Wednesday, March 5th 2025

Apple Unleashes New Mac Studio With M3 Ultra and M4 Max SoCs

The refreshed Mac Studio is here, and it appears that Mark Gurman's reports were accurate once again. The system was updated with the M4 Max and the M3 Ultra SoCs - and once again, that is not a typo. For whatever reason, Apple refused to fit the Mac Studio with an M4-flavored Ultra SoC, instead settling for an undeniably confusing product lineup. The M4 Max, with up to 16 CPU cores and 40 GPU cores, will undoubtedly have the upper hand in single-core performance by as much as 30%, whereas the M3 Ultra will have superior multithreaded and GPU performance, courtesy of its 32 CPU cores and 80 GPU cores. Moreover, the price gap between the base M3 Ultra and M4 Max SKUs will remain the same, despite the former being based on an older generation.

However, the M3 Ultra will allow the system to be configured with up to a whopping 512 GB of unified memory, with memory bandwidth of 819 GB/s. While that number is not particularly mind-bending for a workstation-class system, the fact that the M3 Ultra's 80-core GPU will have access to over half a terabyte of fast-enough memory is a game changer for select few ultra-high-end workloads. Of course, this amount of VRAM is not intended for the average Joe, but the Ultra SoCs were always meant to be a halo product. The M3 Ultra variant can also be equipped with up to 16 TB of storage - at Apple's ridiculous pricing, of course. Needless to say, Apple's performance claims are as vague as always, and interested customers will have to wait for independent reviews and benchmarks to make sense of Apple's confusing SoC strategy with the new Mac Studio.
In terms of connectivity, the Mac Studio shines. The ports array includes dual USB-A, four Thunderbolt 5, HDMI 2.1, 10G Ethernet, and an audio jack on the rear. On the front, the M3 Ultra variant gets dual Thunderbolt 5 ports, whereas the M3 Max gets dual USB-C ports. Both variants get an SD card slot, of course. Wi-Fi 6e and Bluetooth 5.3 take care of wireless networking - no Wi-Fi 7 for either variant, which is rather disappointing. Prices start at $1,999 for the binned M4 Max variant with 36 GB of memory, and $3,999 for the binned M3 Ultra variant with 96 GB of memory. The highest-end variant with the full M3 Ultra SoC, 512 GB of memory, and 16 TB storage costs a cool $14,099. A more sensible variant would perhaps be the one with an unbinned M3 Ultra SoC, 256 GB of unified memory, and a 2 TB SSD commanding a somewhat more reasonable $7,499 price tag.
Source: Apple
Add your own comment

14 Comments on Apple Unleashes New Mac Studio With M3 Ultra and M4 Max SoCs

#2
kondamin
Might be nice to run a pretty complete llm on.
Posted on Reply
#3
Fouquin
These big chip designs are straining even Apple's budget. They aren't "refusing" to equip a Studio with M4 Ultra, they simply don't HAVE an M4 Ultra ready to ship. They've been behind schedule on the big chips ever since M2 swallowed up almost an entire year of R&D time.
Posted on Reply
#4
Darmok N Jalad
GGforeverFor whatever reason, Apple refused to fit the Mac Studio with an M4-flavored Ultra SoC, instead settling for an undeniably confusing product lineup.
The reason is that M4 Max doesn’t have the UltraFusion interposer connection. M3 Max does. It’s believed that Apple won’t make an Ultra chip every generation
Posted on Reply
#5
Easy Rhino
Linux Advocate
What's amazing is that their M1 chip is still incredibly powerful. It runs OSX and all LLMs and other tasks better and with more power efficiency than anything else out there. I hope we get more Linux spins that work on Apple Silicon because there is a lot to gain from the work of Apple's brilliant engineers.
Posted on Reply
#6
hsew
star-affinityAlso think a lack of Wi-Fi 7 is sad on a product released in 2025, but my guess is that the reason for this is they're focused on building their own Wi-Fi 7 chip and it isn't ready yet:

