renderfarmer, awesome setup! very well executed and thought out. I was honestly expecting an AMD setup, not quite a 2P E5 setup however it made me smile!
Thanks! I actually tested AMD chips when I was planning this farm (Early 2012) and I found them to be a worse value proposition for rendering than Intel's offerings.
My old farm was two dual Nehalem X5650 machines. For the new farm I had a choice of either E5 (Romley) or Magny Cours. I picked up a second hand Magny Cours 6128 off of ebay and an Open box UP board from newegg for testing. I then put together a Romley system (had to buy new as they had just shipped after months of delays). I took three different rendering scenarios (interior, exterior, studio) and ran them all at 4 different outputs 1k-6k on all three machines. On a per GHz basis the E5 came out to 1.09x as fast as the Nehalem while Magny Cours came in at 0.77x. I then spec'd out systems with both chips and I had to go with a quad socket Magny Cours system to be competitive with a similarly priced E5. But there was another big factor in my value decision: resellabilty.
The last straw for AMD was when I tried selling the Magny Cours back on eBay... It was up there for nearly a month before someone bought it from me and I was asking for a very low price. I could flip any intel chip in 15min flat at that kind of a discounted price...
For the past year the farm has been filled with E5-2620 chips. I only recently swapped them out for 8 cores. I'll be putting up the E5-2620s up on ebay tomorrow.
My only question/critique in regards to the nodes and host, is why are you not running these chips in quad channel? Cost or oversight? You could potentially shave quite a bit of time allowing the machine full memory bandwidth.
Great catch. I have the dimms arranged in that way because that's how the mobo manual says to arrange them. I tried doing one stick per bank but it still registers as Dual Channel.
The bright side is that I studied this carefully and my findings were that there is no impact on rendering speed whatsoever irrespective of the RAM configuration.
I ran a rendering test using four RAM configs on the same system.
- 4x8GB ECC RDIMM @ 1600 (dual channel)
- 4x8GB UDIMM @ 1600 (dual channel)
- 4x8GB UDIMM @ 1333 (dual channel)
- 4x8GB UDIMM @ 1600 (quad channel)
- 8x8GB UDIMM @ 1333 (quad channel)
All rendering times were within 0.2% of each other and not necessarily favoring the 1600 or quad channel; statsictially insignificant.