programmer

Bill’s work was tremendously inspiring to me back in the 80s.  One of the greats of the computer industry, gone too soon.  Very sad.  😢

danvm said:

And I have work with some of those workstations, and the performance is not as pathetic as you mention.  Some of them have advantages over Apple Silicon, specially when comparing the GPU. 

You say "workstations" whereas what you quoted said "laptops".  Big difference.  Apple Silicon's performance really shines in power-constrained situations.  If you can run plugged into the wall and with big heat sinks, then the Intel/AMD CPUs and the nVidia/AMD GPUs can burn copious power to deliver serious performance.  They're still inefficient in terms of compute per watt, but you don't notice as watts are cheap in such a situation.

It would be interesting to see a large scale data centre built from ARM-based machines and compared to ones build from Intel/AMD-based machines, and compare the operating costs.  Some of the big cloud vendors offer lower cost ARM-based hosts just for this reason -- they greatly reduce energy and cooling costs in the data centre.  Not Apple's focus though, so we aren't likely to see Apple Silicon based data centres (except perhaps for Apple's own, but they are typically very secretive about that).

sflocal said:

Intel deservedly deserved the boot.  I won't miss Intel on Macs.   What I will do is continue to keep my 2018 Mac Mini for the long, foreseeable future for those rare times I need to run x86 Windows natively.  That's it.  Other than that, it's Apple Silicon to the rescue!

Same here.  My M4Pro mini absolutely wipes the floor with the old Intel one.  The latter is now relegated to a Windows game machine, and that’s it.  Even Mac x86 games run better on the M4.

"Barely fast enough to run ..." -- this is a nonsense statement for any computation that doesn't have some sort of real-time constraint.  If you're willing to wait long enough, even a 680x0-based Mac could run an LLM... if it had enough storage.  The main limitation that will prevent the computation from running at all is storage capacity.  LLMs require a huge amount of state (gigabytes), and if that state doesn't fit in memory it has to be kept in offline storage (disks, flash memory, etc).  Moving it offline would make it many orders of magnitude slower, thus increasingly impractical.  Computers which have no means of accessing large storage couldn't run LLMs at all, but otherwise its just a matter of how long it'll take.  In practice a 680x0-based Mac would probably suffer a hardware failure before it delivered a useful result... but barring that, it would eventually produce an answer.

9secondkox2 said:

Still pretty bummed about m3 ultra instead of current gen - especially in terms of GPU performance. Pretty lame decision. I thought the m3 chips were produced on a more expensive, less reliable version of the 3nm process. 

I would bet there were deeper technical considerations that lead to this.  Not everything is visible from outside, and there are tons of details that could make an M4 Ultra not possible or not sensible or not practical.  You have no idea if it was a lame decision, or the only possible decision.

The M3 process may have yielded lower, however the reason for that is that the M4’s process was actually a simplification.  It’s possible that those simplifications made the ultra impractical on the M4 process, or that they needed to move to the new process as fast as possible and stripping out the ultra connector was a way to make it possible.  Or plenty of other potential reasons.  It may well be a resourcing challenge where the choice was to slow down the M5 generation for marginal gains between an M3 and M4 ultra.  So many considerations, and yet internet opinions…