Binned M4-Pro

[Gabebear]

The Mac mini has a binned M4-Pro option... There aren't any benchmarks I can find yet, but seems like it's a good deal if you would like the double-speed Thunderbolt and better support for large monitors over the regular M4.

I've ordered one, but I wonder what the actual performance downsides will be over the unbinned version. Unlike the M4-Max Apple is allowing any memory config on the binned M4-Pro.

I have two 5K monitors and likely want to add more high-speed storage later.

 

[Andrey84]

What does "binned M4-Pro" mean?

 

[MacCheetah3]

With simple math, it’s a ~16% increase of CPU cores and 25% of GPU cores. However, that assumes perfect scaling and full utilization. Realistically, I’d guess <15% difference. Again, that also depends if you’re actually pushing each processor to its limits.

What does "binned M4-Pro" mean?

Explainer: What is Chip Binning?

You bought a new CPU and it seems to run cool, so you try a bit of overclocking. The GHz climb higher. Did you hit the silicon...

www.techspot.com

 

[Gabebear]

What does "binned M4-Pro" mean?

A “binned M4-Pro” is the M4-Pro with some CPU and GPU cores disabled. “Binning” is when you take slightly defective CPUs and disable the defective parts.

CPUs are designed so they can be binned into different categories. Sometimes perfectly manufactured CPUs have capabilities disabled to make different price points or make them to use less power.

The M4 in the iPad is a binned version where Macs are getting a higher core version.

Many apps can’t use multiple CPUs effectively(they have single thread tasks). The marginal loss of CPU cores generally isn’t a noticeable where having individually fast CPU cores often is.

Losing GPU cores is different. Graphics tasks are generally able to be spread across multiple cores, so the graphics performance does take a noticeable hit.

Apple hasn’t binned the memory support on the M4-Pro(they did on the M4-Max), or the IO support; the M4-Pro supports TB5, which the base M4 doesn’t.

 

[Andrey84]

@Gabebear @MacCheetah3, thanks!

So, this is the binned M4 Pro?

 

[MacCheetah3]

@Gabebear @MacCheetah3, thanks!

So, this is the binned M4 Pro?

View attachment 2447089

Yes.

 

[riptilian]

Enlightening, succinct and practical thread. Thank you all.

 

[MacCheetah3]

There aren't any benchmarks I can find yet

Via Tom’s Hardware, there appear to be Geekbench 6 scores for a few M4 variants.

Geekbench Search - Geekbench

Skimming over the numbers, I settled on these averages for multithreaded/multi-core scores:

M4 (10-core): 14700
M4 Pro (14-core): 22600
M4 Max (16-core): 25700

The potential performance differences from the base M4 or the M4 Pro is up to ~55%.
The M4 Pro to M4 Max is an ~15% increase. The noted Pro and Max, basically, vary by two performance cores, which is the same basic difference between the two M4 Pro offerings. Therefore, we can guesstimate that ~15% increase.

Using U.S. pricing, the cost difference is $1,399 vs. $1,599 (when only choosing the CPU upgrade). That cost differences is ~14%. So, not a bad price as you’re paying for… Well.. What you’re getting or rather what you could benefit.

With all said, including the generally fair value conclusion...

I've ordered one, but I wonder what the actual performance downsides will be over the unbinned version.

Again, ultimately, that will depend if your workflow/usage is an any way near being limited by the CPU (a.k.a. “CPU bottleneck).

 

[Gabebear]

Via Tom’s Hardware, there appear to be Geekbench 6 scores for a few M4 variants.

M4 (10-core): 14700
M4 Pro (14-core): 22600
M4 Max (16-core): 25700

The M4-Pro has double the memory bandwidth of the normal M4; which I'm guessing has a lot to do with the benchmark scores. I want to know the relative difference of the two M4-Pro variants.

