Toggle sidebar Toggle sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

Apple Debuts M5 Pro and M5 Max Chips

An end to all confusion with super, performance etc:

Apple’s Fastest, Most Advanced CPU Cores​

M5 Pro and M5 Max feature a new 18-core CPU with 6 super cores and 12 all-new performance cores.

The industry-leading super core was first introduced as performance cores in M5, which also adopts the super core name for all M5-based products — MacBook Air, the 14-inch MacBook Pro, iPad Pro, and Apple Vision Pro. This core is the highest-performance core design with the world’s fastest single-threaded performance, driven in part by increased front-end bandwidth, a new cache hierarchy, and enhanced branch prediction.

M5 Pro and M5 Max also introduce an all-new performance core that is optimized to deliver greater power-efficient, multithreaded performance for pro workloads. Together with the super cores, the chips deliver up to 2.5x higher multithreaded performance than M1 Pro and M1 Max.2 The super cores and performance cores give MacBook Pro a huge performance boost to handle the most CPU-intensive pro workloads, like analyzing complex data or running demanding simulations with unparalleled ease.

www.apple.com

Apple debuts M5 Pro and M5 Max to supercharge the most demanding pro workflows

Apple today announced M5 Pro and M5 Max, the world’s most advanced chips for pro laptops, powering the new MacBook Pro.
www.apple.com
 
Since there seems to be a little confusion on the difference between efficiency, super, and performance cores

From Apple's Newsroom:
M5 Pro and M5 Max feature a new 18-core CPU with 6 super cores and 12 all-new performance cores.
The industry-leading super core was first introduced as performance cores in M5, which also adopts the super core name for all M5-based products — MacBook Air, the 14-inch MacBook Pro, iPad Pro, and Apple Vision Pro. This core is the highest-performance core design with the world’s fastest single-threaded performance, driven in part by increased front-end bandwidth, a new cache hierarchy, and enhanced branch prediction.
Click to expand...
So my reading of that is:

Efficiency cores : Efficiency cores
All-new cores : Performance cores
Old Performance cores : Rebranded as Super cores (since the new mid-tier cores now carry the old name)
 
Last edited:
  • Like
Reactions: sobiloff and Elwe
neilw said:
💯

Seems pretty straightforward to me. Apple is not likely to just eat the higher DRAM (and other) costs. Margins must be preserved. Bumping the SSDs gives them a bit of cover, and honestly it's a reasonable strategy. Pay more, get more. Not everyone wants the larger SSD but that's life.
Click to expand...
And it won’t be popular, but Apple also might up the prices on the M6 redesigned MacBook Pros next year too. They seem to use redesigns with newer components (OLED, touch screens) as another opportunity to boost prices.
 
If we cut through the nice words and marketing terms this is still an extremely impressive chip. The bandwidth jump alone will make many workflows noticeable faster and the addition of the neutral accelerators makes this chip unlock AI workflows that was hampered before. measurable 3-4x boost where it matters (as proven already on the base M5 chip so this is not hyberbole)
Even more interesting is the new approach to the architecture which opens up for a very interesting "Ultra level" chip. mem BW around 1 TB in actual real world use across CPU and GPU is simple not possible elsewhere. Even a 12 channel Epyc does not have that and since this is unified mem distributing load between GPU and CPU.
Apple is well positioned to be the best local AI machines if this comes out soon.
 
  • Like
Reactions: Axelle838, KeithBN and unphased
DrWojtek said:
?? The Snapdragon came along far later than 1th gen M-series. That's like saying, if I were to release a crappy chip tomorrow called "Wojteks Super chip", and I go "Oh you can't blame me it sucks, its only 1th gen"
Click to expand...
To be honest if you never released a chip before it is not totally unfair. But the problem is not which version or gen, it will in the end be what performance you get per dollar, or if it has some special application that is unique.
 
