M1 Studios are pretty great, M2 isn’t a major jump from M1, once we get Apple Silicon chips on the 3nm node is when we will see a nice leap.
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.
2023 Mac Studio Reviews: Faster M2 Ultra Chip and 8K Display Support
- Thread starter MacRumors
- Start date
- Sort by reaction score
M1 Studios are pretty great, M2 isn’t a major jump from M1, once we get Apple Silicon chips on the 3nm node is when we will see a nice leap.
Interesting (surprising?) that despite an updated chip, the improvement in performance between the original M1 Max and the updated M2 Max and M2 Ultra chips seems to be merely equal to or less than relative increase in the number of processing cores. On a per-core basis it doesn't look like the M2 chips are significantly faster than the M1 chips.
My M1 Max Mac Studio drives 2 Apple Thunderbolt Displays (too low res for you) and one 27" LG Ultra Fine ERGO display (that one is 4K). LG makes TONS of 4k displays for less than $500, but their 5K displays are all over $1,000. However, you can get a refurb in excellent condition for $500:
LG 27MD5KL-B 27" UltraFine 5K Thunderbolt (USB-C)... at MacSales.com
27" LG 27MD5KL-B UltraFine 5K Thunderbolt (USB-C) IPS Display. Up to 5120 x 2880 resolution. Built-in speakers & webcam. For desktops & laptops + power charging up to 94W. Designed for Apple - Mac & PC compatible. *Used, Very Good Condition*
eshop.macsales.com
GPU upscaling to more and more cores is not going well. The Pro to Max bump is 65%, Max to Ultra is about 50%.
Now obviously this is for the specific GB5 Metal tasks, and often becomes less of an issue as you scale a problem outwards (make the problem larger, rather than trying to complete the same problem faster).
But there is the basic problem that co-ordination across long distances costs cycles.
I suspect the time has come for GPUs to start following some of the path of CPUs, namely keep the number of GPU cores the same, but make each GPU core smarter. Local smarts (very local for each core) is easier than long-distance co-ordination.
There are at least two immediate paths that could be followed:
- add a scalar lane to each SIMD, so that scalar work can be done in that lane (and differentiated by the instruction, rather the requiring the overhead of an "if lane 0/endif" that's needed today
- allow for lightweight two-way issue (think Pentium UV dispatch, not fancy OoO machinery)
I suspect going down this path, rather than another round of boosting the number of GPU cores by 50% or so is overall a better match to future process technology. As I understand it (without having studied either closely) this is more or less where both AMD and nV in their individual ways are also headed.
Now obviously this is for the specific GB5 Metal tasks, and often becomes less of an issue as you scale a problem outwards (make the problem larger, rather than trying to complete the same problem faster).
But there is the basic problem that co-ordination across long distances costs cycles.
I suspect the time has come for GPUs to start following some of the path of CPUs, namely keep the number of GPU cores the same, but make each GPU core smarter. Local smarts (very local for each core) is easier than long-distance co-ordination.
There are at least two immediate paths that could be followed:
- add a scalar lane to each SIMD, so that scalar work can be done in that lane (and differentiated by the instruction, rather the requiring the overhead of an "if lane 0/endif" that's needed today
- allow for lightweight two-way issue (think Pentium UV dispatch, not fancy OoO machinery)
I suspect going down this path, rather than another round of boosting the number of GPU cores by 50% or so is overall a better match to future process technology. As I understand it (without having studied either closely) this is more or less where both AMD and nV in their individual ways are also headed.
(a) Are you talking about CPU or GPU?
(b) For CPU we knew that going in! The A15 is essentially the same IPC as the A14. And was designed as such; it was designed to be basically the same design as the A14 but everything optimized for significantly lower power while still achieving slightly higher frequency on the same process. This was valuable (battery life on A15 phones is noticeably longer) and part of the plan.
(There are a few very specific tweaks in A15 that help some specialized code, most notably javascript in Safari; but they are very specific.)
Same transfers to M2. It's basically M1 IPC, but a slightly boosted frequency.
None of this is catastrophic or an indication that Apple is running out of steam. The A15 should have been just an energy-saving tick to be followed by a performing-enhancing tock. It's quit possible that it was never supposed to even be in the M2; the next core would have done that job. But as I keep saying, covid screwed up the timelines and Apple was forced to scramble with what was available.
