SEARCH
diglloydTools™
Reader Comments on Apple Mac Studio M1 Ultra
Related: 2019 Mac Pro, 4K and 5K display, Apple, Apple Mac Studio, Apple Silicon, bandwidth, computer display, Mac Pro, memory, memory bandwidth, Other World Computing, reader comment
Consult with Lloyd on your computer or photographic purchase, backup strategy, archival and storage strategy, etc.
See the in-depth review of the Apple Mac Studio M1 Ultra.
Gene F writes:
The more test reviews I see the more I think software hasn’t caught up to the M1 Ultra yet. For instance, almost no existing routines exploit the difference between the 48 and 64 cores. And there’s already hints of an M2 chip.
MPG: correct that some “catch up” time is needed. We can expect future improvements, both on CPU and GPU support. Just 6 months ago Adobe stated this about Photoshop:
Everything written in OpenCL is being rewritten for Metal on macOS this year. OpenCL and OpenGL are dead. I believe SmartSharpen is at the top of the list to move to Metal off of OpenCL.
However, it is also true that most software has not caught up with the 2019 Mac Pro, or the iMac Pro (now discontinued). Because it’s hard to do well, and that means time which means money.
And it is also true that many computing tasks are constrained to serial (in-order) execution in full (eg hashing) or at one or more choke points in the process (eg focus stacking). Non-parallizable tasks inherently limits the number of CPU cores that can be used productively ("scalability").
Which leaves us with tasks that can benefit from multiple CPU or GPU cores. But it’s a lot more complicated to program software for that than to just run it as one “thread”.
And so the M1 Max and M1 Ultra will take years to catch up, the same reasons that most of the 28 CPU cores on my 2019 Mac Pro so often go unused.
Still, unprecedented memory bandwidth of the M1 Ultra chip should reset expectations for the scalability of certain tasks. With 2.6X the memory bandwidth of my 28-core 2019 Mac Pro, the 20-core M1 Ultra has 3.7X the memory bandwidth per active core (4.5X faster per performance core). Developers that saw diminishing returns because of memory bandwidth need to rethink their assumptions.