innerproduct said:
If we cut through the nice words and marketing terms this is still an extremely impressive chip. The bandwidth jump alone will make many workflows noticeable faster and the addition of the neutral accelerators makes this chip unlock AI workflows that was hampered before. measurable 3-4x boost where it matters (as proven already on the base M5 chip so this is not hyberbole)
Even more interesting is the new approach to the architecture which opens up for a very interesting "Ultra level" chip. mem BW around 1 TB in actual real world use across CPU and GPU is simple not possible elsewhere. Even a 12 channel Epyc does not have that and since this is unified mem distributing load between GPU and CPU.
Apple is well positioned to be the best local AI machines if this comes out soon.
Click to expand...
It's this without a doubt. With Ultra shaping up for 512GB (or maybe more) at 1.2TB/s bandwidth, nothing else is going to come close, and the PC world is upside down with DRAM market conditions for the next few years at least, so even if Apple hikes its RAM upgrade prices by 2x (which they can probably avoid doing) in response to PC DRAM prices being hiked 3x, 4x+, they will not just be simply a clearly superior value, they will be the only offerings that represent any sort of value.

Consider further if you were to try to meet in the middle with a $35k setup consisting of 3x RTX PRO 6000 giving a measly, but reasonable, 288GB, you have so many drawbacks:

- tensor parallel is hard/doesn't exist with non-power-of-two GPU count
- a NUMA situation for the VRAM pool constrained to, at best, 64GB/s from gen5 x16

Apple could charge $30k for a M5 Ultra 512GB Studio but looking at past releases it's probably gonna be coming in at half that, and it will represent a steal for this use case.

I do wonder if even with the x4 increase with M5 for tensor compute that in practice M5 Ultra will remain bottlenecked by tensor compute for prompt processing. They should be working on continuing to increase the capacity of inbuilt GPU tensor units. Because the single win that the above nvidia setup would still trounce apple silicon on is in batched inference where it can bring those GB202 to bear and dominate on total token throughput. One might need 4 of those GPUs to do it comfortably and convincingly, but it's just a fact that even an m5 ultra doesn't hold a candle to one single GB202 on flops. Will need an M5 extreme to achieve that.
 
Last edited:
  • Like
Reactions: innerproduct
Interested in how these chips are packaged. The old Ultra chips were two Max chips packaged together using a local silicon interconnect (LSI)/Silicon bridge to connect various I/Os together. The description for the M5 Pro / Max sounds like there are two dies which makes me think one of the following
  • The Max chip is 2 Pro chips connected together like an Ultra chip
  • Apple have moved to a chiplet design with both dies packaged on a full silicon interposer (and the Ultra variation will be 3-4 dies on a full silicon interposer)
Will be really interested to see the teardowns & analysis on this.
 
  • Like
Reactions: atonaldenim
DrWojtek said:
?? The Snapdragon came along far later than 1th gen M-series. That's like saying, if I were to release a crappy chip tomorrow called "Wojteks Super chip", and I go "Oh you can't blame me it sucks, its only 1th gen"
Click to expand...
I think he's more just trying to say the 1st gen Snapdragon was only aiming for M1/2 performance, and that the 2nd was probably more targeting M3/4. They were not after the most blazing fast top end optimized as Gen2 is usually more about revising the 1st gen to eliminate inefficiencies found after putting the 1st gen out at scale.

Realistically I'd expect Gen3 Snapdragon's to be what really tries to bring a real fight to the M series. That said, I still expect them to be lagging a bit. The whole reason that the M-series managed to have such a big impression when it came out was that Intel was slacking on mobile for years and was mostly focused on the battle they had at the time in Servers and Workstations with AMD, their mobile offerings had been massively slouching. AMD was winning there as a result of that. When they lost Apple as a customer and focused on mobile they gained ground rather quickly.
 
  • Like
Reactions: DEMinSoCAL
jimbobb24 said:
Slow but a step in the right direction. But ideally you have enough RAM to load the model....
Click to expand...
For sure it is, 600 gb/s gets you into mid tier dGPU territory, that's the same as an AMD 9700 AI Pro. I think people will be happy with the pp/gen tok rates for MoE.

Apple still owns that space, because the only other options for large unified memory pools are Strix Halo and GB10, which have horrible memory bandwidth.

Ursadorable said:
Try running a 40GB model on that 5090 and see how slow it crawls.
Click to expand...
Sure, I do it all the time, I run Qwen3 Next 80b at 50 tok/s with MoE layer offloading, similar with GPT-OSS-120b. I still run circles around an M3 Ultra for gen/pp rates (my pp is like 3000-7000 tok/s), and I can run smaller dense models like Qwen3.5 27b at very high tok rates, where as unified boxes cannot. The only place that would choke is large dense models, which is where Mac's choke too, because they have nowhere near the GPU power needed.

