MP All Models 2023 Mac Pro! [MERGED]

[mode11]

I think it's fair to say that Apple Silicon makes for incredible laptops, given that it excels in performance per watt. They cost a lot of money, but given the excellent performance, battery life, screens and general build quality, you get a lot too. The only downsides are the much-noted price of memory and storage upgrades, and well as being almost impossible to repair by anyone other than Apple (again... $$$). Yes AppleCare, but aside from being a further expense, second-hand purchases without AC will be risky (which will surely affect resale value).

The desktop is less convincing, when you take into account the PC you could buy (and especially, build) for the cost of a Studio Ultra. It also lets you scale in the way that best suits your needs e.g. a 1x M3 Max-level CPU + a 2x Ultra-level GPU. Apple can only scale in whole Max-sized chunks (and then, only up to 2x Max).

The issue with big iron is likely a) the limited parallelisation of most software, and b) higher core counts tend to mean lower clock speeds. I'm not up on M3 benchmarks, but a quick Google turned up this analysis by Tom's Hardware, which reckons the M3 Max is about level with a 14900K in GB6 single and multicore: https://www.tomshardware.com/pc-com...it-405-ghz-challenge-raptor-lake-in-geekbench. Impressive given the much lower power draw, but as the K is 'just' a 24 core chip, there's no real reason to be comparing the M3 Ultra to 56-96 core x86 processors, which as you point out, are highly specialised for very parallel (and likely multi-user) workloads. The 14900K is also just a refreshed 13900K; Arrow Lake (and Zen 5) should bring greater advances on the x86 side.

 

[ondioline]

If that's the case, GB6 multicore (which shows the Ultra chips and the M3 Max as EXTREMELY powerful - right up there with very big Windows workstations) is probably closer to the truth in an average workstation use case than GB5 multicore

Yes I agree. Most things people are doing are going to be multi-threaded rather than hugely parallel. Probably dealing with consistent stack and heap allocations, rather than a single process initialization and exit.

Of course I can think of things that are actually parallel, like map reduce. But even with a threadripper, who is running MR on a desktop? lol

 

[avkills]

Regarding PCIe lanes, it is yet unknown whether the M3 Max or Ultra is actually going to have increased capacity. Although we do not even know if the Mac Pro is going to be a thing with M3 chips. I sure hope so, because looking at the M3 Max GPU scores, an Ultra should put it right around or above currently what is available as a single slot solution for 2019 Mac Pros.

What would really be crazy is if Apple released a GPU only PCI card for the Mac Pro; although not sure how that would work with the memory system.

 

[deconstruct60]

Regarding PCIe lanes, it is yet unknown whether the M3 Max or Ultra is actually going to have increased capacity. Although we do not even know if the Mac Pro is going to be a thing with M3 chips. I sure hope so, because looking at the M3 Max GPU scores, an Ultra should put it right around or above currently what is available as a single slot solution for 2019 Mac Pros.

Pretty good chance given the M2 Ultra has non-symmetrical PCI-e backhaul allocations x16 + x8. Usually Apple is pretty OCD when it comes to symmetry. That was odd. The M3 shows up and it is x16 + x16 backhaul into the discrete PCI-e switch like the MP 2019 and things would be back to Apple's usual OCD symmetry set up.

Folks chirping that it has to be PCI-e v5 or bust. Probably not. Even more so since it looks like the RAM implementation is relatively conservative ( same tech) and Apple is leaning more on 'smart' on die caching to squeeze out performance ( that smarter cache is likely an uptick on internal network bandwidth consumption. ) NPU bandwidth consumption going up. CPU consumption going up . even more so stretching the internal network.

What would really be crazy is if Apple released a GPU only PCI card for the Mac Pro; although not sure how that would work with the memory system.

What would be less crazy is if Apple released a whole Mac on a PCI-e card. virtual Ethernet network back to the host over PCI-e v4 and would work just fine with somewhat that can distribute workload out to nodes "in the cloud".

6-pin + bus power is 150W. The M3 Max fits into that envelope (and certainly the M3 Pro does) and for some jobs the new Max does what an Ultra used to do. Add two to a Mac Pro plus an host Ultra are are in 4x Max land for workloads can be split up. ( of course Apple would have to 'buy into' the notion that 6-pin power wires were not 'evil'. )


Make the virtual Ethernet platform neutral ( run on Linux/Windows host) and dramatically increase the places can deploy the card to.

