M3 Pro/Max have lower memory bandwidth

[senttoschool]

Tim Cook is an MBA, which explains why he is so mediocre, and why he makes mediocre decisions when it comes to product functionality.

Cook is such a mediocre MBA suit, that it wouldn't even surprise if did something like hiring McKinsey to made decisions about the M3 product line.

I don’t dislike MBAs. I dislike when when an engineering org is dominated by MBAs.

No one can convince me the MBAs didn’t way over complicate the SKUs to extract every ounce of profit for the M3.

There’s not a single SKU that looks like good value for my use.

The M1 Pro and M2 Pro were great values for me. M3 generation leaves a lot to be desired for people who need a powerful CPU but none of the extras.

 

[bobmans]

Definitely cost-cutting going on here.

They also swapped out 2 P-cores for 2 E-cores and decreased the number of GPU cores. All the cost-cutting resulted in a transistor count decrease from 40B to 37B.

Probably they managed to get their cost/chip to be the same as for the M2 Pro due to this. They're getting more chips/wafer due to die shrink + transistor count decrease, combined this could offset the wafer cost increase of the 3nm process.

 

[krell100]

Meh. I had high hopes for M3. I'm definitely disappointed with both the performance and the SKUs. It performs much worse than what I expected and it's more expensive than what I expected (for the SKU that I wanted).

I'll wait until M4. I feel quite good keeping my M1 Pro longer.

Huh? We have no idea how any M3 SOC performs until they get into users hands.

 

[DrWojtek]

I really don’t see the points here. The memory bandwidth has gone from absurdly high to very high. Probably due to more strategic architecture. Not a problem.

The CPU performance is up by about 20% in single-thread. Perf/watt remains phenomenal. GPU performance is up even more.

RAM is a bit unorthodox sure. So? Pick the amount you want.

One thing I can agree on is that it’s expensive. But having an MBP from 2018 I know how reliable they are. And this one will last a long, long time, if that is what you want.

The thing that annoys me a bit is I would like to go very high on CPU but quite low on GPU. I dont need a strong GPU, and I wouldnt want to pay for it.

 

[JohnnyGo]

I really don’t see the points here. The memory bandwidth has gone from absurdly high to very high. Probably due to more strategic architecture. Not a problem.

The CPU performance is up by about 20% in single-thread. Perf/watt remains phenomenal. GPU performance is up even more.

RAM is a bit unorthodox sure. So? Pick the amount you want.

One thing I can agree on is that it’s expensive. But having an MBP from 2018 I know how reliable they are. And this one will last a long, long time, if that is what you want.

The thing that annoys me a bit is I would like to go very high on CPU but quite low on GPU. I dont need a strong GPU, and I wouldnt want to pay for it.

Couldn't agree more

My guess is the number of GPUs now are a driver of performance as much as the "quality" of new GPU architecture.

In relation to p-cores vs e-cores, everyone likes speed but 95% of the time we're reading / typing on a computer, how much speed do we really need. My guess is that MacBooks were already fast enough, hopefully this change will allow for even better battery life. That is something I'm always looking for.

 

[Eric_Z]

I really don’t see the points here. The memory bandwidth has gone from absurdly high to very high. Probably due to more strategic architecture. Not a problem.

The CPU performance is up by about 20% in single-thread. Perf/watt remains phenomenal. GPU performance is up even more.

RAM is a bit unorthodox sure. So? Pick the amount you want.

One thing I can agree on is that it’s expensive. But having an MBP from 2018 I know how reliable they are. And this one will last a long, long time, if that is what you want.

The thing that annoys me a bit is I would like to go very high on CPU but quite low on GPU. I dont need a strong GPU, and I wouldnt want to pay for it.

I don’t remember if it was on Anandtech or somebody enrolling around at Realworldtech (or both?).

But they never could saturate the bus on the M1 (And iirc M2), they were never even close. No matter what pathological example they used. Of that holds true for the M3, then reducing the bandwidth is a solid engineering move. Reducing complexity and cost for something that has no purpose than being a big number on a piece of paper.

 

[livmatus]

It is because Apple is using 6GB and 12GB LPDDR5 chips in addition to 8GB and 16GB. 18GB = 3x 6GB, 36GB = 3x 12GB, and 48GB = 3x 16GB. Each channel is 50GB/s. M3 has 2 channels for 100GB/s, Pro has 3 for 150GB/s.

welp that's up to discussion ... memory chips are usually in pairs

M1/2 has 2 chips
M1/2 Pro has 4
M1/2 Max has 4 but bigger
M1/2 Ultra has 8 (2x Max)

so that 18GB option is more likely binned from 6GB chips (do they even make 6GB chips?), and 36GB option are binned 12GB chips...

... OR ... they really have 3 chips and it would(?) explain the lower memory bandwidth opposed to M2 series

 

[reklex]

Tine to start dismissing specs and look at real world performance.

