Testing Apple’s 2026 16-inch MacBook Pro, M5 Max, and its new “performance” cores

M5 Pro Max's "performance" CPU cores definitely aren't just rebranded E-cores.

See full article...

 

[tipoo]

I'd also like to know if the M5 Pro/Max give up any efficiency in low-to-idle use compared to the E-cores.

I wonder if they waited till it was overlapping on a perf/watt curve while being able to just scale up beyond the E-cores, or if there's some sacrifice to efficiency in coasting range. With a 7-issue width, caches between the two, and around 70% the performance of the Super Cores, I would guess for light use it can't get down as low as the E-cores can which just have less execution hardware for light work, but with power gating (which Apple has heavily used since AS and before) it might be close, I'd curious to see these perf/watt curves charted out between E/new P/SC.

 

[telenoar]

Great, thorough review of the CPU/GPU aspect. Thanks! I loved that it focused on video encoding.
My users generally fall into two categories. "Anything since the M2 Pro is plenty fast", and "need moar speed!!!!". For the latter, we've been keeping tabs on Apple's GPU performance and whether they could/would ever catch up with nVidia's flagship RTX of the month.

I wish the Mac Pro was a very different machine — in its current form it is indeed quite useless. But the lack of advancement in the Ultra chips (relative to the new M generations they reliability ship out every year), is rather disappointing.

Having the GPU semi-separated from the CPU using the same "Ultra" technology makes a lot of sense; it remains to be seen if and whether new Ultra options come out and how many dies they'll link. I know I'd be interested in a 80-core or 160-core M GPU option.

And there's also the downside that Apple's newest processor requires Apple's newest OS — those of us living in the real world of creative apps / media / entertainment often can't be on the latest version. A 15% speed increase ain't worth it if the app your livelihood relies on crashes every day.

 

[THT]

Still curious on whether the lineup consists of 3 chips: 1) CPU chip, 2) Pro GPU chip, 3) Max GPU chip, and there is only one silicon bridge per Pro or Max chip.

Could it be the M5 Max is the CPU chip, and 2 Pro GPU chips and uses 2 silicon bridges?

Then, does the CPU chip have 2 silicon bridge (Ultra Fusion) ports, so you could do two CPU chips and 2 Max GPUs chips, using 3 silicon bridges for the M5 Ultra?

 

[KingArthur10]

Love these writeups. Makes me long for the old Pentium v PPC teardown articles from Ars of yesteryear. Spent many an hour scouring through CPU architecture deep-dives.

 

[thomsirveaux]

Still curious on whether the lineup consists of 3 chips: 1) CPU chip, 2) Pro GPU chip, 3) Max GPU chip, and there is only one silicon bridge per Pro or Max chip.

Could it be the M5 Max is the CPU chip, and 2 Pro GPU chips and uses 2 silicon bridges?

Then, does the CPU chip have 2 silicon bridge (Ultra Fusion) ports, so you could do two CPU chips and 2 Max GPUs chips, using 3 silicon bridges for the M5 Ultra?

My understanding is that there's only one Fusion connection, between the CPU/etc die and the GPU/etc die. The 40-core Max GPU is its own die, and not just two 20-core Pro dies fused together.

 

[tipoo]

TLDR; So they took some performance cores and over clocked them a little and gave them a new name.

Emphasis on the DR. The new Performance cores are an all-new 7-issue design which is not some insubstantial overclocking, the new middle core architecture is truly a third and different core type.

The E cores and Super Cores are the same, latter with a rebadge from Performance, but the new Performance cores are truly new. This is what's interesting here. The Super Cores aren't really overclocked, it's just the move from E cores to a new middle core that's interesting.

Apple marketing may not have done it service with the double retcon to remember.