Doing a card that only worked in Mac Pro is doomed. There are exceedingly few places to deploy to. Extremely likely not a viable long term market due to small size. If Apple did reverse course and do a discrete GPU card , then they would at least need to pull in some other Macs along with the Mac Pro into the deployment picture.


P.S. Apple already does 'virtual Ethernet' over Thunderbolt connections (likely layered on PCI-e data movement before encapsulated onto TB for transport). Virtual Ethernet over PCI-e slot directly shouldn't be a major 'moonshot' technology project. There are mini-clusters that already work with current software when hook them up via TB. There is work to be done here , but shouldn't be any major paradigm disconnects from what they are already doing now.

 

[deconstruct60]

Honestly, they're selling a USD$700 device, with a $25/month cellphone plan that won't use a lot of data because it's not browsing web pages etc, and no doubt there'll be an accessory ecosystem, and fashion tie ins...

If there conversational/generational AI has a substantial cloud component it to it, then that $25/mo is not going to just 'data' on the network. It is also for compute on the network services also. With a battery that small ... somewhat skeptical running large, billion element LLM on the device for 100% of the workload all the time when interacting with it. ( Yeah it is asleep when you don't wake it with a tap. )

Also pretty good chance a relatively small company like this is outsourcing some of the smart services and/or infrastructure here. That will cost money on an ongoing basis.

 

[deconstruct60]

Having just watched Humane's AI Pin demo video, I honestly don't think Apple has much of a future in anything.

Humane just threw the entire core of Apple's business - devices to *look at* your digital self (messages, notes, interpersonal relationship management) under a bus, and when you don't need that stuff to be on the device you use to also do "looking at" tasks, that really shifts the lockin gravity away from Apple.

That is a whole lot of "I'm mad at Apple" wishful thinking. This thing might make a dent in the smartwatch market , but cellphones ? Probably not .

How do you take a decent deliberate picture with a camera with no view finder? There is zero way of doing decent photographic composition with this thing. None. Cellphones laid waste to the digital camera market not solely because they had just screens , but because they had three things... compute power , a view finder , and increasingly better cameras. This thing doesn't disrupt that at all because missing a very critical piece if actually care about the quality of the picture you are taking at all. I suppose the super genius AI on this thing will always do the best photo composition for you automagically. ( Probably not. ) . That is the only way not having a viewfinder doesn't matter ... because the AI is doing it all for you.



It does have a screen... just a monochrome screen on your palm whose's detail fidelity they really don't want you to pay much attention to. So the color that is on the smartphones hasn't been critical all along? Probably not.

Also pretty much presumes that users has pretty much a flat chest also. Which isn't always true over a billion or so folks.

It is a threat if Apple keeps shipping a relatively weak Siri 'smart assistant' subsystem. It won't be just these folks eating away at Apple though if Apple continues to stumble there.

Just as 'smart' (or smarter assistant) , plus 'tap to talk' earbuds and cellphone is just as 'screen less'.

The Silicon Valley 'spin' on this product is largely anchored on the "cellphone screens are primary source of bad things" assumption. Screens in and of solely themselves are not really the primary root cause the problems. It goes to a zone where there are no apps at all to 'throw the baby out with the bathwater'. It isn't solving/addressing the problem as much as avoiding the problem. 'Feature phones' aren't clawing back major share from smartphones either.


They have made it very easy to 'hot swap' batteries probably because many folks will need to hot swap every day if actually use the device substantially. ( one factor driving larger phones is that many folks have expectation of not having to charge or swap batteries for a day's use. So likely not going to get deep traction with those folks either. )


NFC payments / transit / tickets ... fail. QR code ticket / boarding passes ... fail.

No one "cool" "influencer" or "trendy" is going to be seen dead with an iPhone once this pin is released.

ROTFLMAO ... trendy cool influencer people who don't take selfies ... Gee how many of those are around???? The whole subculture is about self centered indulgence and promotion. The value proposition when they can't take a picture of themselves is going to implode pretty fast.