M2 Ultra Mac Studio review: Who needs a Mac Pro, anyway?
- Thread starter JournalBot
- Start date
FWIW, I have an M1 Ultra model, and it doesn't whistle, or make any other unwanted noises either.
Upvote
4
(4
/
0)
The MacPro is just a Studio with some slots, better cooling and more power. Little development costs and so it's easy to sustain. I suspect there is more to come but it was not fully cooked yet.
Upvote
12
(12
/
0)
It's almost as though they were designing it for a VERY PARTICULAR market...
Upvote
15
(18
/
-3)
"Killer feature" means it kills its competition.. It can't be a killer feature if the competition has had that exact same feature for decades.
Upvote
-13
(8
/
-21)
Actually it's a "good riddance Johnny Ive" feature. I'm not sure they fired enough thinness fanatics though.
Upvote
21
(23
/
-2)
What is an issue is co-ordination between the CPUs, eg in the form of snoops, locks, and even occasional TLB flushing. You can see the effects of this in how GB5 multi-core (independent threads) gets almost 2x scaling from Ultra, whereas GB6 (interacting/co-ordinating threads) gets only 1.5x
Apple has a set of patents for various scalability technologies to improve the delays imposed by snoops, locks (primarily a substantially more sophisticated cache protocol) but that stuff has not been implemented yet.
The same issue is present for GPUs. In that case it appears that some of the scalability technologies have been implemented, though in a very preliminary form. In the case of GPUs the primary issues that are somewhat improved are
- affinity (enqueue later work items on cores that did earlier work on the same kernel)
- workstealing (after the initial enqueueing of workitems to cores, allow work to move from a core that is still busy to a core that has finished all its items)
Again the GPU work will probably be improved in the next Ultra. My guess is that the M2 Pro/Max/Ultra were never meant to exist, the transition should have been to a basic M2 then an M3 series based on N3. But covid screwed up the timelines, so Apple was forced to scramble, which included making the best changes they could in the time available to the most obvious limitations of the M1 Ultra.
Upvote
9
(10
/
-1)
No question a Studio is louder than a Mini. A Mini is dead silent till you put your ear right up to its exhaust. A Studio, even without the whistle (which is extremely annoying, but tape helps a lot), makes fan noise which is audible from a couple of metres. It's slight, but it's never completely silent like a Mini.
Upvote
2
(5
/
-3)
You can connect up to 6 Pro Display XDRs. I believe it’s even more displays (8?) with lower resolutions than 6K.
Upvote
2
(3
/
-1)
That description is very generous to Apple. Supposedly, the x16 slots AND 3 of the other PCIe slots are all being fed off of a single PCIe 4.0 x16 connection coming from the second die. That is the sort of connectivity shenanigans I expect on a hobbyist single-board computer, not on a Mac Pro.
I think Apple has to sort out the Ultra + Ultra design or some similar "2 socket" -type config for the Pro to be viable product. There is just no compelling reason to spend extra on a Pro at present (with the possible exception of requiring a rackmount option).
Upvote
3
(16
/
-13)
Keep in mind the M2 Max in the Studio has double the memory bandwidth of the M2 Pro in the Mini. 200GB/s for the M2 Pro and 400GB/s for the Max. (And it doubles again for the Ultra, to 800GB/s)
I think people really underestimate the performance benefit of the increased memory bandwidth, especially given the unified memory architecture of Apple Silicon.
A Mac Mini upgraded to an M2 Pro (12,19,16), 32 GB RAM, and 1 TB of storage will set you back $2199. For the same $2199 you can get a Mac Studio with an M2 Max (12,30,16), 32 GB RAM and 1 TB of storage. You're getting more and faster GPU cores, double the memory bandwidth, more ports, and 10GigE for the same price.
Upvote
36
(37
/
-1)
I guess there may be a use case for a machine with PCI-slots that can't support a PCI-GPU but beyond studio work that probably requires specialist PCI cards I'm not sure what it is.
A Mac Pro with the ability to run multiple M2s would have had a major USP over a machine like the Mac Studio, which is, whilst I definitely lust after one, clearly a prosumer machine. But alas, no.
Honestly, what is the Mac Pro for? All it really has going for it is PCI slots, and there is such a thing as a Thunderbolt PCI enclosure, so it's not like it's impossible to use PCI cards with the Studio.
I am saving up for a Studio (or a Macbook Pro, not 100% decided yet) and am expecting to have got the money I need together by the time the M3 machine launches. Will be interesting to see if there is a Mac Pro by then, and if there is what it's doing to distinguish itself from the Studio.
