I'm retired from being a Defense Contractor and been to many places and seen them with mine own eyes walking back to my section for a new Unix crypto device worked! So don't assume you know everything!
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple in the server market again? Moore's law is dead confirms that Apple is making its server chips
- Thread starter Xiao_Xi
- Start date
- Sort by reaction score
If Apple is doing this, it is so they can run their services on their own systems, and do it more efficiently and at less cost than using outside service providers.
But I dont think they can replace Amazon. Amazon's server scale is beyond Apple's reach.
The M1 die ? No.
Even the M1 Max. No.
If talking pull just the P cores out and build an entirely new un-core portion of the die "top to bottom" is it really based on the the M1 anymore? No; that isn't all that accurate connotation. If chop off 80% of the die and replace it with something else, then pretty much have something effectively new. It is myopic perversion to view the M1 die as primarily just being the P cores and the rest is "window dressing".
The soldered on RAM die thing won't scale in capacity to match Graviton 2. The I/O doesn't scale. The core count doesn't scale. The internal bus network probably doesn't scale to 64+ P cores.
Could it beat Graviton 2 on some single threaded drag race (Geekbench ST ) . Probably But is that the point in a high multiple tenant workload load set up ? No.
Their Mac Pro with a custom chip is still going to be substantively skewed to providing best single user performance as a single user workstation.
The workloads for a high end server in a cloud environment where trying to service workloads of multiple clients concurrently is substantively different.
TLB+CLR thrash. Graviton2 gets to about 100-108 ns latency at over 16MB Likewise full random is fine until past 16MB. ( the log scale graph)
The M1 Max with next generation memory subsystem, next generation process node ( N5 instead of N7) and twice the memory channels is in the same zone for full random above 16MB.
Neoverse N2 will be better. Even Neoverse N1 could be done better than what Amazon did.
Ampere 80
The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster
Latencies here under 100ns even out to over 525MB. Graviton2 isn't what Apple's has to beat to be competitive.
Amazon probably has the least expensive server CPU costs out there; not the best.
Single thread drag racing on relatively small test depths above ... yes the M1 will be substantively better. It is tuned that way. That isn't a mainstream cloud service heavy workload though.
Anything Apple makes for the server will almost certainly be Apple-user focused first. Apple isn't going to compete with AWS to host your Postgres/web services. They're going to offer Apple Silicon Cloud to Apple users to accelerate their workflows first.
Think about businesses that used to have racks of Macs in a closet. Those will hugely benefit from an Apple Silicon Cloud.
Then think about the worker who can carry around a Macbook Air with 20 hours of battery life and be able to tap into a 40/128 chip with a single swipe on the touchpad.
Apple can also offer application developers access to integrate cloud acceleration directly into their apps. For example, the application is installed locally on your computer but for certain functions on the app, you can have it use Apple Silicon in the cloud.
Apple Silicon isn't going to be competing with Graviton2 or Neoverse. Those are designed to be low-cost, low ST performance chips used in microservices and web servers where handling the number of tiny requests is more important than raw performance.
Think about businesses that used to have racks of Macs in a closet. Those will hugely benefit from an Apple Silicon Cloud.
Then think about the worker who can carry around a Macbook Air with 20 hours of battery life and be able to tap into a 40/128 chip with a single swipe on the touchpad.
Apple can also offer application developers access to integrate cloud acceleration directly into their apps. For example, the application is installed locally on your computer but for certain functions on the app, you can have it use Apple Silicon in the cloud.
Apple Silicon isn't going to be competing with Graviton2 or Neoverse. Those are designed to be low-cost, low ST performance chips used in microservices and web servers where handling the number of tiny requests is more important than raw performance.
Last edited:
AWS's Graviton2 uses Neoverse N1 CPU microarchitecture.
What do you mean by that? Does it mean that another company could have designed a better SOC than AWS?
How could server chips for Apple internal use be? Could Apple be designing chips for running Apache Cassandra clusters or the App Store?
Coming in late to this thread, but wanted to contribute.MLID is reliable? Don't think so.
Is Youtuber "Moore's Law Is Dead" legit?
Randomly found this guys channel a few days ago. He claims to know a bunch of industry insiders and supposedly has a ton of info that he "can't talk about". But he only has 70K subs so I'm a little suspicious. Does anyone know if this guy is for real?linustechtips.com
He's just a clickbait youtube guy who spits out lots of guesses (often contradicting himself), hypes the hits, and pretends the misses never happened.
“Moore’s Law Is Dead“ isn’t clickbait. The host often misses specific contexts because his focus is largely on gaming.
