs, on the other hand, are benchmarking all the way down into the low 7-minute range for overclocked i7-3930Ks (for comparison, I have a non-overclocked 3930K machine that benchmarks just 2 seconds faster than my 4771 machine, at 10:44), and into the 6-minute range for the fast stock Xeon 8-cores (e.g. you can find where a single 8-core E5-1680 v2 is beating a machine that has four 6-core AMD 8431s).
Regarding multi-threaded CPUs, the question really only applies to using Intel i7 vs i5 hyperthreaded/non-hyperthreaded, and my answer would be that with Maxwell, you should currently use i7s with hyperthreading enabled. I had not actually tested this in quite some time, so I just did; using an i7-4930MX, the time with HT enabled is 11:01 with a benchmark of 422.21, while with HT disabled, the performance degrades to 15:25 with benchmark 301.77.
The amount of memory depends not on how many cores you have, but on how much memory your scenes typically require; grass, hair, etc can take a lot of memory, so you'd need to do some testing to find out what you typically need.
GPU-wise, I'd not tie up my money in GPUs, since a) they're not useable at this point in time, since b) you don't know whether to purchase NVIDIA or AMD, and since c) cost per performance always goes down -- if/when Maxwell was to support GPUs, you could add cards to your existing machines to leverage the capability.
Power supply and cooling I can't comment on much, as I don't overclock my machines. Typically, I purchase a so-called "bare bones" machine, which already has motherboard, power, cooling, switch wiring, etc, figured out.
On SSDs, they are good, since you need to load resources from disk, and write large MXI files throughout the rendering process; I personally always buy two SSDs half the size I need for a drive, and pair them in RAID 0 (with a proper backup plan in place). On that disk, I would run the OS, programs, and the rendering process, with a large spinning disk being used only for backup and archive.
All that said, there is no definite answer on which approach is best. With a multi-CPU machine, you only have to purchase memory, drives, etc, once, and only have to maintain and back up that one machine. On the other hand, you could build a few cheaper i7-based machines now, and then add more as it becomes economically feasible to do so -- since Maxwell can cooperatively render across any number of machines, together they'll eventually do the work of that multi-CPU machine, or even more, but along the way, you will buy all those bits of hardware multiple times, and thereby increase the likelihood of hardware failure by the total number of machines. On the other hand, such a failure in the multi-CPU machine takes out your rendering capability entirely, where with multiple machines, you could lose one, and still get an important job out.
Here are a couple of examples that come in at or below the $3400 USD budge you mention in your other thread; first, a dual-Xeon machine:
- (1) SuperMicro SYS-7037A-i Mid-Tower
(2) Intel Xeon E5-2630 v2 Ivy Bridge-EP 2.6GHz
(1) Kingston 32GB (4 x 8GB) ECC Unubffered DDR3 1600
(1) Samsung 840 Pro 256GB SSD (system & rendering)
(1) Western Digital Black 7200RPM 1TB HDD (storage)
This machine runs around $2900 USD and should run benchwell somewhere in the 5-minute range. Within the budget you could also use E5-2640 v2 8-cores, but they clock at 2.0 GHz, and may or may not actually perform much differently. To step up above that, you enter into the $1300+ USD range, per CPU, and end up far outside the budget. On the other hand, here's a machine built using a single i7 quad core:
- (1) Shuttle SH87R6 LGA1150 Bare-bone
(1) Intel Core i7-4790K Quad-Core 4.0GHz (4.4GHz Turbo)
(1) G.Skill Ripjaws 32GB (4 x 8GB) DDR3 1600
(1) Samsung 840 EVO 250GB SSD (system & rendering)
(1) Western Digital Black 7200RPM 1TB HDD (storage)
This machine is about $1140 USD, and should run benchwell in around 10 minutes, without overclocking, with three such machines turning in a combined benchwell capability in the 3-minute range. However, to some unknown degree, the speed advantage would be lost to network traffic and time spent merging MXIs from the different machines, along with potential networking complications that you don't have to deal with when there is just one machine involved. There is also the fact that unless you use linux, you have to purchase OS licenses for each machine.
Those are my thoughts, anyway -- in the end, I don't think it's possible to say that there is a single best solution (though I don't think AMD is a good choice, unless you are building a rack of 4-CPU servers, which can be price- and performance-competitive with 2-CPU Intel servers); it just comes down to your own preference regarding simplicity on the one hand (single multi-CPU machine), and on the other, fault tolerance and future expandability.