Upvote
1
(2
/
-1)
Yeah, they wanted to do the M2 Extreme. But the chip would have been too big. TSMC would have had to stand up a new manufacturing line for it. Because Apple only sells 1 million Mac Pros in a very good year, the lack of economy of scale would have made it cost way too much. So Apple was going to just wait another year, but the need to move forward with Sonoma and therefore start the expiration timer on Intel Mac support (and dedicate 100% of their efforts going forward to ARM macOS) forced their hands.
Now as a Linux and ChromeOS guy - meaning not an Apple one - I see this controversy being much ado about nothing. Despite people who claim otherwise, Apple never has at any point produced the fastest or most flexible laptops, desktops or workstations. You have always been able to get notebooks superior in many ways to a MacBook Air. You have always been able to get ultrabooks superior to a MacBook Pro. You have always been able to get workstations and HEDTs superior to a Mac Pro. And as for the workstation and server market, the biggest problem going forward with the Mac Pro won't be power but the lack of ECC memory meaning that the Mac Pro is going to have to be classifed as an HEDT now (as should the Mac Studio).
So, I don't see the big deal. The new Mac Pro is 3 times faster than the 2019 model. That is all that should matter. People who wanted or needed more power than than ... never bought the Mac Pro in the first place or ditched it a long time ago. Take Hollywood. Even Pixar switched to Linux for rendering etc. quite awhile ago.
The M3 Extreme will land when Apple is able to build it on the same nodes as their smartphone and other chips. (People who deny that this is what they are doing ... consider the M2 iPad Pro.) My uneducated guess is that this will be possible on TSMC's 2nd gen 3nm process in 2025. When that happens, then the Mac Pro will have CPU and graphics performance competitive with the best Intel and AMD workstations. The RAM - 384 GB max and non ECC - will still be an issue, but there is a huge difference between 192 GB RAM and 384 GB (256 GB is the cutoff for a ton of applications).
It isn't just Apple who has run into this problem. Google pulled the plug on their ARM-powered Pixelbook when they found out that the costs associated with doing anything but reusing the existing smartphone chip production would cost too much. MediaTek makes "for Chromebook" chips but only using outdated components and legacy fab lines to keep the costs down, and the result is still $600 Chromebooks that perform no better than the $300 ones running Pentium chips. Qualcomm made ARM laptop chips that perform "ballpark" with AMD Ryzen 3 and 5, but they cost so much that no OEM buys them. ARM Holdings is gettting desperate so they are offering to do the R&D to design the chips on behalf of the chipmakers, meaning that they would just have to license and manufacture them, in order to get ARM chips out there that are competitive with x86 pricing.
Now as a Linux and ChromeOS guy - meaning not an Apple one - I see this controversy being much ado about nothing. Despite people who claim otherwise, Apple never has at any point produced the fastest or most flexible laptops, desktops or workstations. You have always been able to get notebooks superior in many ways to a MacBook Air. You have always been able to get ultrabooks superior to a MacBook Pro. You have always been able to get workstations and HEDTs superior to a Mac Pro. And as for the workstation and server market, the biggest problem going forward with the Mac Pro won't be power but the lack of ECC memory meaning that the Mac Pro is going to have to be classifed as an HEDT now (as should the Mac Studio).
So, I don't see the big deal. The new Mac Pro is 3 times faster than the 2019 model. That is all that should matter. People who wanted or needed more power than than ... never bought the Mac Pro in the first place or ditched it a long time ago. Take Hollywood. Even Pixar switched to Linux for rendering etc. quite awhile ago.
The M3 Extreme will land when Apple is able to build it on the same nodes as their smartphone and other chips. (People who deny that this is what they are doing ... consider the M2 iPad Pro.) My uneducated guess is that this will be possible on TSMC's 2nd gen 3nm process in 2025. When that happens, then the Mac Pro will have CPU and graphics performance competitive with the best Intel and AMD workstations. The RAM - 384 GB max and non ECC - will still be an issue, but there is a huge difference between 192 GB RAM and 384 GB (256 GB is the cutoff for a ton of applications).
It isn't just Apple who has run into this problem. Google pulled the plug on their ARM-powered Pixelbook when they found out that the costs associated with doing anything but reusing the existing smartphone chip production would cost too much. MediaTek makes "for Chromebook" chips but only using outdated components and legacy fab lines to keep the costs down, and the result is still $600 Chromebooks that perform no better than the $300 ones running Pentium chips. Qualcomm made ARM laptop chips that perform "ballpark" with AMD Ryzen 3 and 5, but they cost so much that no OEM buys them. ARM Holdings is gettting desperate so they are offering to do the R&D to design the chips on behalf of the chipmakers, meaning that they would just have to license and manufacture them, in order to get ARM chips out there that are competitive with x86 pricing.
