So then I’d close up the Mac Pro line and shift everyone to the Mac Studio. Just about every use case for the Pro you mentioned can be handled via Thunderbolt.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple Starts Selling Refurbished 2023 M2 Ultra Mac Pro at a Discount
- Thread starter MacRumors
- Start date
- Sort by reaction score
So then I’d close up the Mac Pro line and shift everyone to the Mac Studio. Just about every use case for the Pro you mentioned can be handled via Thunderbolt.
The fact there are certified refurbs so soon after the first release should be a clear message to Apple to add external GPGPU support with Silicon, drastically increase the CPU options and add a new Afterburner 2.0. You put in a much improved FPGA beast that can shared across a unified bus and with up to 2TB of DDR5 memory and CXL 3.0/PCIe 5.x or 6.x and people will actually buy them.
It's obvious Pros would want CXL 3.0 and have Apple join the rest of the industry.
It's obvious Pros would want CXL 3.0 and have Apple join the rest of the industry.
Wait…are you saying they’re offering a free iPhone 13 mini with the refurbished 2023 M2 Ultra Mac Pro?
For the millions spent on R&D / Tooling etc ... I'd be surprised if they break even.
Apple replaced OpenCL and OpenGL/Vulkan with Metal. Metal entangles both of those objectives into a coupled stack. OpenCL is computational engine independent. Apple tossed that 'feature' aside. Metal is basically an embrace-extend-extinguish substitute for open standards. Apple got rid of Nvidia , but tore a multiple pages out of Nvidia 'playbook' . It is about a taller wall and herded (proprietary ) deep optimization work for Apple's GPU/NPU.
Does Metal have to be 'so entangled'? No. Is Apple going to do the work? Probably not. ( didn't really need to completely deprecate OpenCL either. Or at least hand off responsibility. Kind of like XQuartz. ) . Very similar issues at the driver level. There isn't an abstraction for a "graphics GUI" device in DriverKit as there was in IOKit. Can be a PCI-e card but need something that is independent of graphics specifics.
Intel's problems with getting their GPU drivers up to speed and applications to optimize for is suggestive as to why Apple's draconian stance has merit. It is getting more code optimized for Apple GPU quicker. There is no other choice as to where to spend optimization time on. So any Mac application's time/resource budget just goes to them.
Intel has OneAPI but their fingers are on the scale there to tilt it toward their offerings. Nvidia has a giant CUDA wall. Microsoft keeps coming up with their own "Compute" APIs. AMD has HIP. Kronos still caught with a committee structure that has all of those folks are members and keep the Open Standards pulling in multiple directions. Khronos managed stuff is almost like GPLv3 with Apple now; can't run away fast enough. As long as most of the major players are Balkanized , Apple isn't going to swim upstream against that 'current'. They do their own wall also.
CXL seems to be in somewhat similar boat. Originally it was an "Intel advocated" thing, but it looks like Apple has missed the boat on the industry consensus that is coalescing around that. It is a hardware standard, but common hardware will probably work better with common standards on top. If Apple isn't going to do the common APIs then why ware they going to do the common hardware?
Is Apple going to miss out on the high end AI/ML training compute revolution? Probably . Do they care? I don't think so. Inference they are trying to cover with their own stuff and APIs along with some more standard libraries sprinkled on top of the Apple library foundation.
What would be a substantive value add for the Mac line up ( not just the Mac Pro) would be an upgrade to the hypervisor/virtualization libraries that Apple provides to support assigning a PCI-e card that macOS doesn't want to bother with solely to a guestOS that does want to put in the effort. Apple is pointing at virtualization only as the supported path to other OS instances. It is a hypervisor feature they are lacking that is present elsewhere. It would be a little bit more limited on TBv4 systems , but over long term TBv5 would ease that limitations and newer Mac Pros would work even better.
If Apple doesn't want to do something... give someone else the 'hooks' to do the work instead.
You mean a connector that would bring PCIe to outside of the motherboard? Something like… Thunderbolt?
Not really.
Thunderbolt backhaul is x4 PCI-e v3.
MP 2023 backhaul is x16 + x8 PCI-e v4 .
That is a HUGE gap. To say those are equivalent is a farce. There is already cards that are 'limited utility" under TBv4. ( any x8 PCI-e v3 ). The number of useful cards in the > x8 PCI-e v3 is only going up in number.
