The Max chips, (M1, M2, M3 12 P-core) has 400 GB/s. The 10 P-core M3 Max is reduced to 300 GB/s. M2 Pro was 200, M3 Pro has 150. For base chips, M1 had 66.67, M2 and M3 100, M4 has 120 GB/s.
If you look at these numbers, M4's 120 GB/s is probably what Apple thinks as the necessary minimum for Apple Intelligence. M3 Pro barely has more.
What you need depends on what you do. I dug into this because I do astrophysical simulations on very long timescales. I repeat the same simple calculations on a large amount of numbers a lot of times. Which means the most time the CPU spends is waiting for RAM. For me RAM speed is king. For you, it could be completely irrelevant. For most people it's not really relevant, that's why the get away with pushing people who need it to more expensive products (or a PC).
I saw some CEO guy say on an interview in 5 years AI will do 90% of your job so you can chill. Dude when an AI can do 51% of my job they've already fired my ass. I will be chilling alright
PBS Newshour did a story on that. To me it seems that the most practical solution is the route Apple appears to be taking. That is, use small language models on device for common simple tasks and large language models off device for more complex tasks. This route keeps most of your data private and is more efficient energy and complexity wise. At leas that is my uneducated thoughts on the subject. You don’t need to consult a team of PhD math professors to total the costs of the groceries you are buying. Your kid can do that for you.I’m not completely sold on the whole AI thing. First of all, it’s super power hungry. Especially for the half baked tripe it’s kicking out right now. Has anyone considered the carbon footprint of all these AI generated responses on Google I didn’t ask for? The processing is using a LOT of electricity for the sake of making bad term papers and cartoonish cat pictures.
I don’t see how this is going to be “carbon neutral” going forward.
It's a bit unclear if you're talking about the NE, the GPU, or the CPU. The bandwidth impact on the CPU p-cores between 200 GiB/s and 150 is minor, and likely zero between 400 and 300.
A little of both. My work is confined to the CPU because that's what other members of the group can work. But the thread is about AI, and how Apple deliberately started to cut back their cheaper hardware's performance years before AI became hyped up.As you can see, aggregate used bandwidth quickly drops off after more than two cores, and never exceeds 250 GiB/s. Maybe that has been pushed to close to 300 GiB/s, but I doubt it.
Now, if you're talking about code running on the GPU or the NE, that's a different story.
Before AI takes over humans, corporations and business will format our brain to follow the specific processes and standardized lifestyle across the globe! For example, if you want to reach a place from A to B, you need to follow the specific path to reach and it will be tutored in your smart devices and you will happily accept it. Like that in every walk of chaotic life, these corporations in the name of standardizing the process, they will prescribe the lifestyle for all human beings that can behave in a predictable manner under the process. This is already accomplished in some way or the other in many fields. It has its own plus and minus but largely help these AI and ML revolution centre around these standard processes. If you try to do something out of the box, you may be termed as a psycho or mentally unfit sort of terminologies. Corporations want predictable behaviour so that they can control all with a touch of a button!Don't worry/look forward to that.
Yes, they might fire you, but no, AI won't magically be able to perform your tasks. Not in 5 years, probably not in 50 years.
Tim, Steve and Jonny? Sort of a quasi-MAGI thing going on there (obligatory NGE reference included).We could have an entire thread just to debate whose engrams to use 🤭🤓
how Apple deliberately started to cut back their cheaper hardware's performance
Yes, Intel has their own version of this and AMD is already using TSMC’s SoIC technology.Isn’t this what Intel is doing with their “tile” design? It also lets them fabricate the less important portions of the chip on cheaper nodes like 6nm.
First of all, geekbench is a terrible benchmark, that intentionally scales badly. Second, you've picked an M3 Max, with 4 more P-cores and 20 GB more RAM.I dunno about that. Overall, each M revision has improved performance.
If we take M2 Pro vs. M3 Pro, for example: p-cores are down from 8 to 6, e-cores up from 4 to 6, memory bandwidth is down from 200 GiB/s to 150. Yet, in every single workload, the M3 Pro wins by between 7% and 47%.
Because it is.So I don't know why a lot of people seem to have the impression that the M3 Pro is somehow a downgrade.