Upvote
15
(22
/
-7)
OrangeCream
Ars Legatus Legionis
That was never the meaning when I was growing up... meaning, just like SD cards are a 'killer feature' because the existing user base never grew up with a Mac with one... the term 'killer feature' is now very different than the definition it originally had (VisiCalc in 1979 was considered a 'killer feature' in that people bought a Mac just for that app).
https://en.wikipedia.org/wiki/Killer_applicationThe definition of "killer app" came up during the questioning of Bill Gates in the United States v. Microsoft Corp. antitrust case. He had written an email in which he described Internet Explorer as a killer app. In the questioning, he said that the term meant "a popular application"
The term 'killer app' is being used by me as similar to 'killer feature'
https://userpilot.com/blog/killer-features-saasKiller features are innovative and solve customer problems in a way that no other product does
https://www.linkedin.com/pulse/killer-feature-ori-feldsteinadditional features are worth nothing without customers and users, and the “killer features” the main purpose is to do exactly that - to attract and retain customers. The “killer feature” is the feature that will make a customer knock on your product’s door, and stay there to enjoy many other features.
If a killer feature kills the competition it's because the competition actually isn't the competition since they lack the feature (as it is implemented)
Upvote
4
(10
/
-6)
As far as I'm aware Thunderbolt does not multiplex a particular PCIe generation because it has nothing to do with the physical PCIe links. It tunnels PCIe packets, and that with a certain bandwidth and latency.
In Apple Silicon Thunderbolt never goes through any PCIe links at all but is wrapped internally on-chip from the internal data buses directly and the already complete Thunderbolt signal comes from the SoC and is only amplified and conditioned for the external port.
Only an external Thunderbolt adapter then decides which link generation to use and how many of those to offer to the actual PCIe devices.
So you could just as well consider Thunderbolt 4 to wrap 2 PCIe-4 lanes as 4 PCIe-3 lanes. As far as Thunderbolt is concerned there is no difference.
Upvote
9
(11
/
-2)
Yeah, that's how I wound up with the 2022 Studio. I was in the market for a Mini, but looking at the M1 Minis, once I upgrade the RAM (even just to 16GB), the GPU, and added a dock for the additional ports that I needed for my usage...I was already well on my way to the price of a Studio. Figured screw it, buy the Studio today, it's more machine than I need now but will probably last upwards of twice as long as the Mini (especially at 32GB of RAM).
Upvote
13
(13
/
0)
Would love to see some benchmarks of compilation speed (e.g. Linux kernel or gcc bootstrap) of the Studio vs. a decent x86_64 workstation. For those of us who live in the world of building (and rebuilding) software, Cinebench and geekbench are unfortunately not that representative...
Upvote
12
(12
/
0)
But that is a different test.
Sheer CPU core scaling is its own metric and applies to most parallel workloads, so between low-interdependence, really parallel threads, while interlocking and messaging latencies are a different one which applies to different workloads (and arguably to fewer ones).
How do you know that? Apple usually hold their cards very close to their vest in matters of crucial competitive technologies until they release products which could be analyzed by competitors.
And the M-class SoCs are apparently using quite advanced caching strategies already to achieve the observed performance!
Chip and memory channel affinity is likely one of the major issues there as well, especially with the Ultra and possibly beyond. And bandwidth contention in general.
The alternative theory is that the planned M2 GPU (presumably with ray tracing and geometry shaders) was just too power-hungry and so the entire core generation had to be delayed to the A17 / M3 on 3nm and the A16 / M2 "plan B" was interjected instead.
Last edited:
Upvote
0
(2
/
-2)
All 8 Thunderbolt ports together are about the equivalent of 16 PCIe-4, but at higher latencies and you'd need to re-aggregate them and then you'd have all of them used up.
And the Mac Studio M2 Ultra only exposes 6 of them, so only equivalent to 12 PCIe-4.
Yes, definitely worth a look when it arrives!
Upvote
2
(2
/
0)
Sure, a monolithic M1/M2 Extreme would have been too big to manufacture (as far as I'm aware it would even exceed what the ASML E-UV lithography machines could have supported) so that was never in the running, most likely.
But the crucial issue was according to credible rumours the actual new core design which exceeded specs and so was kicked off the entire A16 / M2 generation, which was then reverted to a mere tweak of the A15 / M1.
Still very nice as we can see now, but unsurprisingly more power-hungry at its higher clocks and core counts.
