Ability to use more than 64 threads

[CAMERON_SENSE]

Sure, you can`t get exact results. But for machines that are to be used for rendering (or at least that is the most demanding task), I would still use render benchmarks rather than PassMark. You can see how close you get to theoretical performance across a few different renderers and then average the results or something. PassMark on the other hand, will be wrong from the beginning as it just gives an idea for general performance, not the specific tasks that you plan to use the computers for.

I tested something like twelve different benchmarks which I did on machines with: 2 sockets x 4 cores, 2x8, 2x12 and 4x8. As far as I remember both Cinebench, POVRay and Vray had pretty consistent results with very close to theoretical performance. Mental Ray couldn`t handle the four socket machine at all, but scaled very well otherwise. Vray is the render I use for work, and a few scenes have later proved to scale poorly on four sockets. Those are exceptions though.
The rest of the benchmarks (including PassMark) favoured less cores and higher frequency to a MUCH larger extent than any of the renderers. Since fewer cores often means higher frequency, this would give something like an inverse result of what is correct :P.

On the other hand, if you have a range of different software apart from rendering, it would of course be wrong to base the results on rendering benchmarks alone ...

[ramy02]

depends on how you use your farm. If its render jobs through bb or other render manager, your best bet to run multiple instances of slave on a single node. Single instance won't scale well, you are most likely run into problems. Same goes for dr for that matter.

Really we would like to use the dual 18 core cpus for workstations as well. So simply just rendering through 3dsmax UI is what we're after, and what I'm curious about. Just don't know if VFB will support 72 buckets from just one machine.

[peterguthrie]

has anyone tested turning off hyperthreading on a new monster machine..? I would, but I'm in the middle of a render

[vlado]

Really we would like to use the dual 18 core cpus for workstations as well. So simply just rendering through 3dsmax UI is what we're after, and what I'm curious about. Just don't know if VFB will support 72 buckets from just one machine.

Heh, the VFB is the least problem; after all it does show hundreds of buckets for DR already

I will try to post a build with processor groups support sometime next week.

Best regards,
Vlado

[ramy02]

I would be very interested to know how vray handles 2 X E5-2699 without DR.

[CAMERON_SENSE]

*Double post

[CAMERON_SENSE]

has anyone tested turning off hyperthreading on a new monster machine..? I would, but I'm in the middle of a render

TUrning off HT would typically reduce performance by 20-25%, so thats really not an option ...

[ramy02]

Does anyone know how v-ray handles 2x 12 cores (48 threads) or 2x 16 cores (64 threads) on a single machine? We may go that route.

[CAMERON_SENSE]

Does anyone know how v-ray handles 2x 12 cores (48 threads) or 2x 16 cores (64 threads) on a single machine? We may go that route.

2x12 cores is not a problem at all. It scales very close to 100%. Just make sure all the memory settings are optimal and reduce the bucket size. I found that a bucket size of around 10 (+\-) is optimal for 48 threads.

[im.thatoneguy]

What is the rhyme/reason behind picking a specific bucket size or is it purely experimental on what works best per scene? Are small buckets to keep the L1-4 fetches to RAM consolidated so that all cores are working on similar data and each core's cache isn't requesting totally different data and bottlenecking the RAM pipe?

[vlado]

I don't think it has anything to do with RAM caches, but more with CPU load balancing.

Best regards,
Vlado

[FSGFX]

reduced the bucket size from 64 to 12 on my 48 thread machine.... adds ~50% rendertime... ?

[CAMERON_SENSE]

What is the rhyme/reason behind picking a specific bucket size or is it purely experimental on what works best per scene? Are small buckets to keep the L1-4 fetches to RAM consolidated so that all cores are working on similar data and each core's cache isn't requesting totally different data and bottlenecking the RAM pipe?

Its just experimenting for the specific scene, but for 48 threads its generally around size 10. Vlado kan probably elaborate on why

[CAMERON_SENSE]

reduced the bucket size from 64 to 12 on my 48 thread machine.... adds ~50% rendertime... ?

All I can say is that by reducing the bucket size, things generally renders faster. If I use a low core \ high frequency machine of the same CPU-architecture as a bar for 100% theoretical performance, I get very close to optimal performance on a 48 thread machine

[vlado]

Anyone that wants to test a build with support for processor groups, please email me to vlado@chaosgroup.com with your 3ds Max version number so that I can get you download links.

Best regards,
Vlado