If this is true, then the M3 Max is badly designed, because it lacks E-cores compared to the M3 Pro...Plus, the M2 Pro and especially M1 Pro were starved for e-cores, giving it poorer battery life. The M3 Pro mostly rectifies that.
While it also decreases the P-core count in the base model, to only three. The last time I remember about anyone selling a 3-core CPU was in 2010. No, I don't buy into that they gave it 6 e-cores. It's the same what Intel does: spam E-cores so the benchmarks look good, while holding back on cores people actually use for their tasks. No wonder these are called "cinebench-cores".And M4 makes this moot anyway. It increases the clock and instructions per clock and e-core count. I guess we'll see how they configure the higher ends, but I'm not too concerned.
Care to share? It went over my head, lol!It took me a few minutes to get that....lol.
Doesn’t matter, the M3 Pro is often faster than the M1 Max. The bandwidth reduction everyone whines about doesn’t really seem to be an issue at all.First of all, geekbench is a terrible benchmark, that intentionally scales badly. Second, you've picked an M3 Max, with 4 more P-cores and 20 GB more RAM.
Because it is.
If this is true, then the M3 Max is badly designed, because it lacks E-cores compared to the M3 Pro...
While it also decreases the P-core count in the base model, to only three. The last time I remember about anyone selling a 3-core CPU was in 2010. No, I don't buy into that they gave it 6 e-cores. It's the same what Intel does: spam E-cores so the benchmarks look good, while holding back on cores people actually use for their tasks. No wonder these are called "cinebench-cores".
Star Trek TOS S2E24Care to share? It went over my head, lol!
I was thinking that too, BUT, if the purpose is AI tasks and those tasks occur in burtsts instead of sustained load then it probably doesn't matter. Apple will charge their premium for a ranch-style that can sustain/cool? -- that would actually help prevent "consumer" hardware from being used in a server environment and protect their enterprise profits (if Apple goes for it)Wouldn't this make cooling more difficult? Since the chip would have less surface area to transfer heat.
This article only uses the CPU. The memory bandwith is shared between the CPU, GPU and NPU.Doesn’t matter, the M3 Pro is often faster than the M1 Max. The bandwidth reduction everyone whines about doesn’t really seem to be an issue at all.
Go to EclecticLight.co and read the articles there - this is just one of many -
![]()
Evaluating M3 Pro CPU cores: 4 Vector processing in NEON
Differences in vector processing performance between the M1 Max and M3 Pro, and in their use of power. Their frequency control is more complex.eclecticlight.co
You are correct, but this is likely SoIC-P, which is new and Apple will likely be the first consumer silicon to use it.Yes, Intel has their own version of this and AMD is already using TSMC’s SoIC technology.
You are correct, but this is likely SoIC-P, which is new and Apple will likely be the first consumer silicon to use it.
SoIC-X has been around for a while, and AMD and others are using it for high performance. I think a good way to understand the difference (as I understand it) is in the names, which follow the same protocol as N5P, N4P, N3P, N2P versus N5X, N4X, N3X, N2X.
Apple won’t use SoIC-X for the same reason they haven’t used the X process nodes. They’re not efficient. They will use SoIC-P, for the same reasons they have used all of the P process nodes. They’re the height of efficiency.
SoIC-P was announced in May, there will likely be more details about it in September. See Anandtech for a good summary.
M5 is a good guess, because there’s no sign of it in what we currently know about M4, but it’s not completely out of the question that we could see it in the M4 Ultra.
First of all, geekbench is a terrible benchmark,
that intentionally scales badly.
Second, you've picked an M3 Max, with 4 more P-cores and 20 GB more RAM.
Because it is.
While it also decreases the P-core count in the base model, to only three. The last time I remember about anyone selling a 3-core CPU was in 2010.
No, I don't buy into that they gave it 6 e-cores. It's the same what Intel does: spam E-cores so the benchmarks look good, while holding back on cores people actually use for their tasks. No wonder these are called "cinebench-cores".
This article only uses the CPU. The memory bandwith is shared between the CPU, GPU and NPU.
I'll say one last time: this thread is about ai performance, which utilises the gpu too. A mid level pc has around a 100 for the cpu and 300+ for the gpu. Yes, they have to copy between system ram and vram, which is a drawback compared to apple silicon, but that's mostly affects startup time. Because of unified memory, adding vram in a mac was ridiculously cheap compared to standalone gpus, even with these abhorrent upgrade prices. Cutting the bandwith is just about milking more money from people who'd use their mac for ai without increasing further the ram upgrade prices, they've simply forced them to skip the pro and go for the max.Yes, that's the point. The high memory bandwidth is mostly useful for the GPU cores, not the CPU cores. The decreased memory bandwidth and p-core count has a negligible impact on CPU performance.
