MacBook M1 Max with 32GB memory bandwidth

[t2jd1967]

So after first having ordered an 14" M1 Pro 10/16/32GB/1TB, I has second thoughts as seeing quite a few folks on the shipment thread ordering the 24GPU variant instead. So I changed my mind, cancelled the order and instead went for that version, leaving everything else the same. During ordering, I noticed that I had to choose the M1 Max instead of the M1 Pro (hadn't really thought of that) and since the price difference was less then I expected, decided to fork out the additional 200 Euro.

Now all the specs I have read since then say that the M1 Max has double the memory bandwidth than the M1 Pro. But is that actually true if you order an M1 Pro vs an M1 Max both with 32GB, since they may both use 2 x 16GB memory modules? If not, is all this extra silicon on the M1 Max just there to provide 8 GPU cores and nothing else, thereby potentially drawing more power without much everyday gain?

I probably won't change my mind again and am replacing an 2013 15" MacBook Pro with dedicated GPU and hoping to use it for the next 7/8 years or so, so it is more out of curiosity then anything else.

 

[Pressure]

The M1 Max uses 4 chips for a 512-bit memory bus, whereas the M1 Pro uses 2 chips for a 256-bit memory bus.

 

[chabig]

Now all the specs I have read since then say that the M1 Max has double the memory bandwidth than the M1 Pro. But is that actually true if you order an M1 Pro vs an M1 Max

Yes. It's true. Trust the specs.

 

[t2jd1967]

It is Apple's way of writing the specs where my doubts actually began.

M1 Pro Scary fast.​

Up to 10-core CPU
Up to 16-core GPU
Up to 32GB of unified memory
Up to 200GB/s memory bandwidth

M1 Max Scary faster.​

10-core CPU
Up to 32-core GPU
Up to 64GB of unified memory
Up to 400GB/s memory bandwidth

The specs are still correct if the M1 Max only delivers the bandwidth at the maximum memory configuration.

 

[t2jd1967]

This is from the AnandTech website, but it hardly makes it much clearer.

The packaging for the M1 Max changes slightly in that it’s bigger – the most obvious change is the increase of DRAM chips from 2 to 4, which also corresponds to the increase in memory interface width from 256-bit to 512-bit. Apple is advertising a massive 400GB/s of bandwidth, which if it’s LPDDR5-6400, would possibly be more exact at 409.6GB/s. This kind of bandwidth is unheard of in an SoC, but quite the norm in very high-end GPUs.

Apple Announces M1 Pro & M1 Max: Giant New Arm SoCs with All-Out Performance

www.anandtech.com

I guess I will have to wait until someone does a teardown of this exact configuration before we know the answer, although logically it would make a lot of sense to use 4 x 8GB modules if only to fuel the 32GPU cores on 32GB configuration.

 

[Basic75]

Any news on this front? Does the M1 Max always have the full four memory channels with 400GBps? "Always" is regardless of 24 or 32 core GPU, and regardless of 32 or 64GB RAM. It's possible if they have 8 and 16GB RAM chips, M1 has a single channel with either, M1 Pro has dual channel, M1 Max has quad channel. It would be stupid to use 2x16GB for the 32GB M1 Max. And now that I think about it, the M1 Max 24 core GPU surely still has all 4 memory channels for otherwise it couldn't arrive at 64GB.

 

[nquinn]

The anandtech article shows that the cpu's alone can only hit around 220-240 GB/s on the MAX chip, so the 200GB/s Pro chip isn't held back much at all.

Those are also only in the most demanding workloads as well with majority of cores firing.

In the end for non-gaming and non-ProRes type workflows the MAX 24/32gpu chip probably will be like 98% similar to the pro. Benchmarks will show us results soon.

 

[t2jd1967]

After some consideration, I have decided to return the M1 Max with 10/24/32/1TB (still in original packaging) and order 10/32/64/2TB instead. Since I will be using this laptop again for 8 years (it is replacing a maxed out 2013 15" rMBP), I didn't want to risk running out of memory or disk and thought I'd chuck in the extra 8 GPU cores in for good measure.

Still interested to see what any tests on the 32GB M1 Max configuration might turn up.

 

[nquinn]

After some consideration, I have decided to return the M1 Max with 10/24/32/1TB (still in original packaging) and order 10/32/64/2TB instead. Since I will be using this laptop again for 8 years (it is replacing a maxed out 2013 15" rMBP), I didn't want to risk running out of memory or disk and thought I'd chuck in the extra 8 GPU cores in for good measure.

Still interested to see what any tests on the 32GB M1 Max configuration might turn up.

Ugh this is probably what I should do, but decided that for +$400 to move to 2tb, I'll just save that money and upgrade my next MBP purchase a little sooner (maybe 5yrs instead of 7yrs).

