Mac: M3 - *Hardware accelerated RT (Part 1)

Raytracing in this instance.

1 Like

Thats why Nvidia called it RTX… they can own it.

Wanting to understand the impact BVH variations have on renders per the Metal BVH2 commit, made me research some info out there…

I’m assuming the numbers that follow the BVH acronym is dependent on the split ration for the bounding boxes, 2 keeps dividing by two, 4 by 4, 16 by 16, and nothing to do with an updated version numbering?

https://psychopath.io/post/2017_08_03_bvh4_without_simd

Downloaded the build from
Jan 20, didnt crash, bmw gpu benchmark ran at the same speed as the initial metal release. Guessing that latest fix isnt in there just yet. Kinda curious about it though :slight_smile:

idk. smdh.

5 Likes

:joy::joy::joy::joy::joy::joy::joy::joy::joy:

ikr! roflmao!

2 Likes

yes same here :frowning:

switch to shader tab and back and it looks like this.

1 Like

This is quite interesting too
https://blender.it4i.cz/research/bounding-volume-hierarchy-bvh/

Still not sure what that patch will mean for us I guess we might se in tomorrows build.

1 Like

Yup. There’s a bug report active now.
If you don’t go into solid mode in the shader workspace it’s fine, but as soon as you do it will effect all other workspaces.

You can change the timeline window to a Shader Editor with the pull down and it will be fine.

1 Like

Yeah, exciting… from what it sounds like BVH2 wasn’t even being used on ARM Macs with the new Metal builds. Tonight’s build should solve that.

2 Likes

Yes I never realised that there were even different BVH versions, I guess one learns something new everyday :wink:

I could not find however what difference to expect from BVH to BVH 2, if it is about 10% that already would be a decent improvement.

It sure will be exiting to see what else they will fix and optimise.

I noticed that 32 GB sure seems to have a bit of headroom on GPU memory while playing gourd a bit and retesting some things from the Air with 16GB.

On 16GB I could not even render the Sinosauropteryx Prima on default settings.
However on 32GB the video memory still has some room left according to iStats.
I am not sure how to see how much free space is left however or how OSX manages that.
Screenshot 2022-01-20 at 20.35.24

I found this today as I was curious about something, the M1 GPUs are Metal GPUFamily Apple 7, after reading through that I could not resist to ask Michael Jones a question.

2 Likes

Or Apple likes products owners can use without much of an issue.

Still remember my inlaws going to BestBuy for their PC issues or the set of problems my students have with their machines …

That’s the other side of the coin.

While I don’t think it will be an issue did anyone notice that once you start blender and even if you use no audio, that core audio will use like 13-14 % of cpu resources (2nd number is the cpu use in %).

_coreaudiod 373 14.5 0.2 409284640 53328 ?? Ss 3:05PM 9:31.37 /usr/sbin/coreaudiod

also interesting that it does not show in the activity monitor but only in the terminal with the command “ps aux”

Same behaviour in 3.0 or 3.1 alpha.

Isn’t 13%-14% CPU use actually not that much? If you have an M1 processor, that would account for just 1.75% of the whole CPU being used (since the M1 has 8 cores and for it to be fully used the indicator should be of around 800%). I wouldn’t know what normal behaviour is for that process though but 1.75% seems low.

Assuming the BVH fix is now included I’m testing the build from Jan 21. against the one from mid December I normally use on my base MacBook Air with 7-core GPU.

BMW27 GPU only: gets up from 2:30 to 2:47 :slightly_frowning_face:
Classroom GPU only: goes down from 6:25 to 6:03 :slightly_smiling_face:
EDIT:
Monster_under_bed GPU only: from 12:17 to 10:37 :grinning:

Definitely the classroom scene has a more complex bvh structure but why does the simple BMW scene render slower than before?

3 Likes

Yes true, let me put it that way why does it not run on the intel mac while blender is running?

ops correction, it is running but uses 1,1 % instead of 14% (and the intel I checked it on has only 8 cores too).

Which makes it 12-13 % more for what purpose.

Intel for me seems how it should be and Arm sounds like a bug.

for record keeping: the jan 21 build produces the following results overhere:
BMW27 GPU only: from 42 to 47
Junkshop GPU only: from 23 to 26
Classroom GPU only: didn’t track it before, currently: 1:39

I can’t say the m1 is fantastic for cycles rendering, it gets the job done, silently and without slowing down. However, overall viewport speed and for lookdev its working very well.

2 Likes

It sure is disappointing if that is with the Bvh2 patch or perhaps not…

I guess I can confirm the results.

The BMW scene is 8 sec slower for me.

Monster under the bed sees a huge improvement 47 sec faster.

Classroom 20 sec faster.

Mr Elephant 8-9 sec faster

Sinosauropteryx Prima coming soon

So far it is looking like SSS gets a nice boost and more complex scenes.

1 Like

I see no change in the render time for one of my test scenes. Maybe the patch didn’t make it in time or its just not making a difference.