Cycles rendering twice as fast in this benchmark comparing M5 to M4. That’s really promising. I think M5 line has to become quite popular with 3D and AI creators. Performance is catching NVIDIA fast, but unified memory vastly expands possibilities compared to traditional GPUs.
Apple’s M-series will eventually hit the power/heat wall as well.
At this point, architecture and new function types is at best serving to buy a bit of time until we officially hit the barrier on what can be done on silicon-based chips (as advances in the fab process slows down and physics begin to get in the way of further shrinking).
Since chips based on graphene and optics are not even close to production-ready at the moment, it will soon get to where the primary means of increased computing power falls onto the software developers and their will to face all of the legacy cruft and bloat that causes a lot of cycles to be wasted on overhead.
Current speedups are due to adding efficiency systems like multiple BVH build methodes depending on scene type, I don’t think we’re in territory of “make them run faster and hotter” yet.
This is why Apple continues to work on Blender: discover how you can speed up Raytracing and which hardware components you need for that. They can see each part of the engine, which they can’t for game engines. But those reap the benefits too.
In the end this helps Apple for making a case for using Apple hardware.
Seems like with the M5 we finally get into the GPU level we creatives also need + more VRAM than NVIDIA cards.
On Apples website they even feature DrawThings which is an AI render engine that downloads AI models. It seems that the M5 is even more qualified for doing AI calculations.
I’ve been playing around with DrawThings on the Mac, but couldn’t get images out of it that were on par with the online platforms. Speed was good though: around 9sec for 512x512 non-final quality if I remember correctly. Running LLMs like Qwen locally was as fast as web platforms.
Interesting point: another app I have for AI RAW denoising (PureRaw) lets you swap GPU vs NPU processing. On the M2 Ultra, they’re equally as fast. Maybe the added neural cores in the M5 give the advantage to NPU processing. Could be a lack of optimization too.
I have a Question. Occasionally I need to get into Blender’s inner Sanctum (to find preferences or startup file etc.) and the advice I Get is this file path: /Users/USERNAME /Library/Application Support.
But I NEVER see “username/library” portion. Do I need to use a key combination as well (it isn’t indicated)?
This seems to be MacOs and Blender version independent.