Anandtech: M1 Max multithread performance is like Ryzen 5950x

[Adarna]

That would appear to be a surprise to me since the whole point of the leaving of Apple's Chief Architect Gerald William III in 2019, is to form Nuvia where they can actually build Server chips, which Apple allegedly had no plans for.

Key iOS Chip Architect Gerard Williams III Departs Applehttps://www.macrumors.com › 2019/03/30 › gerard-willia...

Now Nuvia is acquired by Qualcomm who failed to commercialize their ARM server business a decade ago... Oops.

Business decisions could change based on available resources.

At the time of his departure my guess is Apple was focusing on converting ~80% of all Macs to Apple Silicon. Then when they got done with that they had to tackle the volume issue of the ~4% of Macs that are the Pro Desktops.

 

[pshufd]

Yup, that's roughly the idea. Put multiple M1 Max SoC on a logic board to increase CPU & GPU core counts and unified memory.

During the PowerPC days multiple CPUs were placed on the same logic board for a primitive dual core processor.

This would be the only economical way to do it.

Unless of course Apple would instead put the M1 Max die destined for single SoC package and then put the multiple M1 Max dies into a multi SoC package instead?

In the early Mac Pro days, there were two separate CPU boards with Intel Xeons and six DIMM slots per PCB. So Apple already has experience with multi-SoC designs.

 

[pshufd]

You are certainly right on this regard. M1 Max's size is already really extreme on N5P; an 16+64 SOC as predicted for the iMac will easily exceed the reticle limit for the manufacturing process and simply be unfeasible.

However, there is currently no clear evidence regarding the existence of high-speed intra-chip connections on the M1 Pro/Max, which is crucial for a multi-processor heterogeneous computer designed The Apple Way. Simply put, it looks like M1 Max can't be easily packaged by two or four to form a bigger chip just yet.

So my prediction would be some large package with CPU&GPU chiplets the AMD Ryzen way. But this makes designing the memory subsystem really difficult, itself being a big selling point of the M1 family architecture. 4 cpu chipsets with 10 cores each and 4 GPU chips with 32 cores each to form a 40+128 config, with memory/IO crammed into another chiplet with 2048 bit LPDDR5??? This would be a nightmare to design with all the cachelines and packaging technologies.

Ahh, sure, there is the Magical, twice-delayed Intel Ponte Vecchio...

https://www.tomshardware.com/features/intel-ponte-vecchio-and-xe-hpc-architecture-built-for-big-data

But is it just me who thinks M1's design topology is too exotic and expensive to scale beyond M1 Max? Like, craaaazy cool for a super low volume halo Mac product but way too expensive to not go into a multi-billion dollar super computer cluster...

The only evidence that I've seen is from the Anand article in that the eight performance cores are organized as two clusters of four cores.

 

[falainber]

I expect 2x and 4x M1 MAX on their big desktops.

M1 Max die is huge. I am not sure they can make it any larger. They can increase TDP with better cooling but iMacs are not known for good cooling either.

 

[pshufd]

M1 Max die is huge. I am not sure they can make it any larger. They can increase TDP with better cooling but iMacs are not known for good cooling either.

I think that it would be a multi-SoC solution.

The current iMac 27 has to handle an i9-10900 which has a PL2 of probably 250 watts and a GPU (Radeon Pro 5700 XT) with a TDP of 130 watts. So the current model can power up to 380 watts for the CPU and GPU. Sounds like it could easily handle 2xM1 MAX.

 

[Pro Apple Silicon]

So the odds of a iMac 27" or larger screen replacement with 500W PSU is very unlikely unless these are multi SoC M1 Max configurations to allow for more than 64GB memory like say 128GB, 256GB, 512GB, 1TB or 2TB using multiple M1 Max SoCs

Multiple M1 Max SoCs is almost certainly the plan for the Mac Pro. Have you looked at the rumored specs? They are literal multiples of the M1 Max.

 

[Adarna]

M1 Max die is huge. I am not sure they can make it any larger. They can increase TDP with better cooling but iMacs are not known for good cooling either.

Check the iMac Pro. The cooling solution inside is for very hot Xeon chips

 

[pshufd]

The iMac Pro with the 28-core Xeon W CPU had a CPU TDP of 255 watts. The PL2 or Turbo Power limit of that CPU is 510 Watts. That's just for the CPU.