 

[Unclebugs]

"Apple is using the same external design for these laptops that it has been using since 2021—it’s aging pretty well, and we still mostly like it, especially compared to late-Intel-era MacBook Pros. There’s just not much else to say about the design that hasn’t been said." With good reason. Apple and the world are waiting for solid-state electrolyte Li-on batteries. When that happens, it will be worth changing the "external design." My money is on a a much thinner and lighter laptop when the new battery tech shows up. Until then, my M1 Pro 16-inch MacBook Pro looks and works just fine. Battery life is still livable for now.

 

[johnsonwax]

So, the new core arrangement appears to be producing more or less the same MT results and power consumption results as you would have expected if they had extended their old P/E core arrangement. (This is orthogonal to the chiplet/GPU change - they could have done that with the old core arrangement.)

So we haven't really seen the payoff with this decision, at least in benchmarks. One possible payoff may be the LogicPro situation where the old E cores weren't really up to handling a low-latency, near realtime compute task where the new M cores look like they should be able to. We'll see if LogicPro allows those Ms to be used, because that's then adding a lot of compute to that product which shouldn't suffer from a drop in S for M cores. And we don't really know what's in store for the Studio and Ultra. If an M5 Ultra M core is on par with an M2 P core, then Apple may be at a point that they can throw a 4K or 8K video stream at it, which they wouldn't have been able to do with the old E cores. And with the Max now at 614GB/s, if the Ultra is roughly the usual doubling, that'd put it in the 1.1TB/s memory bandwidth, and able to feed all these cores.

Geekbench gives the same scores for Pro and Max, so we know it's useless at assessing the benefits of that memory bandwidth difference, so everyone should keep that in mind.

I'd be curious how LogicPro, FinalCutPro, and clang all fare between the Pro and Max, and possibly Ultra assuming we get an M5 sometime in the reasonable future (seems more likely they'll skip M4 Ultra with the new chiplet approach). My sense is that Logic is most sensitive to latency, which is why it refused to use the E cores, FCP is most sensitive to bandwidth and GPU, and clang just to having more cores, but it too tended to not use E cores. So this change up may disproportionately affect each of those products, and I think Pro/Max comparisons may reveal that.

Also Apple claimed roughly double SSD read/write speeds, which will also have an impact on those products, and I haven't seen much testing of that yet.

 

[apoclypse]

One of the issues if you are a Logic Pro or Ableton Live user on Mac is that both don't use the Efficiency cores at all. This should hopefully mitigate that issue and increase performance overall for those DAWs.

It was always a pain having to explain to people (especially during the M3 era) that they needed to buy a Mac with more performance cores even if the chip was older. I had someone ask if they should get an M5 or an M4 Pro. The answer is an M4Pro as that has 8 performance cores, while the M5 only has 4.

 

[kpluck]

If you’re interested in a slightly wider-ranging review of the new MacBook Pros, I’ll point you toward reviews of the M1, M3, and M4 generation models

As a M1 Max MBP owner I appreciate the links to previous performance tests. Unfortunately, at least for the M1, your M1 benchmarks are all different versions of what your are currently using so an apples-to-apples comparison is problematic.

I can't imagine the difficulty involved in trying to test multiple generations of any processor with a full, consistent set of benchmarks. But there has to be a way to have at least a few basic benchmarks between the different generations of hardware so those of us with older hardware, the ones most likely to want to upgrade, have some valid information to base a purchase decision on.

-kp

 

[mjuarez1977]

Great article, best article up to now on performance of these new M5 Pro and M5 Max processors.

I do feel the need to complain: I feel like ArsTechnica (and most journalists out there, honestly) should go back to using zero-based Y axes. It's so confusing to look at a chart with such dramatic results... only to realize the "zero" is actually 12K, and then having to reframe everything mentally to understand.

Anyway, except for that, great job with the article.

 

[johnsonwax]

As a M1 Max MBP owner I appreciate the links to previous performance tests. Unfortunately, at least for the M1, your M1 benchmarks are all different versions of what your are currently using so an apples-to-apples comparison is problematic.

I can't imagine the difficulty involved in trying to test multiple generations of any processor with a full, consistent set of benchmarks. But there has to be a way to have at least a few basic benchmarks between the different generations of hardware so those of us with older hardware, the ones most likely to want to upgrade, have some valid information to base a purchase decision on.