The 3nm process should solve this issue and we should see the result in the A17, then in the M3. (A- and M-families are actually one and the same family, just under different names.)
Upvote
4
(4
/
0)
I can think of two markets without even trying.
Mac forensics where lots of cheap high performance storage isn't just a suggestion, it's a requirement. Sure you can theoretically do Mac forensics on Windows or sometimes Linux workstations, you lose half your metadata evidence in the transition. Any knowledgeable defense attorney is going to have a lot to say about that.
The other are those in the lower end HPC market that need or want extremely parallel systems and lots of storage, but for one reason or another prefer Macs over PC or Power systems. They want a workstation under their desk for local computation without having to submit their job to larger clusters. Their program can be either Mac specific or architecture agnostic.
Before people object that Mac Pros are expensive!!!!!1111.... so are dedicated forensics workstations and HPC workstations. In fact, they're largely on par cost wise.
Last edited:
Upvote
6
(6
/
0)
Apple is most likely using a massive 98-lane PCIe Gen4 switch (Broadcom PEX88096) with dual upstream connections (Gen4 1x16 + 1x8) to the M2 Ultra. Each M2 Max die has 16 PCIe Gen4 lanes, so 32 total for the M2 Ultra, however, 8 are dedicated to the SSD. The 64 downstream lanes for the 6 full-length PCI Express slots as well as all of the other PCIe devices on the board except the SSD are being fed from the switch. This is nothing new, as the Mac Pro (2019) employed a similar design with a PEX8796 96-lane PCIe Gen3 switch with a Gen3 2x16 upstream connection to the host CPU. Compared to the Intel version, the new Mac Pro offers 50% more PCIe bandwidth via the PCIe switch, and 88% more PCIe bandwidth overall, considering that the 8 integrated Thunderbolt 4 controllers are all backed by their own PCIe Gen4 x4 root port. And not many "hobbyist single-board computers" include PCIe switches that cost $778 on their BOM. That and the utterly overkill 1400W PSU are the biggest reasons for the cost difference between the Mac Pro and Mac Studio.
Here's a list of the PCIe devices in the new Mac Pro and the most likely fashion in which they are connected:
NAND to NVMe controller: Gen4 8x1
98-lane PCIe switch: Gen4 1x16 + 1x8
Wi-Fi and Bluetooth module: Gen1 x1 (98-lane PCIe switch, pool B)
10-Gigabit Ethernet: Gen4 2x1 (98-lane PCIe switch, pool B)
2-port 5Gbps USB 3 controller: Gen3 x1 (98-lane PCIe switch, pool B)
2-port 6Gbps SATA bridge: Gen 3x2 (98-lane PCIe switch, pool B)
Apple I/O card: Gen3 x4 (98-lane PCIe switch, pool B)
2-port 5Gbps USB 3 controller: Gen3 x1 (Apple I/O card)
PCIe Gen4 slot 6: Gen4 x8 (98-lane PCIe switch, pool B)
PCIe Gen4 slots 1-5: Gen4 2x16 + 3x8 (98-lane PCIe switch, pool A)
Total host PCIe lanes used: 32 of 32
Total switch PCIe lanes used: 98 of 98
Upvote
28
(28
/
0)
If you look at the 27” iMac, and compare it to the price of a 27” 5k monitor, it’s dirt cheap.
Upvote
18
(18
/
0)
That, and any time you're doing Real Work the cost of the guy using the machine is a couple orders of magnitude more than the machine. Like for budgeting purposes, with full benefits and overhead and everything, each of the personnel on my project costs me about $250K. Now, their computer is part of that overhead, but the point is that an extra $3K or whatever every two or three years for that computer is a rounding error compared to the total cost of that employee to my project.
So if it delivers any value at all it's really not hard to justify.
Upvote
21
(21
/
0)
OrangeCream
Ars Legatus Legionis
I don't understand. What does having a modular desktop have to do with having access to lots of high performance storage (note I omit cheap, because if we're talking about a $7k Mac, the storage is by default not cheap in the first place)?
But my point is that $800 gets you an 8 bay TB enclosure with support for 3.5" drives. You can get a 4U 24 bay TB enclosure for $2.7k; I'm sure there are products that can stuff even more disks into a JBOD enclosure and available to a Mac Studio.
Again, I don't understand. How is having a Mac Pro going to be more attractive to the HPC crowd than the Mac Studio?
Or are you saying that the Mac Pro as it is isn't going to be attractive to either crowd because it's deficient as a solution?
Upvote
-11
(1
/
-12)
S
Somebody did a breakdown versus the Mac Pro 2019 and it turns out they’re 100% identical in PCIe bandwidth, so you’re not losing anything. They both use a PCIe switch to distribute lanes as required.