QED. Keep geekbench to theards talking about browsing experience, and forget it when talking about specific use cases which scale well with more resources. What geekbench does is kinda pointless, because an m1 is perfectly good for what they try to measure.Some people keep claiming this, without substantiating it.
Yes, that's by design. Very little real-world stuff scales well.
And how is that an acceptable generational difference?My bad. I've updated the post. There are indeed some benchmarks where the M3 Pro is slower than the M2 Pro, but only by 4.6%. Overall, it is faster.
"Only in multi-core" is a strange way to start a sentence in 2024.But it's not. Only in multi-core does the M2 Pro even come close to the M3 Pro (which makes sense, since the M2 Pro has 33% more p-cores), and even then, I haven't found an overall benchmark where it wins. Cinebench R23 and 2024, Geekbench 5.5 and 6.2, Blender: all of them show the M3 Pro at least slightly ahead.
A very specific use case, hmm. Like trying to use a Pro CPU for Pro stuff?You'd have to have a very specific use case, and one that's heavily parallelized, for your notion that the M2 Pro is faster to apply.
I'm not optimistic that the 4 core variant will cost the same as the 4 core M3.But it's not a 3-core CPU.
And that's only for the binned base model. The regular M4 has four p-cores.
That's a valid reason to have 2-4 e-cores. Maybe even 6. But not to have 16 like intel does in certain models. Those are mostly there for the benchmarks.That's not why Apple and Intel give a lot of e-cores. Power efficiency is why. A ton of background tasks don't really need high performance. Very little of what people do on their computers actually needs it, and usually only in bursts.
Cutting the bandwith is just about milking more money from people who'd use their mac for ai
QED. Keep geekbench to theards talking about browsing experience, and forget it when talking about specific use cases which scale well with more resources. What geekbench does is kinda pointless, because an m1 is perfectly good for what they try to measure.
And how is that an acceptable generational difference?
"Only in multi-core" is a strange way to start a sentence in 2024.
A very specific use case, hmm. Like trying to use a Pro CPU for Pro stuff?
It has a huge impact. Higher memory bandwidth and speeds are one of the main reasons GPUs are used for scientific computing and AI calculations. The entry level NVIDIA desktop card has 272 GB/s memory bandwitdh, and everyone is complaining about how low it is, and holds back the card.I really doubt it.
I would, however, be amenable to the argument that not offering a 48 GiB (say) option for the M3 Pro is in part about driving people to go for the Max, even when they don't need some of the Max's other features. But the memory bandwidth? I think that's a stretch. It just doesn't have that much impact.
Don't compare it to Intel or Qualcomm. Compare it to the M2->M3 change. I won't even say compare the Maxes, because they got more P-cores. Or compare to AMD. And I'd take a machine with M2 Pro over M3 Pro in a heartbeat. The benchmark probably most relevant to me is the PassMark Physics Test, and the M2 Pro beats the M3 Pro by 60% in that test.Are you asking "how is 17% / 7% / -4% overall change compared to 9 months before acceptable"? Intel and Qualcomm would love to have such a change in less than a year.
And that's before we get to M4. Apple's overall iteration pace with their ARM designs seems… fine to me?
Because those workloads can run on a toaster. Totally pointless to benchmark them then run around happily that your new cpu has faster cores, then don't feel a damn difference while actually doing that stuff. That's why I'm totally baffled at what Geekbench is trying to do with this new direction.Yeah man. Why focus on workloads that, oh, I dunno, most people are affected by.
Yeah, software dev is lightweight on the CPU, contrary to what most people believe. In 2016 I worked at a company where we got i7s - with HDDs. The IT was totally clueless.Even if you heavily do AI (which, my condolences), most of the time, your machine will be doing a ton of single-core stuff, whether that's background processing, or starting Internet fights in a web browser.
I dunno about you, but I do software development. Only during actual builds does some parallelism really come into play, and even then, a lot of it simply doesn't scale to more than a handful of cores for extended periods of time.