Another way to think about it: if you only mostly need storage at home, you can get a used Synology DS220+ nas and 2 x 4tb (8tb total) of storage also for... $400. 8x cheaper. Obviously not apples to apples, but if you just need cheap storage for static media like photos and video then why burn all that money on ultra fast SSD storage.

 

[Basic75]

Ugh this is probably what I should do, but decided that for +$400 to move to 2tb, I'll just save that money and upgrade my next MBP purchase a little sooner (maybe 5yrs instead of 7yrs).

Another way to think about it: if you only mostly need storage at home, you can get a used Synology DS220+ nas and 2 x 4tb (8tb total) of storage also for... $400. 8x cheaper. Obviously not apples to apples, but if you just need cheap storage for static media like photos and video then why burn all that money on ultra fast SSD storage.

Yeah, storage is the only thing that you can reasonably add externally, RAM and GPU not so much.

 

[gt853]

Apologies if this is not the right thread for this, however I didn't want to make a new one - does anyone know the clock speed/CL (cas latency) for the LPDDR5-6400 in these machines? I searched online everywhere for answers but couldn't find anything... I understand they said 400GB/s throughput but I wanted to know the actual memory clock speed. Perhaps you can calculate a theoretical number based on the number of lanes and memory bandwidth?

 

[Kpjoslee]

Apologies if this is not the right thread for this, however I didn't want to make a new one - does anyone know the clock speed/CL (cas latency) for the LPDDR5-6400 in these machines? I searched online everywhere for answers but couldn't find anything... I understand they said 400GB/s throughput but I wanted to know the actual memory clock speed. Perhaps you can calculate a theoretical number based on the number of lanes and memory bandwidth?

I don't think they ever release the CAS latency for LPDDR5 but wouldn't be much different from DDR5 I assume.
Memory clock speed is 3200Mhz. Bandwidth comes from 3200 (memory frequency) x 2 (double data rate) x 512 (memory width) / 8 (bits to bytes) = 409,600 MB/s = so 409.6GB/s

 

[Pressure]

Apologies if this is not the right thread for this, however I didn't want to make a new one - does anyone know the clock speed/CL (cas latency) for the LPDDR5-6400 in these machines? I searched online everywhere for answers but couldn't find anything... I understand they said 400GB/s throughput but I wanted to know the actual memory clock speed. Perhaps you can calculate a theoretical number based on the number of lanes and memory bandwidth?

Latency have regressed somewhat from the M1 but that is alleviated by the much larger 48MB System Level Cache.

These numbers are from AnandTech, who is pretty much the only source to not just run Geekbench.

 

[Macintosh IIcx]

Apologies if this is not the right thread for this, however I didn't want to make a new one - does anyone know the clock speed/CL (cas latency) for the LPDDR5-6400 in these machines? I searched online everywhere for answers but couldn't find anything... I understand they said 400GB/s throughput but I wanted to know the actual memory clock speed. Perhaps you can calculate a theoretical number based on the number of lanes and memory bandwidth?

I think that Fixit have done their first tear down and they normally mention the exact part number of the RAW modules, so maybe you can google that and figure it out. I would actually be interested to see whether there is any difference between the 32GB and 64GB in the CL numbers, could be a bit worse on the 64GB maybe.

 

[Chancha]

The clock has to be 3200MHz as indicated by various benchmarking utilities.
As for the actual listed CL of the chips it does take a teardown to be kind of sure.

 

[nquinn]

After some consideration, I have decided to return the M1 Max with 10/24/32/1TB (still in original packaging) and order 10/32/64/2TB instead. Since I will be using this laptop again for 8 years (it is replacing a maxed out 2013 15" rMBP), I didn't want to risk running out of memory or disk and thought I'd chuck in the extra 8 GPU cores in for good measure.

Still interested to see what any tests on the 32GB M1 Max configuration might turn up.

IMO not worth it at least on the RAM. Apple charges such a huge markup on RAM (+$800 for 64gb or +$400 for 32gb) that you can simply save that money and upgrade a bit earlier. It's very possible in 4-5 years that the base RAM could become 32gb or "free", not to mention all the other upgrades you get. 2tb storage could make sense if you need the space.

These things are always tricky but there's a real balance of - it only costs X extra per year over Y years vs. how much more depreciation do I take on custom BTO features that don't resell nearly as well. Unless you truly need it I'd probably stay with 32gb as a sweet spot for more pro users. SSD storage you have to decide yourself.

 

[nquinn]

Latency have regressed somewhat from the M1 but that is alleviated by the much larger 48MB System Level Cache.

These numbers are from AnandTech, who is pretty much the only source to not just run Geekbench.

I wish there were more reviews around the 80% typical world usage. Like show me web browsing/javascript performance, macOS performance on launching apps, MS Office, etc.

I feel like Apple and every reviewer who is a content producer thinks we are all video editors when that probably makes up like 2% of mbp users.

 

[Basic75]

I wish there were more reviews around the 80% typical world usage. Like show me web browsing/javascript performance, macOS performance on launching apps, MS Office, etc.

I feel like Apple and every reviewer who is a content producer thinks we are all video editors when that probably makes up like 2% of mbp users.

An M1 works just as well for such non-heavily-multithreaded light-weight tasks...

 

[nquinn]

An M1 works just as well for such non-heavily-multithreaded light-weight tasks...

Pro has plenty of advantages that make it worth the money:
- Performance when needed 10-20% of the time for xcode compilation, etc
- Better screen, no big bezels
- Faster SSD
- New keyboard with no stupid touchbar
- Improved speakers

 

[Pressure]

I wish there were more reviews around the 80% typical world usage. Like show me web browsing/javascript performance, macOS performance on launching apps, MS Office, etc.

I feel like Apple and every reviewer who is a content producer thinks we are all video editors when that probably makes up like 2% of mbp users.

Typical computer users don't need the M1 Pro or Max.

The M1 is more than fast enough for those workloads you mentioned and so are computers sold the last 10 years.

With that said I believe you will find M1, M1 Pro and M1 Max to typically be the fastest in javascript and web browsing experience as Apple have prioritised that in their processor design.

Just find some tests of WebPXRT, Speedometer, JetStream, Octane, Kraken and SunSpider.

 

[gt853]

I don't think they ever release the CAS latency for LPDDR5 but wouldn't be much different from DDR5 I assume.
Memory clock speed is 3200Mhz. Bandwidth comes from 3200 (memory frequency) x 2 (double data rate) x 512 (memory width) / 8 (bits to bytes) = 409,600 MB/s = so 409.6GB/s

Thank you, that makes sense. Do you think 3200Mhz is the base clock or max clock under load?

Latency have regressed somewhat from the M1 but that is alleviated by the much larger 48MB System Level Cache.

These numbers are from AnandTech, who is pretty much the only source to not just run Geekbench.

...

That chart is exactly what I was looking for, thank you. I'm not too knowledgable on the inner workings of RAM (I'm not sure if you would even call it RAM), what do you mean by saying it is "alleviated" by the much larger 48MB System Level Cache? I was able to find the article where that image comes from, and one of the things he says (in reference to the 48MB System Cache) is : "While being much larger, it’s also evidently slower than the M1 SLC – the exact figures here depend on access pattern, but even the linear chain access shows that data has to travel a longer distance than the M1 and corresponding A-chips". I'm confused now, what effect does the System Cache have on memory latency?

 

[leman]

I don't think they ever release the CAS latency for LPDDR5 but wouldn't be much different from DDR5 I assume.
Memory clock speed is 3200Mhz. Bandwidth comes from 3200 (memory frequency) x 2 (double data rate) x 512 (memory width) / 8 (bits to bytes) = 409,600 MB/s = so 409.6GB/s

LPDDR5 memory clock is 400Mhz, 3200Mhz is the memory bus clock

Thank you, that makes sense. Do you think 3200Mhz is the base clock or max clock under load?

That’s the base clock which is also the max clock. RAM does not support opportunistic overclouding like modern CPUs do. Although, if I understand it correctly, RAM can be throttled to reduce power usage in scenarios where you don’t need full performance.

 

[Pressure]

Thank you, that makes sense. Do you think 3200Mhz is the base clock or max clock under load?



That chart is exactly what I was looking for, thank you. I'm not too knowledgable on the inner workings of RAM (I'm not sure if you would even call it RAM), what do you mean by saying it is "alleviated" by the much larger 48MB System Level Cache? I was able to find the article where that image comes from, and one of the things he says (in reference to the 48MB System Cache) is : "While being much larger, it’s also evidently slower than the M1 SLC – the exact figures here depend on access pattern, but even the linear chain access shows that data has to travel a longer distance than the M1 and corresponding A-chips". I'm confused now, what effect does the System Cache have on memory latency?

Historically accessing RAM was really slow and caching were invented to store the most needed data ahead of time. That way you could try and prefetch what data was needed next by the processor and speeding up the execution time significantly.

Increasing these caches give you a better "hit" rate as you can store a broader range of prefetched data and if successful spare an "expensive" trip back and forth from main memory to locate the data.

 

[Kpjoslee]

LPDDR5 memory clock is 400Mhz, 3200Mhz is the memory bus clock

Yeah, I know. Just far easier to go by IO clock when calculating bandwidth. You also gotta know the prepatch size with Memory clock while you can just simply double the IO clock when calculating date transfer rate on DDR ram. I think when they refer to memory frequency, they just go by the IO bus clock frequency pretty much all the time.

 

[Basic75]

Pro has plenty of advantages that make it worth the money:
- Performance when needed 10-20% of the time for xcode compilation, etc
- Better screen, no big bezels
- Faster SSD
- New keyboard with no stupid touchbar
- Improved speakers

In case you hadn't noticed, what I was responding to was mostly about benchmarks for web browsing and office applications.