Hi Guys,

I'm needing a new machine for 3D (Houdini (compatible with the upcoming release 14), Maya) and 2D (Nuke) work.
I'm wondering if you have any suggestions for specs and graphics. Budget is about 2500$.

I'd like to get something that has a good compromise between price and power (especially for dynamics, fluid sim, particles..ect)

The systems I use are Windows and Linux.

Advice is appreciated.
Cheers

My recommendations:

Intel i7 5930K or 5920K (both 6-core, $640 or $420)
X99 motherboard with 8 RAM slots ($250-350)
Nvidia GEForce 980 or 970 ($650, $440)
32GB DDR4 RAM (~$550) or 64GB (~$1100)
460GB SSD (~$300)
Cooler Master Hyper 212 Plus CPU heatsink ($25), or CORSAIR Hydro Series H100i for a watercooler ($110)
ATX case with at least 3 120mm fans ($60-120)
2-3GB 7200rpm HD ($120-200)

There's a bit of play there to go for more CPU, more GPU or more RAM. A Xeon is a bit too pricey to recommend for a $2500 system, as is an 8-core i7 5960X ($1200).

Thank you so much for such detailed info!
I was thinking about a Xeon, but you're right, it's way too pricey

I was wondering about the graphic card. Do you think it would higly increase the performances to choose the Nvidia Quadro kxxxx? (one that is not so far from the price of the GeForce that you suggested).

Thx!

For H14, no, Quadro vs. GEForce of the same GPU type makes no difference in terms of performance. Quadro tends to have much more VRAM though, which can be useful for very large scenes or driving multiple high-res monitors.

Hello, could you please share any more info about usefulness of AMD Fire Pro Wx100 row (W7100, W9100)?

The FirePro W7100 has 8GB ram and OpenCL 2.0 support and should support more then 4GB RAM for gpu computation (smoke sim) for much better price then Nvidia. (W9100 has 16GB)

Can Houdini (14) utilize that on those GPUs?
From my old tests NVidia GPU are driver crippled to the 4GB limit for OpenCL :roll:

Can Houdini (14) utilize that on those GPUs?
From my old tests NVidia GPU are driver crippled to the 4GB limit for OpenCL

Yes, Houdini 14 works with both AMD and Nvidia's pro cards, and the higher-end consumer cards as well.

From what I've heard, Nvidia is working on removing that 4G limitation, but I have no idea when a 64b CL driver might appear.

Does Houdini definitely allow full use of 64-bit memory on AMD cards with OpenCL 2.0 supporting drivers then?

Has anyone tried out the OpenCL settings thread example scene to see what resolution they can top out at with 6/8/12 GB vram?
https://www.sidefx.com/index.php?option=com_forum&Itemid=172&page=viewtopic&t=25234 [sidefx.com]

(…and being a bit of an nVidia fanboy myself, any word yet on when they're going to get their asses into gear and support 2.0?)

twod
For H14, no, Quadro vs. GEForce of the same GPU type makes no difference in terms of performance. Quadro tends to have much more VRAM though, which can be useful for very large scenes or driving multiple high-res monitors.

Hi twod. So just to be clear: you mean that my Quadro K2200 4GB has the exact same performance as a GTX750ti 4GB (both Maxwell GM107)?

And do you maybe have any info about display quality? I mean: allegedly a viewport powered by a Quadro should have none or at least fewer display artifacts/glitches compared to a Geforce. You have any experience with this? Ever noticed big differences between Quadro and Geforce?

cheers

sn1jder
Hi twod. So just to be clear: you mean that my Quadro K2200 4GB has the exact same performance as a GTX750ti 4GB (both Maxwell GM107)?

And do you maybe have any info about display quality? I mean: allegedly a viewport powered by a Quadro should have none or at least fewer display artifacts/glitches compared to a Geforce. You have any experience with this? Ever noticed big differences between Quadro and Geforce?

I'd be surprised if there were viewport artefacts outside of the occasional driver bug or compatibility issue. GeForce cards are consumer cards, not defective ones :-)
As I've understood it, buying a Quadro has always come down to the need for a professional hardware support contract from nVidia (along with conservative, solidly tested driver updates), and/or improved double-precision compute capabilities compared to consumer cards which are artifically limited in that area (except for the Titans). As far as I know, double-precision computing is usually limited to scientific applications, and Houdini doesn't actually make use of that, so for all practical vfx purposes, a GeForce is usually directly equivalent to its (5x more expensive) Quadro counterpart.

The top-end Quadro does have 12GB of RAM, which no consumer card has so far, but given nVidia's reluctance to update their OpenCL support, they all might as well be 3GB cards anyway.

VortexVFX
Does Houdini definitely allow full use of 64-bit memory on AMD cards with OpenCL 2.0 supporting drivers then?

Has anyone tried out the OpenCL settings thread example scene to see what resolution they can top out at with 6/8/12 GB vram?
https://www.sidefx.com/index.php?option=com_forum&Itemid=172&page=viewtopic&t=25234 [sidefx.com]

(…and being a bit of an nVidia fanboy myself, any word yet on when they're going to get their asses into gear and support 2.0?)