Don’t forget there will be diminishing returns with most of these configs and M1 and M2 Pro/Max chips are still incredible.

If the M3 line fails to deliver we still have an excellent refurbished and used market to choose from.

 

[senttoschool]

Huh? We have no idea how any M3 SOC performs until they get into users hands.

Based on the slides Apple showed, I was expecting more. If I wanted the best CPU in a 16" MBP, it'd cost $4,000 minimum. The extra GPU cores will never get used by me. It makes the laptop a bit less efficient as well.

Heck, if I just want more than 6 P cores, I'd have to $3,500 minimum.

Previously, I could get the best CPU on a Macbook for $2500. That's a $1000 - $1500 increase.

I think what Apple saw in M1 and M2 generations were that far too many people were happy buying M1 Pros. In my engineering office, all the devs use the Pro chip. Almost no one uses the Max chip. I think Apple saw that and decided to gimp the Pro CPU and up the Max CPU by a bit.

 

[lowkey]

Well looks like I’ll be keeping my 14” M1 Pro MBP a bit longer.

2 generations old but still with 2 and 3 more performance cores than an M3 Pro. That’s really pretty disappointing development of what seemed like such an exciting possibility with Apple Silicon.

I wouldn’t be surprised if the M3 Pros only just beat the M1 Pros in multi core performance.

Add to that the M3 Max in a 14” is up over $5.5k. Seriously Apple?

 

[leman]

I don’t remember if it was on Anandtech or somebody enrolling around at Realworldtech (or both?).

But they never could saturate the bus on the M1 (And iirc M2), they were never even close. No matter what pathological example they used. Of that holds true for the M3, then reducing the bandwidth is a solid engineering move. Reducing complexity and cost for something that has no purpose than being a big number on a piece of paper.

One can't saturate the 400GB/s interface with the CPU, simply because the CPU L2 cache was not fast enough. But it's not there to support the CPU, it's there to support the GPU. And it's fairly easy to saturate the bandwidth using the GPU. I mean, my 32-core M1 Max can easily do it. Depending on the workload, the 30-core M3 Max could get contained by the 350GB/s interface (then again you get things like memory compression etc., which make it a bit less trivial).

 

[Basic75]

I don’t remember if it was on Anandtech or somebody enrolling around at Realworldtech (or both?).

But they never could saturate the bus on the M1 (And iirc M2), they were never even close.

They showed that all CPU cores together on an M1/2 Max can't get past using more than about 200GB/sec RAM bandwidth.

 

[Pasha13]

I didn't think the M3's would come out soon and was worried about my M2 Max, but you know, I'm still happy with mine. Sure M3 is an improvement, but its not so transformational even on 3nm and there are these quirks like memory bandwidth you mention on the M2 Pro and (baby) Max. I am curious about actual game performance with these new GPUs though and the ray tracing. I thought we'd see a game announcement also today.

Not to mention the massive cost increase when comparing a fully spec'ed out M2 Max vs M3 Max.

 

[DrWojtek]

Based on the slides Apple showed, I was expecting more. If I wanted the best CPU in a 16" MBP, it'd cost $4,000 minimum. The extra GPU cores will never get used by me. It makes the laptop a bit less efficient as well.

Heck, if I just want more than 6 P cores, I'd have to $3,500 minimum.

Previously, I could get the best CPU on a Macbook for $2500. That's a $1000 - $1500 increase.

I think what Apple saw in M1 and M2 generations were that far too many people were happy buying M1 Pros. In my engineering office, all the devs use the Pro chip. Almost no one uses the Max chip. I think Apple saw that and decided to gimp the Pro CPU and up the Max CPU by a bit.

Yup, the Pro was the best choice up until now. Now CPU completely scales with GPU.

 

[dalestrauss]

Not to mention 8 core and 11 core and 12 core and 14 core and 16 core options

I'd almost bet it would have been 8/12 but they got more useable chips for binning at 8/11...

 

[senttoschool]

Yup, the Pro was the best choice up until now. Now CPU completely scales with GPU.

Apple could have left the Pro at 8 P cores so it's at least not an obvious downgrade/sidegrade. I might have gone for an 8 P core M3 Pro laptop this generation. Instead, I'm skipping it entirely.

 

[MRMSFC]

M4 will have its own “features”, for example, crazy throttling and overheating or something else. But then the fanboys will explain to you that this is the way it should be, that this is the norm, etc.

Let’s wait until the M4 is out to **** on it, unless you enjoy punching that strawman.

 

[senttoschool]

Let’s wait until the M4 is out to **** on it, unless you enjoy punching that strawman.

I agree. I think M3 generation is a dud in general and definitely a dud for me. I have more hopes for M4. Hopefully N3E node and more architectural optimization rectifiy this mess.