The M3-Pro also had binned versions and the CPU benchmarked about 90% of the un-binned version and the GPU benchmarked about 80% of the un-binned version. IF those relative differences are the same for the M4-Pro then the binned version will be just very slightly slower than the M2-Ultra and faster than the M3-Max

 

[AVBeatMan]

Sorry, but is the unbinned better than the binned then? And how do you tell which is which?

 

[mw360]

Sorry, but is the unbinned better than the binned then? And how do you tell which is which?

'Binning' just refers to the process by which manufacturers provide different CPU specs for the same model. The 20-core and 16-core options represent the two bins in this particular M4 line-up. 'Un-binned' is just a (somewhat nonsense) term for the top spec CPU in the range, i.e. the 20-core.

Edit: Sorry I misread the specs and quoted the GPU core counts and not the CPU core counts but the meaning is the same.

 

[Kotsos81]

Sorry, but is the unbinned better than the binned then? And how do you tell which is which?

Yes, the un-binned is the full, better version: it is the one with more cores, either in CPU or GPU, usually in both.

In the M4 Pro example at hand, we have a binned version with 12 CPU cores (8 Performance or simply P cores and 4 Efficiency or simply E cores) and 16 GPU cores, and an unbinned version with 14 CPU cores (10P+4E) and 20 GPU cores. The cores are all of the same type, M4, their number changes: the unbinned version has 2 more CPU P-cores and 4 more GPU cores. More cores, more processing power for the respective CPU or GPU tasks. That is why the more powerful, unbinned version, is more expensive than the binned one. (Likewise, this is why even the binned M4 Pro chip is more powerful than the base M4 chip, which has 10 CPU cores configures as 4P+6E and 10 GPU cores).

In the so-called binned version of a chip the cores are physically there, but they have been deactivated. This happens for one of two reasons (or both): 1) The fabrication of microchips is a very complex and "delicate" process and in the massive number of produced chips, there is always a small portion of chips with a few (e.g., 1-2) defective cores in CPU or/and GPU. In order to not throw them in the trash and lose, as Apple, a large amount of money, they put them in the market as "binned" versions of the chip. 2) It fosters business segmentation of the products, i.e., a cheaper less powerful version and a more expensive, more powerful one, targeting a larger audience and keeping prices close to the different variants (after updating RAM, SSD storage, etc.) to "push" more customers to the more expensive option ("after upgrades, I am only 200$ away from the expensive, better option").

The same concept applies also to M3 Pro, with the binned version having 11 CPU cores (5P-6E) and 14 GPU cores and the unbinned version having 12 CPU cores (6P-6E) and 18 GPU cores, as opposed to base M3 which has 8 CPU cores (4P+4E) and, I think, 8 GPU cores - all being, of course, M3 cores.

Edit: What a great lost opportunity to write "in order to not throw the binned versions to the bin"!

 

[AVBeatMan]

Thanks guys. So performance wise there's little difference and the extra cores just not worth it unless you know you need them?

 

[mw360]

Thanks guys. So performance wise there's little difference and the extra cores just not worth it unless you know you need them?

Well a 20 core processor is theoretically 25% faster than a 16 core processor but only you know whether that's worth cash to you or whether you would be preceptive enough to notice the difference.

 

[Gabebear]

Well a 20 core processor is theoretically 25% faster than a 16 core processor but only you know whether that's worth cash to you or whether you would be preceptive enough to notice the difference.

The higher-end M4-Pro CPU has 16.6% more cores(14 vs 12), the GPU has 25% more(20 vs 16)

In reality it would be really hard to ever even make a benchmark show near that much of a difference because other limits kick in(memory, pixel fill, disk).

Both the memory and disk access rates for both M4-Pro variants(binned and unbinned) are approximately double what they are for the regular non-Pro M4. Those 200% differences will be noticeable in most workflows where the relatively minor difference between the binned and unbinned might not matter at all.

 

[mw360]

The higher-end M4-Pro CPU has 16.6% more cores(14 vs 12), the GPU has 25% more(20 vs 16)

In reality it would be really hard to ever even make a benchmark show near that much of a difference because other limits kick in(memory, pixel fill, disk).

