Hey everyone,

All of the issues I am now having with OS X and Houdini has pretty much forced a move to build a PC. To begin, I will be specifically interested in doing large dynamics simulations (pyro, flip, grain, bullet, rbd, cloth)

I've read everything I could find on the forums in terms of Houdini specific setups, and I have a few questions, if you have the time.

1st off, i'm willing to spend around $3-5k ~ with lower being more preferable, but the wiggle room to make “worthwhile” parts choices if necessary.

2nd, i'd like (if possible), to build something that is a bit “future-proof” with the ability to upgrade down the road. Especially, if there are cost reduction choices that I could make right now, that I could then upgrade later down the road.

Just to get it out of the way, i'm thinking for the GPU: EVGA GeForce GTX Titan X (down the road I could get a second for SLI)
other options are 2 x GTX 980's or 1 x GTX Titan Z

The CPU is where i'm having the most questions. Here are my thoughts:
i7 6700K Skylake or i7 5960X or (here is where i'm not sure about)… Should I go for an LGA2011-v3 MOBO DUAL Socket, then get 2 x Xeon e5 (lower end), possibly ES 2630 2.4 GHz 8 core and 20 MB Cache. Another option is to get a higher level Xeon e5 for similar price of the 2 x cheaper e5 processors. Then only install it in one socket of a dual socket, (i'll have to make sure I find a MOBO that will not be affected by this). And then, upgrade with a 2nd CPU later on, when I have some more money.

This also opens up the question (more specific to Houdini) and it's performance with Hyper Threading, dual processors, clock speed, cores, and all that.
So does Houdini benefit more from a CPU with higher Ghz and less cores, or lower Ghz and more cores? In particular, related to the basic concept of the CPU's that i'm debating up above (or if there are better choices, please point me to those, thanks!)

Overall my thinking is this, to satisfy the “longevity and upgrade” requirement that i've self imposed, I feel like if I went the dual Xeon route, it will also give me the ability to upgrade MUCH more RAM over time. For the moment I could go 64 GB or similar, then later up to 128+ also, if I went with the “cheaper” dual xeon's, i could upgrade to 2 better ones when prices drop.

On the other side, if I went with something like the new i7 6700K Skylake, it let's you upgrade to max of 64 gb (so that's better than the previous 32 GB limit). But then at that point, it won't really be “upgradeable” anymore, because I could just max that out right now. The only real upgradeability to the system would then be the CPU if they come out with a higher core Skylake later on, that still fits that same MOBO. However, there are no details on that yet.

Well that's about it for now. I appreciate your help. Cheers everyone.

Multi-threading is very well used in Mantra, quite a few nodes have good threading, and I would think H15 expands upon what is threaded.

SLi is not useful, but you can run one card for the display and another for openCL processing. Titan X would be the go for it's ram size and cores. You mentioned large sims though, so you'll most likely always run out of GPGPU ram regardless and have to process on the CPUs.

Memory efficiency keeps getting better but then people always just add more detail, so more ram is always good!

Memory capacity is probably the single most important thing for a Houdini build. Large sims and renders consume a lot of RAM and if you run out, things slow down real fast. Memory speed doesn't matter quite as much (ie, don't get the expensive DDR3 2400 packs). Go for 64GB if you can.

In terms of clockspeed vs. cores, that's a tougher question. For simulations, cores is very important, but beyond 8 threads you start seeing diminishing returns. Clockspeed helps everywhere, especially those parts that aren't threaded (and when you run into a single threaded operation, you really want it to get done fast). Luckily modern processors have good boost clocks, so even if you're running a 8-core CPU, a single CPU can still boost up to near-4GHz levels. The 5960X is a good balance for a workstation, IMO - single threaded performance, multi-threaded performance, and price.

The Titan-X is a good choice if you're considering doing OpenCL-accelerated sims, with its 12GB of VRAM. If your not, consider the 980Ti and save $350 (it's a cut down TitanX w/6GB and 90% of the shaders).

Thanks guys,

I'll definitely keep this all in mind. Yeah, I definitely felt the Titan X will be the best route in terms of my GPU.

As for RAM, i'm definitely aiming for larger amounts, which is why I was leaning toward the LGA2011-v3 motherboard, dual socket, with xeons, so I can utilize more RAM and have some room for upgrading.

Whereas, if I go with the 5960X, this will be a single CPU machine, and the ram utilization would be less.

Thanks for the thoughts! I will definitely keep this all in mind, while i'm doing my search. Cheers

Okay so here are a few options i've put together with pcpartpicker:

Dual Intel Xeon E5-2630V3 2.4GHz 8-core:
http://pcpartpicker.com/user/Epheks/saved/fKwv6h [pcpartpicker.com]

Single Intel Xeon E5-1660V3 3.0 GHz 8-core:
http://pcpartpicker.com/user/Epheks/saved/hHCWGX [pcpartpicker.com]

Intel i7 6700K 4.0GHz Quad-Core Skylake:
http://pcpartpicker.com/user/Epheks/saved/Dsd2FT [pcpartpicker.com]

Each system has 64GB of RAM. The main difference is upgradability. In the Dual and Single Xeon setups, there is more room to upgrade in the future, whereas the i7 is basically maxed out.

Any thoughts are welcome! Also, if you notice anything in the parts list that might bottleneck the system, please point that out.