www.macrumors.com/2025/02/21/iphone-17-air-wifi-7-rumor/
If any workload that the target buyer for this machine chooses to run actually depends on a network connection, they'd rather have the more reliable, consistent, and honestly faster 10g Ethernet port anyways.
Posted on Reply
#7
duckface
This is great for generating images with AI, and multi-model models, but I don't know about the processing capacity, Nvidia should have added a method of using the storage memory as an increase for the graphics card's vram
Posted on Reply
#8
Scrizz
duckfaceThis is great for generating images with AI, and multi-model models, but I don't know about the processing capacity, Nvidia should have added a method of using the storage memory as an increase for the graphics card's vram
I mean... that kind of defeats the purpose of doing the things on the GPU to begin with.
Posted on Reply
#9
kondamin
hsewIf any workload that the target buyer for this machine chooses to run actually depends on a network connection, they'd rather have the more reliable, consistent, and honestly faster 10g Ethernet port anyways.
He base configuration isn’t that expensive that You would want to be able to squeeze every ounce of performance out of it.
so having a clean desktop with just 2 wires is just fine.

now if I were spending 14k on a fully decked out machine, yeah I think I’d kinda want 40g sfp ports on it
Posted on Reply
#10
Dristun
Would be interesting to see benchmarks of this vs Digits. Though this can be specd with 512GB, which is wild. Sad that no option exists to build something like this yourself in PC world. Only AMDs upcoming APUs might be able to access so much memory but their GPU part will likely be slower, while Intel has nothing. Oh well.
Posted on Reply
#11
Tigerfox
kondaminnow if I were spending 14k on a fully decked out machine, yeah I think I’d kinda want 40g sfp ports on it
There is no such thing as 40G SFP. 40G is QSFP+ or higher only, it is essentially four links of 10GbE SFP+ in one port. But 40GbE is ad good as dead, atleast 100GbE with QSFP28 is used today, which is 25GbE SFP28 quadruppled, while 200GbE over QSFP56 is widespread and 400GbE over QSFP112 or OSFP is already available.

But yeah, I would expect atleast 2x25GbE on a machine like that. Atleast QNAP is developing 2x10GbE-over-USB4-NICs and allegedly even 2x25GbE.

512GB of soldered down memory is insane. I bet the 16TB of NAND are soldered down, to?
Posted on Reply
#12
Daven
With up to 6 TB5 ports (120 Gbps) for external expansion, I think Apple will use the new Studio as the ultimate high end and discontinue the Mac Pro with internal PCIe slot expansion. Big, bulky desktops are not ‘Apple’ enough. Plus Apple has never really been enthusiastic about large desktops as they have swung widely around the form factor over the years. The Mac Pro is also the last design from the Apple-Intel era which prioritized cooling Intel’s hot Xeons and upgradeable components. The Mx SoCs run way more efficient with soldered components.
Posted on Reply
#13
kondamin
DavenWith up to 6 TB5 ports (120 Gbps) for external expansion, I think Apple will use the new Studio as the ultimate high end and discontinue the Mac Pro with internal PCIe slot expansion. Big, bulky desktops are not ‘Apple’ enough. Plus Apple has never really been enthusiastic about large desktops as they have swung widely around the form factor over the years. The Mac Pro is also the last design from the Apple-Intel era which prioritized cooling Intel’s hot Xeons and upgradeable components. The Mx SoCs run way more efficient with soldered components.
a product you can charge 2K extra for by just putting it in a larger box than the product in the small box.
Totally sounds like an apple thing
Posted on Reply
#14
star-affinity
hsewIf any workload that the target buyer for this machine chooses to run actually depends on a network connection, they'd rather have the more reliable, consistent, and honestly faster 10g Ethernet port anyways.
Yeah, most of the time I also think that's the case for such a machine. :)
Posted on Reply
Add your own comment
Mar 28th, 2025 08:06 CDT change timezone

New Forum Posts

Popular Reviews

Controversial News Posts