Page 1 of 1

MacMini: Losing AVX2, Intel and RAM

PostPosted: Tue Nov 16, 2021 2:50 am
by solotibia
Hi,

As the MacMini line loses its Intel models and moves to become a pure Apple silicon (M series) range, what elements of Hauptwerk audio will be negatively impacted through the loss of AVX2?

As the current M1 processor RAM in a MacMini maxes out at just 16GB, it seems that the remaining M1 and i5 MacMini's are now of little use in the Hauptwerk world. Right now the only replacement for a 32GB, 6 core i7 MacMIni might be limited to a MacPro, MacBookPro laptop. Or will I need to move to a W11 machine with its many unknowns?

Does anyone know if the M1 Pro processor (used in the latest MacBookPro) is about to become part of the MacMini range roadmap? I see snippets of news about about a MacMini Pro being released before Christmas. I have sent Santa a letter.

Cheers,

Ian

Re: MacMini: Losing AVX2, Intel and RAM

PostPosted: Tue Nov 16, 2021 5:46 am
by mdyde
Hello Ian,

[Topic moved here.]

solotibia wrote:As the MacMini line loses its Intel models and moves to become a pure Apple silicon (M series) range, what elements of Hauptwerk audio will be negatively impacted through the loss of AVX2?


- The main part of Hauptwerk that specifically takes full advantage of AVX's ability to process 8 floating point numbers simultaneously per instruction is the audio engine, but only (for any given pipe) if you have either of the 'Perspective mix: source persp to output persp 3/4' adjustments turned up above zero on for the pipe on the voicing screen. On a CPU without AVX the audio engine would require twice as many CPU instructions to process the pipe's audio stream in that case, which may negatively affect polyphony. (If you don't have those two voicing adjustments turned up, the audio engine only needs to process 4 floating point numbers simultaneously, so that won't apply.)

- The convolution reverb engine will also take advantage of AVX/AVX2/AAVX-512 for processing 8 or 16 floating point numbers simultaneously (for DFFT/IDFFT operations) if available. On Apple's ARM-based Macs Rosetta probably won't benefit from that, so performance of convolution reverb might currently be less good on Apple's ARM-based Macs than Intel ones. (However, in the longer term when we hope to make a native port to Apple's ARM-based CPUs, which would be likely to regain most or all of that performance benefit.) Of course, that's only relevant if you actually use convolution reverb in Hauptwerk.

- AVX2's fused multiply-accumulate instruction will be used by signal processing in the audio engine and elsewhere for applying filters, which gives a small but worthwhile benefit. Currently, Apple's Rosetta on ARM-based Macs will use the SSE3-optimised Hauptwerk executable (not the AVX2 one), which won't benefit from that, although I understand from M1-Mac owners that performance is still very good regardless: viewtopic.php?f=16&t=19285 . (As above, in the longer term a native port to Apple's ARM-based CPUs should be able to regain that small performance benefit.)

- Performance gains from other general compiler optimisations in the AVX2-optimised Hauptwerk executable above the SSE3-optimised (e.g. to organ loading speeds) would potentially be lost currently (but might be regained in a longer-term native ARM port). However, as above, reported performance under Rosetta is still currently very good.

Hence, in summary, the main performance considerations currently are likely to apply if you either use output perspectives 3 or 4 on the voicing screen, or if you use convolution reverb in Hauptwerk. (There should be no direct effect on audio quality in any case -- only potentially indirect, e.g. if you had to disable to some audio engine quality features in Hauptwerk to get enough performance, but that probably wouldn't apply).

solotibia wrote:As the current M1 processor RAM in a MacMini maxes out at just 16GB, it seems that the remaining M1 and i5 MacMini's are now of little use in the Hauptwerk world. Right now the only replacement for a 32GB, 6 core i7 MacMIni might be limited to a MacPro, MacBookPro laptop. Or will I need to move to a W11 machine with its many unknowns?


Aside from Apple's recently-released 'M1 Pro' MacBook Pro (which supports up to 64 GB of RAM), I believe it's expected that Apple will release comparable Mac Mini and iMac models next year.

Re: MacMini: Losing AVX2, Intel and RAM

PostPosted: Tue Nov 16, 2021 9:35 am
by larason2
The mac mini range will eventually move to M1 Pro/Max or M2 processors, but it is unlikely they will before the end of the year. Most predictions I’ve seen say sometime early 2022. Apple has said it will move the whole mac product line to custom processors by the end of next year, so I would wait if possible.

Re: MacMini: Losing AVX2, Intel and RAM

PostPosted: Wed Nov 17, 2021 5:04 am
by solotibia
Hi Martin,

Many thanks for your valuable insights into Hauptwerk's use of AVX2. I think that I can breathe a sigh of relief.

I don't use output perspectives 3 or 4 on the voicing screen, or Hauptwerk convolution reverb. So, should I need to replace my current Hauptwerk computer, with one of the mooted M1PRO MacMinis there will not be a loss of the fabulous audio quality that I currently enjoy.

I'm very happy with my three year old 6 core i7/32GB/1TB MacMini which replaced two earlier MacPros, and I can now see that there could be a replacement pathway with another MacMini that uses the M series Apple chipsets. Even more so when a native M series version of Hauptwerk is released.

Cheers,

Ian

Re: MacMini: Losing AVX2, Intel and RAM

PostPosted: Wed Nov 17, 2021 5:09 am
by mdyde
Thanks, Ian. Excellent. You're very welcome.

Re: MacMini: Losing AVX2, Intel and RAM

PostPosted: Mon Nov 22, 2021 3:57 pm
by Romanos
it seems that the remaining M1 and i5 MacMini's are now of little use in the Hauptwerk world.

Not so. I have an m1 mini with 16gb ram, and I run HWVI via rosetta just fine. True, the ram limits you to smaller instruments, but I'm working on a continuo project, and for something like that, it's perfect. Rock-solid stability, and plenty or ram for smaller instruments or even medium instruments with single perspectives loaded. It's really the duplication of multiple perspectives that gobbles up ram the worst. I don't like rear samples (too wet/muddy for me) so I almost never load them, so that helps tremendously.

I suspect the moment we get the second version of an m1 mini with higher ram thresholds the minis will be an appealing option for HW users with more modest requirements.