Cheers

do you work for you own or do you work in a team?

i would prefer the cheap solution because you always can ad another system as a batch system.

for the dual socket system you are coming to a price range where you can use a barebone that safes you time and it's sometimes cheaper:
http://www.newegg.com/Product/Product.aspx?Item=N82E16816139087 [newegg.com]
it's a 900watt PSU but it should be enough for your setup.

This would be for my solo work.

The cheaper option with the ability to get a second system if needed is appealing and honestly, if I need to upgrade, then I can just build a new “cheap” system that is more up with the times, and turn the old one into a secondary system.

I just kept thinking the dual xeon, would be a better all-in-one, longer lasting method, but yeah, a more powerful single cpu, maxed out, for less money, is probably the way to go!

Thanks for your thoughts.

I'm closing in on my final build. After researching numerous benchmark comparisons, devouring forums and having discussions about possible builds, i've been steered towards a single high clock, high core cpu instead of a low clock, high core dual xeon.

I would like to know if anyone has used or is familiar with the following cpu's and which one would give me the best performance in Houdini.

Intel i7 6700k (Skylake) - Quad Core, 4.0GHz, 8M Cache ($339.00)
Intel i7 5930k - 6 Core, 3.5Ghz, 15M Cache ($594.00)
Intel i7 5960X - 8 Core, 3.0 Ghz, 20M Cache ($1059.00)
Intel Xeon E5-1650v3 - 6 Core, 3.5GHz, 15 M Cache ($586.00)
Intel Xeon E5-1660v3 - 8 Core, 3.0GHz, 20M Cache ($1080.00)

The 6700k, 5930k, 5960x all support 64GB of RAM –> Which I will be getting. The E5-1650v3 and the E5-1660v3 supports 768GB of RAM.

Many people have already said that the i7 6700k can easily be OC'd to a comfortable 4.5 or 4.6GHz. So i'm curious if having such a high clock speed would outweigh it's 4 cores and 8M Cache.

As previously mentioned, this system is going to be an All in One, for doing big sims and rendering (both which are multi-threaded). Is the lower clock, higher core cpus going to actually perform better than the insane 6700k clock with lower core? Also, how much does the L3 cache size help? Is there going to be a huge benefit having say 20M cache vs the 8M Cache?

Thanks everyone! I can't wait to get this thing purchased and built up. I really appreciate any final thoughts before I get this thing going.

just for a note the 20mb has to feed 8 cores the 8mb has to feed 4cores so you will get a very small benefit.

8MB Cache L3 / 4 Cores = 2 MB per Core
20MB Cache L3 / 8 Cores = 2.5 MB per Core

just for fun made a hole system of the price difference from these two cpu's (render node):
http://pcpartpicker.com/p/Br7dhM [pcpartpicker.com]

mandrake0
just for a note the 20mb has to feed 8 cores the 8mb has to feed 4cores so you will get a very small benefit.

This make no sense. L3 is shared.

Can you share more of your interpretation of CPU cache uses as L3 feeds L2, then L1 then registers.

5960x or 6700k?

6700k 4.0GHz base, 4-core, 8M cache ($360)
5960x 3GHz, 8-core, 20M cache ($1060)

Can anyone say with confidence one of these will perform better than the other? I'm primarily going to be doing large sims. I know Twod mentioned diminishing returns above 8 threads.

One of these bad boys is going to be paired with 64gb ram, gtx titan x and i'm going to call it a day! So what's the final vote everyone? Thanks for all your time.

It's almost impossible to say with a direct test as there a lots of moving parts in this equation.

Using OsX's ‘powermetrics’ command, I'm not seeing Turbo boost at all in Houdini, based on this, the base frequency is the one I would calculate total CPU power.

An unknown is that H15 is released next month and we don't know what has been accelerated.

Obviously single threaded parts will be faster on the 4GHz by 25%, and Mantra should be 1/3 slower on the 4GHz

The last part is you as the operator. Your skill can have an enormous part of the equation. Are you running tests currently?

I just finished my build I sold my Mac Pro and built an insane pc

5960x overclocked to 4.5ghz (stable) I had it running 4.8ghz but it heated up too quick.

Dual Titan x .. Don't run in sli for Houdini or rendering engines

Proper liquid cooling.


http://markpamatat.com/Build/Alpha_Centuri_Power.jpg [markpamatat.com]

http://markpamatat.com/Build/Alpha_Centuri_No_Power.jpg [markpamatat.com]

Nice! Can you please run some benchmarking software

Passmark - demo version for 30 days
http://www.passmark.com/products/pt.htm [passmark.com]

Cinebench
http://www.maxon.net/products/cinebench/overview.html [maxon.net]

I didn't run those but I did run the 3dmark tests and was put on the hall of fame for #19 fastest system ever tested with Dual GPU cards.

http://www.3dmark.com/fs/6127366 [3dmark.com]

http://www.3dmark.com/hall-of-fame-2/fire+strike+3dmark+score+extreme+preset/version+1.1/2+gpu [3dmark.com]


Mark

not a fan of the Passmark test.. just got ~7000 … But it doesn't even read the right amount of ram and stuff and I don't know just not feeling that test.. ill run a cinebench.

The passmark I did on my 24-7 overclock setup. so it was set much more conservative then my testing with 3dmark. 3dmark I was at 4.8ghz, on this test I was running 4.4 with non overclocked ram.

Cinebench OpenGL test I got 208.48 Frames Per Second

Cinebench CPU I got a 1750 on my 4.4ghz setting