 

[Pasha13]

Based on the slides Apple showed, I was expecting more. If I wanted the best CPU in a 16" MBP, it'd cost $4,000 minimum. The extra GPU cores will never get used by me. It makes the laptop a bit less efficient as well.

Heck, if I just want more than 6 P cores, I'd have to $3,500 minimum.

Previously, I could get the best CPU on a Macbook for $2500. That's a $1000 - $1500 increase.

I think what Apple saw in M1 and M2 generations were that far too many people were happy buying M1 Pros. In my engineering office, all the devs use the Pro chip. Almost no one uses the Max chip. I think Apple saw that and decided to gimp the Pro CPU and up the Max CPU by a bit.

Indeed.

A few months ago with the student discount (which anyone could get) it was $4279 for a fully loaded 16" MBP and 2TB storage after the $150 apple gift card they were offering with a qualified purchase.

Now, if you still want at least a 38-core GPU, it's $1270 more, since 96GB of unified memory is no longer an option with that config (you'd need to get 128).

If they did offer the highest-spec M3 Max and 96GB of unified memory, it would theoretically be $5k before tax, or $870 more than the comparable M2 Max version (again, assuming 2TB of memory).

IMO, the M2 Max was a steal just before the Fall. Not exactly an apples to apples comparison, but now you're looking at like a 20-30% cost increase depending on how you want to view it. Do you get a 20-30% increase in performance? Not likely.

 

[CWallace]

Apple could have left the Pro at 8 P cores so it's at least not an obvious downgrade/sidegrade. I might have gone for an 8 P core M3 Pro laptop this generation. Instead, I'm skipping it entirely.

The feeling I got from the keynote is Apple is trying to better match SoC configuration to workload needs to maximize power efficiency (and therefore increase battery life).

So with M3 PRO having 6E+6P, for workloads that primarily need the E cores, they have two more to work with and therefore the power draw is lower than if they had to be run on the P-cores (like on M1/M2 PRO). If you have a workload that primarily uses P-cores, then MAX is there with 12 P-cores and 4-E cores.

 

[leman]

The feeling I got from the keynote is Apple is trying to better match SoC configuration to workload needs to maximize power efficiency (and therefore increase battery life).

So with M3 PRO having 6E+6P, for workloads that primarily need the E cores, they have two more to work with and therefore the power draw is lower than if they had to be run on the P-cores (like on M1/M2 PRO). If you have a workload that primarily uses P-cores, then MAX is there with 12 P-cores and 4-E cores.

I think there are multiple motivations. What you describe is certainly one. But we also see Apple move to a new hexa-core cluster architecture, most likely to improve die area utilization and reduce manufacturing costs. In general, I feel that reducing costs is a common theme with these new chips.

 

[ApplesAreSweet&Sour]

I don’t dislike MBAs. I dislike when when an engineering org is dominated by MBAs.

No one can convince me the MBAs didn’t way over complicate the SKUs to extract every ounce of profit for the M3.

There’s not a single SKU that looks like good value for my use.

The M1 Pro and M2 Pro were great values for me. M3 generation leaves a lot to be desired for people who need a powerful CPU but none of the extras.

Just as with iPhone, more and more of the genuine, no asterisk improvements are moved to the highest of the high-end, and less and less to the mid-tier and entry-level products.

More and more with every new chip generation.

Knowing Apple, I expect there to be a few more compromises for most of the M3 and M3 Pro MacBooks Pro that we'll only know about once they are properly reviewed and taken apart.

The different configurations and prices seem particularly terrible this year. Like you almost cannot mix and match without steep price increases, much larger than before?

M1 Macs are still the goats of the Apple Silicon era when considering price and value.

 

[Basic75]

But we also see Apple move to a new hexa-core cluster architecture

The M3 Max has two 6-core P-core clusters?

 

[Allen_Wentz]

This is by far the most complicated spec scheme I've seen from Apple since I switched in 2008.

At least with Inte... ...How does any professional operate with such restriction in this day and age?

Easy. It ain't rocket science. Apple makes choices available, choose. I chose 96 GB RAM and 2 TB SSD for my M2 MBP. Now with M3 you can choose even more.

Note that any professional should be getting more than 32 GB RAM for a 2024-2029+ MBP life cycle box, so that means a Max chip, which promptly limits your additional choices to not be some "most complicated spec scheme."

I do agree it can be hard to decide on buying 1/3 more GPU for $300 or not; and hard to decide on 64 vs. 96 vs. 128 RAM. But those choices are not complicated, just difficult. IMO having choice is good.

IIRC having paid $400 for 2 MB of third-party memory in the past, so today's $400 for each +32 GB does not really faze me. I am just thrilled that tech advance allows us UMA RAM, and in such large amounts in our laptops.

 

[leman]

The M3 Max has two 6-core P-core clusters?

So it seems!

https://forums.macrumors.com/threads/m3-chip-generation-discussion-megathread.2393667/post-32679976