Unified is nice for massive models like GLM-5 that cannot be quantized down enough to fit anything else. But given that the scene is all in on MoE now, layer offloading works like a champ, memory pressure is actually coming down, not going up, and smaller models are becoming amazing (seriously, Qwen3.5 0.8b is wild, it smacks around old 32b dense models). Especially with how Qwen3.5 now handles context/attention... 256k context costs almost no memory.

I could also send my 5090 off to be re-balled with 96gb, since it's fundamentally the same GB202 die as a Blackwell 6000 96gb. It's crossed my mind a few times. I see lots doing it with 4090 (48gb swap).
 
Last edited:
DEMinSoCAL said:
You're saying I can get a base model MBP14 for $1999 like last year? No? You're saying it's $2199 for a base model now? That is the definition of the price going up.
Click to expand...
They are forcing certain specs now. The 16" M5 pro w/ 48GB ram is the same price as the 16" m4 pro.

The M5 max is higher because it forces you to get the 2TB drive. So it didn't go up they are just making you get next tier upgrades

The base model MBP14 starts you off with 1TB drive instead of the 512GB drive.

The prices are the same just Apple is forcing spec upgrades.

Like I've literally been speccing options for the past 2 weeks.
 
  • Like
Reactions: Tdude96 and neilw
DEMinSoCAL said:
To be fair, you're comparing 2nd Gen Snapdragon to 5th Gen M-series. It's not surprising Apple is ahead.
Click to expand...
Last year when Apple released M5, articles supporting that Snapdragon chip (which I think was released about the sam time) were comparing Snapdragon to M4 and it wasn’t that impressive IMO as they generally stick way more cores to make up for lack of performance to even play in the same park.
 
The_Gream said:
Last year when Apple released M5, articles supporting that Snapdragon chip (which I think was released about the sam time) were comparing Snapdragon to M4 and it wasn’t that impressive IMO as they generally stick way more cores to make up for lack of performance to even play in the same park.
Click to expand...
Snapdragon X2 Elite made significant gains over the X1 generation. I'm not arguing more cores, etc. I'm just saying M5 is a more mature product than SD X2 so it's hardly a proper comparison. Since X2 actually performs similar to an M4 in many tests, I'd say they're doing pretty good.
 
  • Like
Reactions: aaronage
I love the naming. Very intuitive. Not sure off better which one is actually better. Professional seems elite. But the word maximum implies the highest you can go. But I have some better nomenclature.

What about MAXIMUM PRO?

I think it really adds umph to it.

How about ULTRA EXTREME PRO MAX?

They should go with that for their top line flagship stuff.
 
  • Like
Reactions: DrWojtek
AltecX said:
I think he's more just trying to say the 1st gen Snapdragon was only aiming for M1/2 performance, and that the 2nd was probably more targeting M3/4. They were not after the most blazing fast top end optimized as Gen2 is usually more about revising the 1st gen to eliminate inefficiencies found after putting the 1st gen out at scale.

Realistically I'd expect Gen3 Snapdragon's to be what really tries to bring a real fight to the M series. That said, I still expect them to be lagging a bit. The whole reason that the M-series managed to have such a big impression when it came out was that Intel was slacking on mobile for years and was mostly focused on the battle they had at the time in Servers and Workstations with AMD, their mobile offerings had been massively slouching. AMD was winning there as a result of that. When they lost Apple as a customer and focused on mobile they gained ground rather quickly.
Click to expand...

Considering Qualcomm themselves were claiming their first gen was faster than the M3 (by cherry-picking certain benchmarks) the I’m not sure how you can say they were targeting M1/M2.

13 pages of comments in this article alone shows that nothing upsets the Apple haters more than a discussion on their processors.

www.macrumors.com

Qualcomm Claims Snapdragon X Elite 21% Faster Than Apple's M3 Chip

Chipmaker Qualcomm has claimed that its new Snapdragon X Elite PC processor is 21% faster than Apple's latest M3 chip in multi-core performance, although questions remain about its thermal profile. Demoing PCs with the new chipset, the San Diego manufacturer told Digital Trends that the...
www.macrumors.com www.macrumors.com
 
  • Like
Reactions: DrWojtek
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back