Both the memory and disk access rates for both M4-Pro variants(binned and unbinned) are approximately double what they are for the regular non-Pro M4. Those 200% differences will be noticeable in most workflows where the relatively minor difference between the binned and unbinned might not matter at all.

Yep, indeed, though I prefer not to indulge this 'binned/unbinned' language because there's no need for it. All processor lines are binned and as we can see from this thread it's unhelpful to suddenly start using new terminology which imparts less information than just stating 12-core or 14-core.

 

[DaveEcc]

Yes, stop calling the full version unbinned. It's still in a bin. It may be the top-tier fully-featured bin, but it's still a bin.

 

[Alameda]

Sorry, but is the unbinned better than the binned then? And how do you tell which is which?

Here’s what happens:
Semiconductor chips are made by applying chemicals to silicon, then shining light through a mask and a lens which focuses the light into minute traces onto the silicon. If any dust particle gets in the way of the light, or if there’s any chemical or surface impurity anywhere, that specific spot on the chip won’t work. The larger the chip, the more likely this error occurs.

For instance, if you get an average of 5 defects per 12” wafer, the larger the dies, the more expensive the defect. If you are making small chips and you get 2,500 chips per wafer, your defect rate is probably 5/2,500, or 0.2%. But if you design large chips and you get 250 chips per wafer, the defect rate is 5/250, or 2%. The probability that a given defect hits an ARM core is very high, so I can recover most of the defects.

The chips are tested after fabrication to make sure these defects aren’t present. In a large multicore chip, if an error is present in, say, Core 7, that core can be disabled and the chip will still work perfectly. Of course, the chipmaker can similarly disable cores on perfectly good dies in order to create the number of lower cost chips they want.

I’m simplifying, of course, but that’s the gist of it.

 

[AVBeatMan]

Here’s what happens:
Semiconductor chips are made by applying chemicals to silicon, then shining light through a mask and a lens which focuses the light into minute traces onto the silicon. If any dust particle gets in the way of the light, or if there’s any chemical or surface impurity anywhere, that specific spot on the chip won’t work. The larger the chip, the more likely this error occurs.

For instance, if you get an average of 5 defects per 12” wafer, the larger the dies, the more expensive the defect. If you are making small chips and you get 2,500 chips per wafer, your defect rate is probably 5/2,500, or 0.2%. But if you design large chips and you get 250 chips per wafer, the defect rate is 5/250, or 2%. The probability that a given defect hits an ARM core is very high, so I can recover most of the defects.

The chips are tested after fabrication to make sure these defects aren’t present. In a large multicore chip, if an error is present in, say, Core 7, that core can be disabled and the chip will still work perfectly. Of course, the chipmaker can similarly disable cores on perfectly good dies in order to create the number of lower cost chips they want.

I’m simplifying, of course, but that’s the gist of it.

Thanks for the explanation. If that's the simplified version I'd hate to hear the "full" explanation! 😵‍💫

 

[Alameda]

C

Thanks for the explanation. If that's the simplified version I'd hate to hear the "full" explanation! 😵‍💫

Think of a camera sensor -- they have very small ones used in phones, and larger ones used in DSLR's, as you can seen in the picture below. They're made on silicon wafers, just like other computer chips. If you compare the size of a "full frame" sensor to the 1/2.5" sensor, you can easily see that you can make about 25 of those small sensors using the same silicon area as a Full Frame sensor, right? So the full frame camera sensor should cost 25 times more, right?

But here's the problem: A single defect which makes the sensor useless would cause ONE of those 1/2.5" sensors to be bad, but it would make the entire Full Frame sensor bad. A single defect increases the cost substantially more when the chips (the dies) are larger. This is why those processors get binned -- it lets manufacturers use the chips despite manufacturing flaws. This also saves consumers a ton of money.