I believe you have to set the environment variable GPU_FORCE_64BIT_PTR to 1 in order to get 64b OpenCL addressing on AMD cards. It defaults to 32b addressing.

As for Nvidia, I can't comment on their unannounced products or software. Sorry

sn1jder

twod
For H14, no, Quadro vs. GEForce of the same GPU type makes no difference in terms of performance. Quadro tends to have much more VRAM though, which can be useful for very large scenes or driving multiple high-res monitors.

Hi twod. So just to be clear: you mean that my Quadro K2200 4GB has the exact same performance as a GTX750ti 4GB (both Maxwell GM107)?

And do you maybe have any info about display quality? I mean: allegedly a viewport powered by a Quadro should have none or at least fewer display artifacts/glitches compared to a Geforce. You have any experience with this? Ever noticed big differences between Quadro and Geforce?

cheers

The Quadro drivers tend to have much fewer regressions (bugs) than the GEForce counterparts, since they aren't being continually re-optimized for new triple-A game titles. That's probably their biggest strength. You also get unlocked geometry processing, 10b monitor support, quad-buffer stereo support (for 3D monitors), a lot more VRAM, and unlocked 64b FP support as VortexVFX mentioned. Houdini supports Quad buffer stereo, so if that's a requirement for your work, you'll need a Quadro.

Hardware-wise, Quadros have more memory but are usually clocked a bit lower than GEForces to fit within certain thermal and power tiers. Extra connectors for the pro market are often available as well.

Finally, your customer support with a Quadro should be a lot better, from what I've heard. This is often an unmentioned but very important feature for small shops and studios alike.

twod
I believe you have to set the environment variable GPU_FORCE_64BIT_PTR to 1 in order to get 64b OpenCL addressing on AMD cards. It defaults to 32b addressing.

Cheers, yeah, I saw that mentioned a while back - have you confirmed that it definitely works with Houdini then? It sounds great in theory, but I've yet to see anyone post saying they've attempted/succeeded in actually utilising above 4GB vram on an AMD card.

I'm tempted to pick up an 8GB Radeon to try this out if it definitely works, but it's a big chunk of change to put down without knowing :-)

VortexVFX

twod
I believe you have to set the environment variable GPU_FORCE_64BIT_PTR to 1 in order to get 64b OpenCL addressing on AMD cards. It defaults to 32b addressing.

Cheers, yeah, I saw that mentioned a while back - have you confirmed that it definitely works with Houdini then? It sounds great in theory, but I've yet to see anyone post saying they've attempted/succeeded in actually utilising above 4GB vram on an AMD card.

I'm tempted to pick up an 8GB Radeon to try this out if it definitely works, but it's a big chunk of change to put down without knowing :-)

I haven't confirmed it, no. I have an older FirePro W8000 with “only” 4GB VRAM, so hard to tell if it's working or not.

Well, I'm very glad to announce that it not only works, but appears to work without even needing to faff about with environment variables.

I bit the bullet and ordered in a Radeon R9 290X with 8GB vram so I could test this out. With the latest AMD Omega driver set installed, Houdini instantly recognises it at an OpenCL 2.0 device, with 64-bit addressing, 8192 memory, and a 4048MB maximum allocation (Not sure what that affects, but GeForce and Intel drivers report it as 25% of the total memory pool, whereas it appears to be near 50% here)

I managed to hit 150 megavoxels (532x532x532 res) before it ran out of memory. So, it does indeed seem that AMD is the way to go if you want heavy OpenCL sims!

Nice! Post speed results too

Try mucking around with this hconfig -H setting

HOUDINI_OCL_MEMORY_POOL_SIZE
Houdini uses a memory pool for allocating device memory through OpenCL for
improved performance and to decrease fragmentation. This variable is the
percentage of total device memory to allocate for the pool. For example, the
default setting of 0.125 will allocate 1/8 of device memory for the memory
pool. Note that if the specified OpenCL device is 32-bit, the total device
memory is clamped to 4GB before being multiplied by this percentage.

Thanks, I'll do some speed tests soon and add them to that thread with the test scene for comparison. It's plenty fast anyway :-)

Thanks for the suggestion - I tried out the HOUDINI_OCL_MEMORY_POOL_SIZE variable, but it appears it's hard-coded to max out at 0.5 (presumably by the GPU). If I set 0.45, hconfig will report it as 0.45… if I set it as anything higher than 0.5, it'll report it as 0.5. For whatever reason, the allocation limits seem to unnegotiable. I think it's something to do with how much can be allocated in one go, so it doesn't necessarily limit the sim to half the vram, and even a basic sim has to fit two scalar fields and a vector field into memory, so I'd hazard a guess that at most it'll need to allocate about 20% in one chunk. Absolutely no idea if I'm even barking up the right tree with that assumption though.

At the end of the day, I'm much relieved to be able to fully able to max out my 8GB with a typical sim :-)

This is the tree you want to bark at.

https://www.sidefx.com/index.php?option=com_forum&Itemid=172&page=viewtopic&p=139633#139633 [sidefx.com]

Interesting, so it sound as though setting it lower/zero might actually slightly increase available ram, at the expense of speed. Might have to test that at some point.