-kp

This kind of ties in with my observation that the M5 Pro/Max having the exact same cores but double the memory bandwidth on the Max creates an opportunity to test the impact of memory bandwidth on various tasks - something we don't get to see often - because when do you ever get exactly the same cores and clocks but twice as many memory channels? I've got a few communities that have noted how much faster an M Pro/Max is than a 285K or 9950 relative to what benchmarks suggest it would be and like yeah, that's 8 RAM channels vs 2 right there. How critical is that to FCP, or clang, etc. In some, I'd expect almost not at all and in others loads.

 

[johnsonwax]

Great article, best article up to now on performance of these new M5 Pro and M5 Max processors.

I do feel the need to complain: I feel like ArsTechnica (and most journalists out there, honestly) should go back to using zero-based Y axes. It's so confusing to look at a chart with such dramatic results... only to realize the "zero" is actually 12K, and then having to reframe everything mentally to understand.

Anyway, except for that, great job with the article.

Never feel guilty about zero based axes. As a retired data scientist, I'll tell you that's a hill worth dying on.

 

[tipoo]

Makes you wonder if these "efficiency" cores have significant advantages over the "performance" core - or if they are simply smaller - aka "space saver" cores.

I have a M3 pro laptop, and I was originally swayed by the concept of having 6 "efficiency" cores saving battery life.

New's flash - I don't think they do.

The battery life results are also looking very good, and there are some articles about how splitting the chip allows larger total die area at higher yield and lower cost, so maybe the e-cores did also more optimize for die space as well and we're not giving up any efficiency trading them for the new middle cores


Edit: Wow, even with no E-cores the idle is actually lower than the equivalent M4 tier
https://creativestrategies.com/research/m5-max-chiplets-thermals-and-performance-per-watt/

 

[fridder]

Assuming we see updated Studios at WWDC with an m5 Ultra, I wonder if this new strategy will allow them to do a more interesting update, like all super cores etc. With better thermals and no need to worry about the battery life they could do some interesting things

 

[johnsonwax]

Assuming we see updated Studios at WWDC with an m5 Ultra, I wonder if this new strategy will allow them to do a more interesting update, like all super cores etc. With better thermals and no need to worry about the battery life they could do some interesting things

I kind of wonder at what point the typical Studio workload is more dominated by GPU than CPU. With the old approach, CPU/GPU had to scale together, but now they don't.

 

[dmsilev]

I kind of wonder at what point the typical Studio workload is more dominated by GPU than CPU. With the old approach, CPU/GPU had to scale together, but now they don't.

Yeah, you could imagine an M5 Ultra version of the Fusion interconnect with four rather than two ports. Then theoretically you could have anything from 3 CPU dies and one GPU to 1/3. A straight-up upgrade of the M3 Ultra would be a 2/2, but there would at least in principle be those extra possibilities. Not sure the 3/1 would have all that many customers, but the 1/3 definitely would be attractive for GPU and ML and so forth uses.

 

[MilanKraft]

Love these writeups. Makes me long for the old Pentium v PPC teardown articles from Ars of yesteryear. Spent many an hour scouring through CPU architecture deep-dives.

Indeed! I remember well the legendary, Bad Andy-fueled Mac forum threads as well, during the era when everyone was debating whether the PPC architecture could be scaled for performance each generation the way Intel's best at the time could... or whether the unthinkable might ever happen... Apple moving to Intel. Funny to consider the progress since then, with Intel firmly in the rearview. Progress marches on, as they say.

There was also an industry guy whose user name I've forgotten, that used to do battle with BA and others, over the finer details of CPU design, manufacturing, etc. For us mere mortals, those who only knew enough about CPU design and photolithography to be dangerous, much popcorn was had. Those were the golden days of internet forums and the rest. Much of the ad filth, profiling, and other shit-tacular practices didn't exist yet (not coincidentally, social media also didn't exist yet... and Amazon was a trendy, Seattle-based book seller.)