 

[pshufd]

The die size for M1 MAX is reported at 435 mm^2 (Anand).

The Intel W-3175X (iMac Pro 28 core Xeon W) is 4,294 mm^2 (Intel). No wonder they charge $3,000 for the CPU. Good old 14 nm. Could you imagine an Apple Silicon SoC with 10 x M1 MAX?

 

[Patchwork]

I think John Siracusa tweeted the Apple silicon roadmap back in May where he correctly predicted the M1 Pro and M1 Max as Jade C-Chop and Jade C-Die respectively. Given the accuracy of his prediction then the his other predicted chips may well end up the Mac Pro. The other chips were Jade 2C-Die and Jade 4C-Die, so effectively two or four M1 Max glued together, hence 40 CPU cores and 128 graphics cores at the maximum configuration. The Jade 4C-Die would comfortably be well below the power usage of the current Mac Pro, which is probably where the rumour of a smaller Mac Pro or Cube like Mac started.

 

[leman]

However, there is currently no clear evidence regarding the existence of high-speed intra-chip connections on the M1 Pro/Max, which is crucial for a multi-processor heterogeneous computer designed The Apple Way. Simply put, it looks like M1 Max can't be easily packaged by two or four to form a bigger chip just yet.

You are certainly correct about this, but it's not like we have any evidence that such connections do not exist either. We know that Appel patented advanced chip interconnects some years ago. We also know that manufacturing a 1000m2 SoC just for a workstation is probably not economically feasible. Finally, we have rumors about core counts of Mac Pro systems.

None of this is conclusive, for sure, but it's a very realistic path forward. Besides, this is where the industry is moving anyway. It is not silly to expect that complex multi-tile packages will be released by multiple vendors at approximately the same time (Apple in 2022, Intel in 2023).

A few months before the M1 there was news of an ARM-based supercomputer running Linux.

Supercomputers have been running on all kinds of exotic hardware for a while. This doesn't really mean much. If you put together thousands of CPUs, you'll obviously get a fast cluster.

Would not be surprised if Apple would transition their data center, cloud, iCloud, App Store, etc to a M1 Max-based server farm.

These operations would benefit the most from the performance per Watt of Apple Silicon.

Way too expensive. I doubt that they can produce that many chips. Not to mention that M1 Max is mostly a large GPU which will be completely wasted on a cloud server. And of course, the biggest issue — software. Are they supposed to run macOS on these?

 

[raknor]

I can fully, 100% understand this benefit as it relates to battery, but this is such a weird thing to care about when you're plugged in.

Context is important. Read the comment I am responding to .. I'll lay it out in simpler terms since you seem to have a hard time grasping things.

The person I responded to claimed the M1 Max based MBP is pulling 120 Ws and is Desktop territory.
I pointed out the MSI laptop is pulling more than double.

So what does that make it?

You seem to have trouble with numbers 1.3 mm 3.1 mm.. 120W 256W.

Like, are you crying for the environment while your numbers are crunching while corded? Do you think "gosh, this is so fast, but if only the power draw from my wall socket was lower", just ... who thinks that? Is it hard on your electricity bill or something? Was it ever?

Yes it was. My iMac Pro drew 200Ws idling. Working from home that is not an insignificant power draw. Only idiots are not concerned about climate change, conserving power or saving money.

Do you care about this in other areas, like televisions and refrigerators? Why or why not?

Yes. See above.

 

[raknor]

Yes it does. These are gaming laptops, nobody buys them to use on the laps. My point is that laptop and desktop power envelopes overlaps and while M1 Max in this case is used in the laptop its power consumption crosses into the desktop territory.

Wrong. 120-140W power draw from the wall is middle of the pack laptop territory.

Their room for straight scaling up the chips by increasing the number of cores/GPUs is rather limited (also because the chip is humorous). BTW, even M1 is used in the desktops.

Wrong again. Desktops is a very vast area. There are intel NUCs based on 15W U series chips.

Making up random terminology is exactly that.. random terminology.

 

[leman]

I can fully, 100% understand this benefit as it relates to battery, but this is such a weird thing to care about when you're plugged in. Like, are you crying for the environment while your numbers are crunching while corded? Do you think "gosh, this is so fast, but if only the power draw from my wall socket was lower", just ... who thinks that? Is it hard on your electricity bill or something? Was it ever?