Putting aside Slot 1 and Slot 3 of MP 2019 because those are covered a different way , the MP 2023 is a bandwidth net increase over what the MP 2019 did with the remaining six slots. Thunderbolt wasn't any more viable in 2019 for what the MP 2013 fell short on all those years . That is even less true that the workstation norm is moving to PCI-e v5. "Leaning too hard" on Thunderbolt is an even worse call now in 2023 than it was in 2013.
And if Apple wanted to stop ignoring CXL, Thunderbolt is in a even deeper swamp of misguided "equivalent substitute".
There is a camp of folks that "the only useful PCI-e card is a GPU card". I didn't not outline that use case or philosophy at all.
IF the M3 Ultra bumped the backhaul to a symmetric two x16 PCI-e v4 then that would crush the MP 2019 backhaul outright with obsessing over the slot of old Slot 1 and 3 . [ the on die backhaul to the GPU cores in M2 is way higher than old Slot 1. ]
Apple could make a bigger dent in the Max memory cap if they had a decent chiplet design and could go to three dies. [ for while there was talk of a quad die set up, but that seems doubtful. Especially if Apple has 'short term' , non-longetivy issues with TSMC N3B ). I don't think Apple is going past pushing monotholic dies as chiplets in M3 mainly because M3 probably isn't going to last too long. Over intermediate term it is cheaper. But if at M4/M5 can could do a larger package then it isn't going to have paid well to have dumped the Mac Pro chassis earlier.
It is also somewhat pointless to push for > 200GB of RAM when there is no ECC to go with it. Apple probably isn't doing much there in the short/intermediate term either .
[ Until the decoupled themselves from actual laptop optimized dies that will probably continue to be an issue for next several years. ]
M3 may not even get an Ultra ( for Mac Pro or Mac Studio) depending upon how long a roll out would be. To long and it would might be better to skip to M4 series (and reverse the rollout order. )
If fixated on the modular RAM / Graphics, then there really isn't volume here to support diverging far off of what the rest of the Mac line up is paying for. Intel era the server market largely paid for the chips Apple just happen to use for Workstations. Those largely just 'fell out" as a side effect. If the main driver is gone, then should be little reason to expect that now.
That’s two years after introduction. With this one, what’s it been, three months?
How do you get rid of an investment like the mac pro in three months?
The new Mac Pro. Faster than ever...at getting tossed into the refurb pile when users realized it's not upgradable and they might as well get an equally fast Mac Studio for half the cost and 1/10th the space.
wow, refurbs already? which ones are these? Luke Miani's, Max Yuryev's, iJustine's? Coz I know my boy MKBHD wouldn't buy one. He's still rocking an M1 series MacBook even though he can surely afford to "upgrade".
Metal is far more encompassing. OpenGL is deprecated on all platforms of UNIX, Linux and Windows. Vulkan oversight is from Nvidia who themselves doesn't even want to fully implement it. It's the Java of Graphic stacks and keeps becoming even more and more entangled.
Metal is very clean and extensible with distinct areas 3rd party OEMs don't have to bother with unlike OpenGL which has a massive hit on performance.
Metal does more for computational processing than OpenCL via Metal Shaders
AMD has HIP which is growing very well as a replacement to CUDA.
Apple is serving OS X consumers. Windows with DX12, etc., is serving Windows users. Linux/FreeBSD UNIX, etc., have Wayland a poor man's OS X WindowServer and they utilize support for legacy OpenGL and complain constantly that AMD doesn't extend HIP support for the Radeon 5000, all 6000 and even the rest of the current 7000 GPUs when it is primarily being developed for MI Series of Compute Solutions. They complain support is for the Workstation class GPUs when everyone knows those are the folks with large professional support contracts who pay for the driver support.
Nvidia has no CPUs for consumers so it better support CUDA across both consumer and enterprise clients or see no one using their GPUs for anything other than games that aren't Professional line cards.
AMD and Nvidia offer different approaches because they have no OS to maintain.
OpenCL is available via POCL and more projects via the LLVM Project who has oversight of OpenCL while recognizing Apple's Trademark and ownership of it.
Apple has no incentive to open up Metal as some Agnostic platform independent series of services for any and all cards.
Please like, comment, subscribe, ring the bell, stay to the end of the video, use my affiliate links, buy me a coffee, thanks to today’s sponsor…..
They can have all the GPU performance they want, if they don't support CUDA, they rule out a lot of potential customers. And it won't support CUDA, unless you can pop in an Nvidia GPU. It's been 16 years, a lot of stuff won't be ever ported to anything else, no matter how fancy slides other GPU manufacturers make. And this is the market segment where the Mac Pro itself would be the cheapest part. But I don't expect them to realise this until they only consider youtube wannabes as pro users.
Yeah outside of literally trying to sell fewer so they can just shrug and say "no one wants this", I don't understand the pricing at all. It's less capable than the Intel model, and they decided to price it higher. Mind-boggling. It still has a useful place in the lineup, but not if you're basically paying for another Mac Studio just to get some slots.
The pricing? Apple charges a substantial amount for RAM ( $400/32GB ) and storage ( $400/TB) .
The 2019 model started with 32GB of RAM and 0.5TB of storage. To level up the 2019 to the 2023 model you'd need to add 32GB of RAM and 0.5TB of storage.
The gap is $1000 and there is $600 of that gap right there.
The Intel model started with 8 CPU cores and now getting minimum of 24 Cores.
GPU about the same thing. 580X and something substantially better than a 580X.
so getting at least 8 more cores (if only count P cores ) and a better GPU for $400 at 'Apple prices'.
( note to go from 60 to 72 GPU cores costs $1K ( so about $83/core) and CPU cores on M2 Pro costs $300 for 2 cores ( $150/core ) )
As for no one wants this, Apple stated at the first Mac Studio introduction the the most common selling CPU option on MP 2019 was 16 cores and the most common GPU bought was W5700. If you go back to see what a >= 64GB RAM , 16 core CPU , W5700 priced at in 2020, you'd find the new Mac Pro is priced a several hundred under that. That is where the pricing comes from.
Free to not like the prices, but Apple is rigidly consistent with their BTO pricing across all of their products. (e.g, the SSD costs per TB has been the same for more than several years. It doesn't change. ). The entry level RAM and storage capacity went up, so the price goes up.
In the bottom 30-50% of old MP 2019 demographics they are pushing more value/$ of what you get out of the box. At the top end of the old spectrum, yeah Apple is going to loose people. But if the objective is to sell enough if follow the most common bought configurations then probably going to still capture most of those.
The $3,000 price gap between Studio-Ultra and Mac Pro.
xMac Studio Echo III $1,550
And again over half that gap is already gone for dramatically lower bandwidth and three fewer slots. If skip buying 2-8TB of Apple SSD can close the rest of the gap.
Part of Apple's problem is that they left the MP 2013 comatose at the $2,999 price point for a very long time and the $2,500 (and lower once 'used' was the only option) for the MP 2009-2012 models that lots of folks are price anchored at far lower price points. MP 2019 price point probably had the post transition in mind for where Apple set the entry price target. Apple wanted to move it up. (didn't hurt that mainstream high end workstation market was moving things up also. )
Performance wise, out of the respective newly open Apple shipping box .. an 8 core Xeon 3223-series with 32GB RAM and a 580X is outperforming this new box? At what?
With the newer system you have at least 8 more CPU cores ( that are better). A better GPU . 32GB of more RAM , and 0.5TB of more storage space.
Can say it has more possible potential if spend even more money on additional parts for the MP 2019 entry model. Yes, you can by 2 generations older 32GB DIMMs , a PC market 5700 or marked down 6800 , and a 1TB drive for less than $1K. But the gap won't be $1K large anymore.
If Apple added an "More than Ultra" SoC option to the Mac Pro , that isn't going to lower the entry price point. ( e.g. a 3 die chioplet ).
It’s a Mac Studio with the same processing power and some PCIe slots… that costs between $2,000-3,400 more for an equivalent configuration.
…And we already know Apple has to be making SOME kind of a profit on the Mac Studio to begin with. (Hell, they’re probably saving costs by just reappropriating the same chassis as the Intel Mac Pro.)
So, yeah. Overpriced. And a massive profit margin.
And also useless. Just buy a Mac Studio and an external Thunderbolt PCIe enclosure.
Nobody needs 4 PCIe slots. Especially when you can’t put a dedicated GPU in one of them.
The Mac Pro is so incredibly stupid. And a cynical and thoughtless product on Apple’s part.