Somebody did a breakdown versus the Mac Pro 2019 and it turns out they’re 100% identical in PCIe bandwidth, so you’re not losing anything. They both use a PCIe switch to distribute lanes as required.
Upvote
3
(5
/
-2)
Some of those people were hoping for a 3-year upgrade cycle to bring a faster Mac Pro (check!) along with similar or better memory capacity (boo!). About a year ago, it looked like Apple was going to fail on memory capacity, when talk of an Extreme chip was being bantered about without any talk of ECC. UMA just wasn't going to cut it.
Upvote
0
(0
/
0)
Not quite. I think the killer feature is in reference to other Mac products. Like the Mac Mini having all ports on the back - which is an easily solved issue, but for some users easy access to ports is what makes this a better Mac than other Macs. A bit like how 3.5m jacks aren't on all products so ones with them are seen as more useful to some segments.
Edit: Auto fecking incorrect
Upvote
0
(0
/
0)
Upvote
2
(2
/
0)
I always wonder how much of Apple’s own needs go into planning high end hardware that they sell. I suspect that Apple was probably the biggest purchaser of iMac Pro for its large pool of software engineers.
The Mac Pro with several video capture cards and scads of PCIe4 SSDs would seem to be a great machine for recording content for the Vision Pro. What kind of camera will be used to take advantage of what can be done with the Vision Pro? How many Mac Pros would it take to stream a live sporting event in 4k VR?
Between Apple’s sports ambitions and their own AppleTV+ service I think they will be using a lot of Mac Pros for 3d capture. And I’m sure they are hoping to spur other companies to make content as well which would make a bigger potential market for them as well.
The Mac Pro with several video capture cards and scads of PCIe4 SSDs would seem to be a great machine for recording content for the Vision Pro. What kind of camera will be used to take advantage of what can be done with the Vision Pro? How many Mac Pros would it take to stream a live sporting event in 4k VR?
Between Apple’s sports ambitions and their own AppleTV+ service I think they will be using a lot of Mac Pros for 3d capture. And I’m sure they are hoping to spur other companies to make content as well which would make a bigger potential market for them as well.
Upvote
10
(10
/
0)
What do you use that much RAM for on an Apple machine?
I have a 512G 2U box to save EC2 costs, but that is (1) a linux machine and (2) would cost as much as a brand new BMW, spec-for-spec on the Intel Mac Pro.
I like to develop and write research code on Mac OS, and use various Macs for light but IO-heavy work, but I don't know why anyone needing that much power would ever use a Mac today, especially since support for both CUDA and various parallelization implementations have degraded over the years.
Upvote
7
(7
/
0)
“Refurb” is such a dirty word in most of tech, I avoid it everywhere…..except for Apple.
Apple has built a long standing reputation for selling refurbs that are essentially indistinguishable from new. It may be that many are simply open box with no actual defects, but there are low defects nonetheless.
In fact if I look through the service history of the Macs I have owned, if I did not label the columns and you looked at the data, you would say that the column with more defects must have been the refurbs. No, that would be the ones I bought new. Maybe I got lucky but I have had less defects with Apple refurb than with “new” ones. When it’s like that, and they sell refurbs at 15% off new, the one single remaining reason to buy new is when you need a very specific config, you need it right now, and it is not in the Apple Refurb store.
Upvote
17
(17
/
0)
...this makes no sense whatsoever.
They built a new machine with the express intent of soft killing it? What?
Upvote
17
(17
/
0)
And Apple seems to see that as a valid enough target market to build the Mac Pro for. I’ve read a lot of reviews of the Mac Pro and the usual withering comments afterwards, and there is a definite pattern of puzzled normal home computer users not wanting anything to do with this, and then, often a smaller group of comments from a specific set of studio owners saying “Yes, this is so perfect..my studio will buy 10 or 20 of these.” That is apparently who the Mac Pro is for, not the office and gaming masses, but a niche market that will make bulk purchases of the Mac Pro, where the price is fine because the price is not anywhere near the total cost of the cards and gear they will plug into it.
Upvote
22
(22
/
0)
I don't recall it being model specific. But very machine dependant - one will do it and the identical one beside it won't.
Some have suggested a small piece of electrical tape over a couple of the holes is all it takes to stop it.
Have you tried this on another M-series machine yet? Like a MBP? I'm utterly fascinated to know what the performance is like. Some reports suggest it's practically native, which is jaw dropping if true.
Upvote
2
(2
/
0)
Did you ever use a trashcan Pro? If so, how does the Studio compare?
Upvote
1
(1
/
0)