I think 6×(6K@60Hz) Pro Display XDRs are the sweet spot.
Remind me again, why is such a big scandal that the Mac Pro can't add extra GPUs?
much better deal than the tarted up Mac Mini M2 Pro due to exorbitant upgrade prices. Glad I didn't spring for one earlier.
Honestly a Mac Studio in Midnight color would have swayed me into upgrading. My M1 Max is already stupid fast
Because dGPU wether PCI-E slot or eGPU are gone. Watch the video where basically the Apple Engineer says with the new architecture then GPU cards are thing of the past with the chosen architecture and unified memory.
If you need that level of GPU cards then the new Mac Pro and Studio are not for you and it is off to PC land and either Windows or Linux for you.
considering M1 to M2 wasn't that much of an improvement then not sure why expecting the Max and Ultra to be big improvements per Core.
So you need 1.92TB of DRAM right now? I assume you’re not paying for it yourself once you find a workstation with that much capacity? I’m assuming you’ve decided between Linux and Windows already as well?
I don't understand this at all. If the price is about $0.04/GB, that means ~$40 for 1TB. How is this not price gouging? And more importantly, the failure to charge less for SSD upgrades is one additional factor keeping Macs from being more widely adopted. This decision makes no sense to me.
Attachments
A base or low-end Studio certainly seems like the sweet spot in the lineup if you need more power than the standard M chips. The Mini upgrade pricing makes no sense - it's £100 more for fewer GPU cores and fewer ports than the Studio? And let's not talk about the cost difference of going up to the Mac Pro, for extremely little tangible benefit to 99% of users...
So still getting thrashed by Intel, even in the most favorable bench for Apple? Yikes.
So core count is up 20%, and single-core is up 14%, but multi-core is up 4.8%? I don't think that's quite right.
Other M2 Ultra results average 28322, an 18.7% bump. That sounds more right.
I find it somewhat humorous that some writers see the default 512GB of on-board storage as unreasonable for pro-class users. We have always attached our editing stations to our SAN. We haven't had big drives on our Pros for well over a decade. Local storage is basically for the OS, virtual memory, and cache, and all the big stuff is on our big iron so it can be easily shared, managed, and backed up. Not everyone is a home user making YouTube videos in their garage.
It feels a bit disingenuous to compare the M1 Ultra with 48 GPU cores (capable of 64) to the M2 Ultra with 76 GPU cores (base 60, up to 76). A more fair comparison would be M1-48 vs M2-60 or M1-64 vs M2-76.
Most PC OEMs have similar pricing structures for their physical RAM upgrades. Dell, for example, wants $2200 for 64GB for one of the model servers we use while Crucial will sell it to us for $220 - 1/10th the price. We still ante up for the Dell modules, since we bundle it all into a service contract.
PCI-E Slots.
To make 6 slots of 2 x 16 and 4 x 8 with PCI-E v3, vs PCI-E v4 inside then is about 2.2K in buying external TB chassis, plus they are at the end of the TB bus and higher latency rather then PCI-Slots so you pay a premium. Plus that Aluminium Case ain't cheap.
Basically if you need non-GPU Cards (and move on Apple Engineer in interview basically saying GPU cards not coming) such as Blackmagic SDI or Audio Capture cards then makes better performance in a PCI-E slot then on a TB Bus.
Why are people hyping 3nm so much? I feel like when you look at it even on the Intel side, the Sandy Bridges lasted a LONG time because afterward you only got very minor IPC upgrades. That was what... 32nm? We had 4 GHz easy overclocks back then and when compared to more 6th, 7th, 8th gen Core processors were still able to keep up only with a minor disadvantage.
I think people are setting unrealistic expectations for M3 / 3nm. If you look at A-series CPUs, the advancements in processing power have long slowed down. We're reaching the point of only minor IPC improvements and are thermally limited. M2 uses more power than M1 and part of the performance advantage comes out to higher clocks and higher thermals. M3 isn't going to suddenly deliver +50% performance or thermals because of a process die shrink.
It costs Apple a whopping $25 to go from 512gb to 1TB based on current NAND pricing. And Apple is already saving on the controller as it's built into the SOC. They're using basic NAND in a proprietary package. I get the low storage options on the basic machines, but not on the Studio.