Based on numbers published by Andrei, using the M1 instead of a desktop with the same performance will save me around 400-500 Swiss franks per year. I’d say it’s pretty much significant.

 

[Pressure]

TSMC have many ways to build a chip larger than the reticle limit.

They have 3D chip stacking with either Chip on Wafer or Wafer on Wafer, then they have advanced packaging like Integrated Fan-out with Redistribution Layer and Local Si Interconnect or Chip on Wafer on Substrate which can also be mixed up with Redistribution Layer.

Last year they built a 1,700mm² solution for Broadcom using two 5nm dies.

With these new 3D stacking technologies you can assemble chips with up to 4 times the reticle limit (3,400mm²).

Remember that these larger chips will go inside machines starting at $5,999. The iMac chip can easily be made under the reticle limit. Or they could just slap the M1 Pro and Max in them for now.

 

[raknor]

Andrei of Anandtech is experiencing unusual screen issues

https://twitter.com/i/web/status/1452622025550438403

That's not unusual for LCD. That's a slow motion video capture.. you are not going to see what the video shows with your eyes.

 

[leman]

Yes it does. These are gaming laptops, nobody buys them to use on the laps. My point is that laptop and desktop power envelopes overlaps and while M1 Max in this case is used in the laptop its power consumption crosses into the desktop territory. Their room for straight scaling up the chips by increasing the number of cores/GPUs is rather limited (also because the chip is humorous). BTW, even M1 is used in the desktops.

The 140W charger is for fast charging. These laptops are not consuming 140W. They use approx 50W for CPU-intense work and approx 80W for GPU-intense work. Way less than a desktop (where you’d see power draws of over 250W for similar performance).

 

[EntropyQ3]

The 140W charger is for fast charging. These laptops are not consuming 140W. They use approx 50W for CPU-intense work and approx 80W for GPU-intense work. Way less than a desktop (where you’d see power draws of over 250W for similar performance).

Ryan Smith at Anandtech stated that throughout their testing, they had never seen the GPU-block exceed 45W. (The whole system obviously draws a bit more, but for comparison with dGPUs, I think 45W peak is the relevant number.)

 

[leman]

Ryan Smith at Anandtech stated that throughout their testing, they had never seen the GPU-block exceed 45W. (The whole system obviously draws a bit more, but for comparison with dGPUs, I think 45W peak is the relevant number.)

I just added the CPU to it No harm in being generous. I rather overestimate the power consumption than underestimate it, or the bean counters will haunt me to the end of my days.

 

[Serban55]

Gruman said that the larger iMacs will come with 14/16c+64c options and for the Mac Pro, 32/40c+128c.

where he said that about the imacs?

 

[Serban55]

The die size for M1 MAX is reported at 435 mm^2 (Anand).

The Intel W-3175X (iMac Pro 28 core Xeon W) is 4,294 mm^2 (Intel). No wonder they charge $3,000 for the CPU. Good old 14 nm. Could you imagine an Apple Silicon SoC with 10 x M1 MAX?

wait...the Intel W 3175X is 4 metres ^2 ?? ))) jesus hahahahha the whole vertical imac is not that long =))
i think you meant

Dimension

76.16 mm × 56.6 mm

 

[falainber]

Wrong. 120-140W power draw from the wall is middle of the pack laptop territory.


Wrong again. Desktops is a very vast area. There are intel NUCs based on 15W U series chips.

Making up random terminology is exactly that.. random terminology.

I am not trying to establish new terminology. You are. There is a long list of desktop processors that consume less power than 90W (M1 Max). Hence my point about M1 Max crossing into desktop territory. I am not sure what exactly you are trying to deny. How would you say that M1 Max cosumes more power than a lot of desktop processors?

 

[PeterJP]

wait...the Intel W 3175X is 4 metres ^2 ?? ))) jesus hahahahha the whole vertical imac is not that long =))

1 mm^2 = 1mm x 1mm = 1 millionth of a m^2. So to get to 4000 mm^2, you have about 65mm x 65mm.

 

[Serban55]

1 mm^2 = 1mm x 1mm = 1 millionth of a m^2. So to get to 4000 mm^2, you have about 65mm x 65mm.

Agree, he was saying like the Intel chip is 10 times bigger than the m1 max

 

[PeterJP]

Agree, he was saying like the Intel chip is 10 times bigger than the m1 max

Yes! It's a huge chip. But not 4 meters squared