So let me break this down the best way I can. For iCloud to go 100% internal would require Apple to either rent space in existing data centers to support their users at the edge in a reliable way, which is something they 100% could do. The logistics of that however would cost $100’s of millions if not billions to do so.
Apple largely tries to outsource others do better to those companies. I have 0 doubt Apple will continue to build iCloud on AWS and GCP. Why? Why build data centers when Amazon and Google provide many of the building blocks they need on a consumption basis in a way that scales globally.
I am BEYOND happy to be wrong on this. Truly, I am. I’d LOVE for Apple to build a server product.
Do you know what I see when I look at the various configurations of the Mac Pro though? I see an IO intensive single socket server, that could easily be slide into a rack. What does the Mac Pro have? An option to get it rack mounted.
I enjoy the content from “Moore’s Law Is Dead”, for many reasons. It’s important to remember that their target audience is gamers.
If you show a Gaming enthusiast a rig like the Mac Pro that has a rack mounted option, they’ll probably assume it’s a Server.
To that point, ”Moore’s Law is Dead” host today posted a video covering AMD’s new EPYC chip set to go into mass production video. He says specifically that he doesn’t focus on servers, but he’s covering it because because of the ramifications long term to gaming.
He does have intelligent points, but he does often misunderstand what’s going on outside of gaming and tries to bring it down to his frame of reference.
Is he the most brilliant commentary fellow? No, as someone thats watched him for a couple of months I can tell you that his perspective is narrow. At least in my opinion he says that pretty clearly.
If you show someone who’s focus is on gaming, a machine that can be rack mounted (like the Mac Pro) a system that has 4 M1 Maxs in it, that can be rack mounted, they’ll probably assume it’s a Server. Why? Because Fortnight probably can’t use something that in many cases is faster than the NVIDIA stuff coming out the pipe, let alone 4x that.
Apple is more likely to work with Cloud providers to get open things from their server stack to work well in a Kubernetes and Serverless environment than spend a ton of money building everything themselves to just not be on AWS and GCP.
If Apple wanted to literally own everything for service transactions, they would have built their own fab, and factories for manufacturing a long time ago. TSMC, and Foxconn are both happy to take the money in high volumes and agree to Apples terms. There is little to no reason to think that outside of having one primary massive scale data center (which I believe they did when the sapphire company they put $400 million into failed to produce and turned their facilities into a massive scale data center) that Apple would spend that same kind of money globally when they can do what they do with TSMC and Foxconn and rent it out on a per millisecond basis.
I largely agree, Cloud computing assumes flexibility. I don’t know how macOS is doing it’s virtualization, but I will tell you that as far as a workstation goes, it’s the best solution I’ve used having used workstation laptops and Macs Out of the many years I’ve experimented with it.
It “could” be done. I have little doubt about that. I don’t think it makes sense for them though.
I have no doubt the hypervisor framework grown out in scale could do what AWS and GCP does. But on return on investment, unless they tell an incredibly competitive story, I just don’t see it happening.
It costs a lot of money to build out an actual competitive Cloud platform. More money than what was probably spent to secure 5nm for a couple of years from TSMC.
Using TSMC's chiplet tech it is possible to mask-stitch 9-16 M1 Max dies together to reach 90-160 CPU cores together.
The 17.004cm² die could fit into the 19.7cm² Mac mini enclosure although it would need at a 1-1.4kW PSU to power
The 17.004cm² die could fit into the 19.7cm² Mac mini enclosure although it would need at a 1-1.4kW PSU to power
Form Factor (As of 9 Nov 2021) | AMD Zen 4 Epyc 128 core cpu Rival | AMD Zen 4 Epyc 128 core cpu Rival | Mac Pro iMac Pro | Mac mini Pro Mac Pro iMac Pro | MBP 14" MBP 16" Mac mini Pro iMac 24" iMac Pro | 300mm² Silicon Wafer |
Apple silicon chip | M1 Max Jade-16C | M1 Max Jade-9C | M1 Max Jade-4C | M1 Max Jade-2C | M1 Max | M1 Max Jade-49C |
Launch | >Speculation< | >Speculation< | Q2 or Q4 2022 | Q2 or Q4 2022 | Q4 2021 | >Speculation< |
# of dies | 16 | 9 | 4 | 2 | 1 | 49 |
CPU | 160 | 90 | 40 | 20 | 10 | 490 |