 

[Shiunbird]

I do feel the need to complain: I feel like ArsTechnica (and most journalists out there, honestly) should go back to using zero-based Y axes. It's so confusing to look at a chart with such dramatic results... only to realize the "zero" is actually 12K, and then having to reframe everything mentally to understand.

Apple is the worst on that, especially comparing across architectures.

X is 1.2x faster than Y at Z. Thats really only 20%, and they always ofc focus on tests of large delta.

I had a G5 Quad and the first gen Mac Pro and the Quad had better memory speed scores but Apple only showed scores where they were better.

And ugh... charts where a bar that is 3x longer than another representing 10% more ))))))))

Or log scale charts that somehow try to distract us from the fact that they are log

 

[wrylachlan]

This could actually be really good for the Apple Watch. Hear me out. Up to now, the efficiency cores have been carrying the water for MT performance. Apple has consistently chosen “do more with what we have” over “use less to do the same thing”. And the efficiency cores have been climbing the performance ladder to levels of performance that just aren’t really needed in an Apple Watch.

With the new performance cores filling the multi-threaded role that might free the E cores up to lean in even harder to efficiency. The next gen E cores could unlock some serious battery life.

 

[Pixy Misa Mk II]

Emphasis on the DR. The new Performance cores are an all-new 7-issue design which is not some insubstantial overclocking, the new middle core architecture is truly a third and different core type.

The E cores and Super Cores are the same, latter with a rebadge from Performance, but the new Performance cores are truly new. This is what's interesting here. The Super Cores aren't really overclocked, it's just the move from E cores to a new middle core that's interesting.

Yes, the point there is that the "new" Super Cores are simply renamed M5 Performance Cores.

It's the Performance Cores that are actually new.

Apple marketing may not have done it service with the double retcon to remember.

Yep. That's exactly the problem we're seeing here. Even people who understand the technical details are having trouble with the terminology.

 

[Xign]

Measuring the M5 Max’s CPU power consumption with the powermetrics command-line tool, average power consumption during our Handbrake video encoding test is about 23 percent higher than M4 Max, and because of that increase, the chip uses just a bit more energy overall to do the same work.

This is a bit concerning to me. These are still ultimately laptops and battery life still matters. If this metric is to be taken at face value, it seems that the new M5 Pro/Max mostly gives modest improvements in multi-core performance, but also costs the same amount in power efficiency? I'm not sure I understand why it justifies a new design then. If it uses more power I would have expected a better increase in performance given it's a new design.

Comparing the official specs for the M5 Pro / Max and the M4 Pro / Max though it does suggest the battery life should be similar between the two, so I wonder what gives. It could be that for normal usages the battery consumption is the same, or that Apple is using the high/low end of the testing report range to hide some decrease in battery life between the two. The M5 Max does have a mysterious 2 extra hours for "video streaming" but I'm guessing that's something to do with the video decoding rather than raw CPU compute.


Edit: Nevermind, I just re-read what was written. Given that it took less time to run, and the metrics tool was measuring power usage not energy usage, it made sense that the M5 Max used more power. So seems like as a package both use similar power consumption. I'm still curious about how the idling performance work for the new "performance" cores though.

 

[AER]

It’s a shame the review lacks any real world benchmarks other than Handbrake. I’d love to see things like Lightroom export times, zip compression times, OCR times, etc., and maybe also something involving AI inference. These really need benchmarks that give an idea of real world performance, given the architectural changes.

 

[Legatum_of_Kain]

This is great, I'm just mad that Apple doesn't play nice with Linux.

I would get their stuff in a heart beat if they helped run Linux on their machines once they're EOL (7 ish years) and no longer getting OS updates.

 

[spindizzy]

The benchmarks would be more useful if they included the M1 Pro and the M2 Pro/Max as these are the generations where people will be considering an upgrade. If I have an M3 or 4 Pro/Max I'm less likely to be looking to update unless I'm purely performance driven regardless of value.
Also would have been nice to see the generational improvements in Ultras becnhmarked.

 

[jonbob_newcastle]

Love these writeups. Makes me long for the old Pentium v PPC teardown articles from Ars of yesteryear. Spent many an hour scouring through CPU architecture deep-dives.

Kids these days complaining about film reviews on ars...
I remember when there were nowt here but CPU architectures as far as t'eye could see. It were a simpler time, back when t'background were black and t'text were white.

 

[internet2000]

The naming scheme is a travesty. Could they really not find a better word for these new "middle" cores? Though if that's the only thing people can complain about, they're in a good spot...

 

[THT]

My understanding is that there's only one Fusion connection, between the CPU/etc die and the GPU/etc die. The 40-core Max GPU is its own die, and not just two 20-core Pro dies fused together.

Yes, it is likely. The x-rays will be very interesting to see.

For the M5 Ultra however, one of the chips has to have 2 Fusion ports so that they can double up, and the Ultra will have 3 Fusion bridges, if an M5 Ultra ships. Wonder how they solve the scaling issues with the Ultra, too.

Having 2 Fusion ports allows them to scale a bit in both directions too, be it more CPU cores or more GPU cores, instead of just doubling both CPU and GPU cores like in prior Ultra versions.

 

[YetAnotherAnonymousAppellation]

I'm not interested in Apple products for a variety of reasons (not dissing them, just don't want them), but I would be interested in a "deep dive" into how the new fleet of ARM-based Windows machines fit into the existing Windows hierarchy. Any chance of seeing something like this as an Ars article?

 

[rcombs]

I'd be very curious to see single-core benchmarks run on an M5 E-core and an M5 Pro P/M-core (which should reflect real-world performance of background tasks, and give an idea of how the new "medium" cores stack up against the existing tiers), either via core affinity or QoS (eg taskpolicy -b [program]).

 

[Dano40]

Great, thorough review of the CPU/GPU aspect. Thanks! I loved that it focused on video encoding.
My users generally fall into two categories. "Anything since the M2 Pro is plenty fast", and "need moar speed!!!!". For the latter, we've been keeping tabs on Apple's GPU performance and whether they could/would ever catch up with nVidia's flagship RTX of the month.

I wish the Mac Pro was a very different machine — in its current form it is indeed quite useless. But the lack of advancement in the Ultra chips (relative to the new M generations they reliability ship out every year), is rather disappointing.

Having the GPU semi-separated from the CPU using the same "Ultra" technology makes a lot of sense; it remains to be seen if and whether new Ultra options come out and how many dies they'll link. I know I'd be interested in a 80-core or 160-core M GPU option.

And there's also the downside that Apple's newest processor requires Apple's newest OS — those of us living in the real world of creative apps / media / entertainment often can't be on the latest version. A 15% speed increase ain't worth it if the app your livelihood relies on crashes every day.

If you are working on Adobe or Autodesk family of programs. You won’t have to worry it will take those companies five years to make any sort of improvement to their programs, if past history is any indication. Speaking from experience….

 

[just6979]

"M5 Pro and M5 Max both use the same 18-core CPU die, but Pro uses a 20-core GPU die, and Max gets a 40-core GPU die"

Are the different dies, or just different bins of the one big die? Are they actually different sizes (in mm²)?

 

[thomsirveaux]

please, for the love of god, throw out Geekbench """benchmarks""" out of the door. It has been very well documented, all over the Internet, that Geekbench is a joke of a benchmark for ages. Geekbench also makes users expect unrealistic results (e.g the MacBook Neo with the A18 Pro being thought to have single-core performance of M3/M4, but in reality it only barely matches M2; or the M5 Ultra somehow beating out the flagship workstation Threadripper 9955WX for some reason).

Geekbench is one tool in the toolbox. I think it does a reasonably good job of representing what normal "bursty" usage feels like. But there's a reason it's not the only thing we run and I'll have more to say about that in, oh, say, 12 hours or so.