performance cores | 128 | 72 | 32 | 16 | 8 | 392 |
efficiency cores | 32 | 18 | 8 | 4 | 2 | 98 |
GPU core | 512 | 288 | 128 | 64 | 32 | 1568 |
Neural Engine core | 256 | 144 | 64 | 32 | 16 | 784 |
memory bandwidth | 6.4TB/s | 3.6TB/s | 1.6TB/s | 800GB/s | 400GB/s | 19.6TB/s |
Max Memory | 1024GB | 576GB | 256GB | 128GB | 64GB | 3,136GB |
Hardware-accelerated H.264, HEVC, ProRes, and ProRes RAW | 16 | 9 | 4 | 2 | 1 | 49 |
Video decode engines | 16 | 9 | 4 | 2 | 1 | 49 |
Video encode engines | 32 | 18 | 8 | 4 | 2 | 98 |
ProRes encode and decode engines | 32 | 18 | 8 | 4 | 2 | 98 |
Peak Quoted Transistor Densities using TMSC's 5nm (2020) at the same die size 171.3 million transistors per mm² | 912 Billion | 513 Billion | 228 Billion | 114 Billion | 57 Billion | 5.19 Trillion |
Estimated Die Size | 17.004cm² | 12.753cm² | 8.502cm² | 6.3765cm² | 4.251cm² | 30cm² |
Peak Quoted Transistor Densities using IBM's 2nm (2025) at the same die size 333.33 million transistors per mm² | 1.774 Trillion | 998.24 Billion | 443.66 Billion | 221.83 Billion | 110.92 Billion | 10.1 Trillion |
Estimated AMD Ryzen 9 5950X Performance | 16x | 9x | 4x | 2x | 1x | 49x |
Estimated RTX 3080 Performance | 16x | 9x | 4x | 2x | 1x | 49x |
Their Mac Pro with a custom chip is still going to be substantively skewed to providing best single user performance as a single user workstation.
The workloads for a high end server in a cloud environment where trying to service workloads of multiple clients concurrently is substantively different.
TLB+CLR thrash. Graviton2 gets to about 100-108 ns latency at over 16MB Likewise full random is fine until past 16MB. ( the log scale graph)
The M1 Max with next generation memory subsystem, next generation process node ( N5 instead of N7) and twice the memory channels is in the same zone for full random above 16MB.
Neoverse N2 will be better. Even Neoverse N1 could be done better than what Amazon did.
Ampere 80
The Ampere Altra Review: 2x 80 Cores Arm Server Performance Monster
www.anandtech.com
Latencies here under 100ns even out to over 525MB. Graviton2 isn't what Apple's has to beat to be competitive.
Amazon probably has the least expensive server CPU costs out there; not the best.
Single thread drag racing on relatively small test depths above ... yes the M1 will be substantively better. It is tuned that way. That isn't a mainstream cloud service heavy workload though.
Latency to RAM depends on the RAM... and obviously regular DDR has lower latency than LPDDR. But why is it even that interesting? If you are hitting cache so rarely that RAM latency starts to matter your CPU performance is already down the drain anyway...
Make that Mac mini taller, say 9.8", and you would have room for that PSU & a heat sink filling the remaining interior volume; something about the size of the G4 Cube...! ;^p
For 16 maxed-out M1 Max SoCs that would be a bargain...!
16-way M1 Max MCM
160-core CPU (128P/32E)
512-core GPU
256-core Neural Engine
1TB LPDDR5 RAM
6.4TB/s memory bandwidth
the Ampere 80 ( and 128) are also based on N1. And Ampere (another company) does a better job with an internal interconnect that scales better . That is what the rest of the post covered.
[ Some folks might read Ampere and think Nvidia's codename for their current GPU microarch family. There is a company with that name.
https://amperecomputing.com/altra/
Unlike Amazon/AWSl, they are quite willing to sell to individual cloud services vendors. Got a major stack of cash... you too can buy some. No impediment here to Apple just buying something. Or commissioning a semi-custom future version that they slap their logo on for internal use. ]
@deconstruct60 Thank you!
AMD has two types of GPU: CDNA for scientific computing and RDNA for gaming/rendering.
Which type of GPU should Apple's server chips have: CDNA-like GPU or RDNA-like GPU?
AMD has two types of GPU: CDNA for scientific computing and RDNA for gaming/rendering.
Which type of GPU should Apple's server chips have: CDNA-like GPU or RDNA-like GPU?
One would think that answer would depend on what Apple plans on doing with these hypothetical servers...?
Crunching data, probably CDNA-like; on-demand streaming gaming, gotta be RDNA-like...?
Well, I'm not retired. An M1 SOC Mini is not a competitive supercomputer node. Sorry. Too few cores, inadequate memory, limited to ethernet.... I don't think HPE is worried. Vastly different market. I thought you were joking.
If I'm parsing "mine own eyes walking back to my section for a new Unix crypto device worked!" correctly, that's not a supercomputer.
Last edited: