ARM & Mac Pro

[Takuro]

Exactly, that's why I'm not sure not having access to CUDA would impact the workflow. Are there equivalent solutions for AMD, or Python/C++ libraries that leverage GPU?

OpenCL. You can use Metal too, but I have no idea how mature it is.

OpenCL for macOS - Apple Developer

If you’re using OpenCL, which was deprecated in macOS 10.14, we recommend that you transition to Metal and Metal Performance Shaders for access to a wider range of capabilities.

developer.apple.com

As somebody who worked on a project that involved both Cuda and OpenCL, Cuda is a lot better though. I know it's not supported by Macs anymore, but it was a lot less of a headache. In a quick Google search, it seems like the gap has narrowed in the last 10 years since I touched it, though, and there are a lot more prewritten examples of how to implement common mathematical formulas that tend to get ran on big data sets, like Fourier Transforms and junk.

Sidenote: I do not miss college comp sci.

 

[leman]

Exactly, that's why I'm not sure not having access to CUDA would impact the workflow. Are there equivalent solutions for AMD, or Python/C++ libraries that leverage GPU?

If you work with machine learning, maybe look at CoreML. It is Apple's set of libraries for GPU-accelerated machine learning. If I remember correctly, there is a converter from TensorFlow to CoreML. Note that CoreML is still not as feature rich then other frameworks.

OpenCL. You can use Metal too, but I have no idea how mature it is.

There is not much point in using OpenCL, it is as good as dead. Metal is an excellent and very easy to use API, but I don't know how useful it is for someone who wants to use existing ML tools directly.

As somebody who worked on a project that involved both Cuda and OpenCL, Cuda is a lot better though.

Well, of course it is. Nvidia effectively killed OpenCL by pushing its proprietary CUDA API instead.

 

[Woochoo]

but like the others said, ARM is not quite there yet for such huge amounts of RAM + cores. Not sure on IO support (latest PCIe) also one has to take in consideration that the software won't be up for it for a couple of years.

I'd suggest you to do some research instead of repeating what others said, that just propagates misinformation and is misleading to other people.

ARM is there already for huge amounts of RAM and cores, not only that, there are ARMs out there with 80 cores already. Amazon implemented their own alternative server chips that performs on par with the other 2 options they offer (Intel's Xeon and AMD's Epyc).
2 weeks ago an ARM supercomputer took the crown as the 2020's most powerful supercomuter, and trust me those manage way more RAM, cores and lanes than anyone here will ever have in their workstation in at least 30-40 years.

So how come "ARM is not quite there yet"? We all should be better doing a bit of research before doing such bold affirmations instead of repeating what others said.

 

[johngwheeler]

I'd suggest you to do some research instead of repeating what others said, that just propagates misinformation and is misleading to other people.

ARM is there already for huge amounts of RAM and cores, not only that, there are ARMs out there with 80 cores already. Amazon implemented their own alternative server chips that performs on par with the other 2 options they offer (Intel's Xeon and AMD's Epyc).
2 weeks ago an ARM supercomputer took the crown as the 2020's most powerful supercomuter, and trust me those manage way more RAM, cores and lanes than anyone here will ever have in their workstation in at least 30-40 years.

So how come "ARM is not quite there yet"? We all should be better doing a bit of research before doing such bold affirmations instead of repeating what others said.

Yep, ARM is definitely there in the cores & RAM stakes, at least for servers. I posted earlier about Ampere Altra 80-core with support for 4TB RAM, and there are 128-core CPU designs already out there (none deployed commercially yet AFAIK).

Where ARM is a bit lacking still is in single-core performance compared to the best Intel & AMD, but it's not too far off - there are 3.5GHz ARM cores, even though most run at 2.5-3GHz for power reasons.

Some specific workloads (e.g. some encryption standards) are better supported on Intel because there is hardware support. This will probably come on ARM over time.

Software / compiler optimization will be a bigger issue. Developers will need to tune their apps to a new platform which could take many years.

 

[tdar]

Please remember that Apple Silicone is not ARM. It's ARM+. It has an Arm core but with Apple extensions. Like the secure Enclave and The ML engine. Apple will bring all of this to the new Mac Pro. It's not going to be as far away as people think. I expect the entire transition to be done by iPhone Reveal day in September 2021. Remember when we transitioned from Power PC to Intel they said it would take 2 years just like now. It was done in 9 months.
This new Mac Pro is going to be a beast of a machine. NO other commercial workstation will be able to touch it. Not even close.

 

[Macbookprodude]

Please remember that Apple Silicone is not ARM. It's ARM+. It has an Arm core but with Apple extensions. Like the secure Enclave and The ML engine. Apple will bring all of this to the new Mac Pro. It's not going to be as far away as people think. I expect the entire transition to be done by iPhone Reveal day in September 2021. Remember when we transitioned from Power PC to Intel they said it would take 2 years just like now. It was done in 9 months.
This new Mac Pro is going to be a beast of a machine. NO other commercial workstation will be able to touch it. Not even close.

Yes, but if Apple ends Intel short, then it will be a paper weight in say 5 years from now.

 

[kuliand]

That means all those who spend 50k or more on intel 2019 mac pro - you are stuck with a paper weight.

people keep saying this but the majority of Mac Pro purchases are by enterprises who if spending 20k must feel they need top performance so in 4-5 years they will need to buy a new machine anyway to keep having the best possible performance. If you are working in machine learning etc you will always need to be on the cutting edge.

 

[ZeroTemplates]

I'd suggest you to do some research instead of repeating what others said, that just propagates misinformation and is misleading to other people.

ARM is there already for huge amounts of RAM and cores, not only that, there are ARMs out there with 80 cores already. Amazon implemented their own alternative server chips that performs on par with the other 2 options they offer (Intel's Xeon and AMD's Epyc).
2 weeks ago an ARM supercomputer took the crown as the 2020's most powerful supercomuter, and trust me those manage way more RAM, cores and lanes than anyone here will ever have in their workstation in at least 30-40 years.

So how come "ARM is not quite there yet"? We all should be better doing a bit of research before doing such bold affirmations instead of repeating what others said.

First of all, ARM is based on RISC (Reduced Instruction Set) architecture, x86 is based on CISC (Complex Instruction set ), some instructions aren't optimized for it vs the existing ones (not saying in the future this won't happen), here is where the research needs to be put into (not I can have X "inches/cm" of RAM or cores -- it's not a measuring contest).
I don't mind it having even 1024 cores (that's good for all of us), second the request is for research where you need it to work from day one with support for all the instructions (one needs to squeeze as much as possible out of it), this means software support (this isn't quite there yet, software is part of the package).
With apple moving to own silicon (ARM+ in this case) software will follow, but you can't say when or if they will bring on the market the latest and greatest ARM Mac Pro (apple is a company that implements stuff after it's proven -- most of the time --)

 

[Waragainstsleep]

you can't say when or if they will bring on the market the latest and greatest ARM Mac Pro (apple is a company that implements stuff after it's proven -- most of the time --)

Well done for researching RISC and CISC so you sound more knowledgeable but you're not quite there yet. You claimed ARM didn't have the core count or RAM support of Xeon and someone corrected you because there are ARM chips that do. I don't know why you are talking about measuring contests.
There is [plenty of software that doesn't make use of multiple cores very well but the people who need them have software that does. People using CS, editing video or doing CAD renders have software that will use all sorts fo computing power and research scientists typically write their own code so they will be beyond ecstatic with a stable, friendly OS that can crunch their giant calculations or fold their proteins or whatever they are up to.

Apple have said they will transition the whole line within two years (of the first one) so we can say when they will bring out the ARM Mac Pro. It will be the last one because if they upstage the entire fleet of $50k machines ~1 year after launch, the pro market will never forgive them. ~3 years they probably will if the new one is good enough. Apple likely has prototypes of this Mac Pro running in its labs already.

Apple adopts technology when they think its ready for market. They were essentially first with USB, Firewire, WiFi, Thunderbolt, some very important technologies. Some things they have to rely on their partners to support, others they don't adopt due to issues around maturity or power consumption.

 

[leman]

First of all, ARM is based on RISC (Reduced Instruction Set) architecture, x86 is based on CISC (Complex Instruction set ), some instructions aren't optimized for it vs the existing ones (not saying in the future this won't happen), here is where the research needs to be put into (not I can have X "inches/cm" of RAM or cores -- it's not a measuring contest).

What instructions are not optimized? And honestly, we should really stop talking about RISC vs CICS. This dichotomy lost its meaning long time ago. For CPU implementation it does not make any sense (both x86 and ARM CPUs are superscalar architectures that decode instructions into microcode before executing them). For comparing the ISA, it is pointless.

 

[ZeroTemplates]

What instructions are not optimized? And honestly, we should really stop talking about RISC vs CICS. This dichotomy lost its meaning long time ago. For CPU implementation it does not make any sense (both x86 and ARM CPUs are superscalar architectures that decode instructions into microcode before executing them). For comparing the ISA, it is pointless.

For research it's quite needed as that is a niece topic, depending on what one uses it for. Most likely for me and you (normal users) ARM is much better, less heat and less power with more cores (yey) perfect for on the road creators, in this case yes I agree the comparison doesn't make sense.
For the next couple of years the software will be the differentiating point, but now it's very optimistic (bcz apple).
Hackintosh on raspberry pi (one can dream ? of apple pie)

 

[leman]

For research it's quite needed as that is a niece topic, depending on what one uses it for. Most likely for me and you (normal users) ARM is much better, less heat and less power with more cores (yey) perfect for on the road creators, in this case yes I agree the comparison doesn't make sense.

I'm still not sure what exactly you mean with "not optimized" instructions, ARM64 is a well documented and it generally does not contain any obsolete/inefficient instructions (unlike x86-64). Maybe you are referring to wider SIMD registers in the x86-64 architecture? It is true that current Apple ARM chips are limited to 128bit SIMD, while most Intel CPUs have 256bit units and the newest ones even have 512bit SIMD. There is speculation that Apple will include the more advanced SVE instructions with wider (or more) vector units in their upcoming desktop CPUs.

 

[ZeroTemplates]

I'm still not sure what exactly you mean with "not optimized" instructions, ARM64 is a well documented and it generally does not contain any obsolete/inefficient instructions (unlike x86-64). Maybe you are referring to wider SIMD registers in the x86-64 architecture? It is true that current Apple ARM chips are limited to 128bit SIMD, while most Intel CPUs have 256bit units and the newest ones even have 512bit SIMD. There is speculation that Apple will include the more advanced SVE instructions with wider (or more) vector units in their upcoming desktop CPUs.

Yup, I didn't know that they will add it, thanks for the info!

Do you have some recent data where I could check this further (seems I'm having some old information regarding to some hardware features)

 

[leman]

Yup, I didn't know that they will add it, thanks for the info!

To make to clear: nothing has been officially confirmed. But if Apple means business in the pro space, they have to work on their SIMD.

 

[johngwheeler]

Yes, but if Apple ends Intel short, then it will be a paper weight in say 5 years from now.

You do know that the machine will still run as well in 5 years time as it does today don't you? They don't suddenly refuse to boot the day it no longer supports the latest version of MacOS.

The issue will be software support, both from Apple and more importantly 3rd party vendors. Historically Apple has supported new versions of MacOS for 7-8 years after the launch of any given machine, and provided security upgrades for a further 2-3 years. After the PowerPC to Intel transition, these periods were a lot shorter, but I think we will see a longer support period for Intel, because there are millions of Intel Macs in daily use, and Apple is still releasing new ones.
[automerge]1594076229[/automerge]

people keep saying this but the majority of Mac Pro purchases are by enterprises who if spending 20k must feel they need top performance so in 4-5 years they will need to buy a new machine anyway to keep having the best possible performance. If you are working in machine learning etc you will always need to be on the cutting edge.

As you say refresh periods for hardware for enterprises are typically 3-5 years, but longer for really expensive hardware. For Machine Learning, I would think that most businesses are using cloud services because these are updated much more frequently, and it's a fast moving technology - you don't want to be running 5 year old hardware when you can just rent time on one and get the latest and greatest.

 

[nick9191]

people keep saying this but the majority of Mac Pro purchases are by enterprises who if spending 20k must feel they need top performance so in 4-5 years they will need to buy a new machine anyway to keep having the best possible performance. If you are working in machine learning etc you will always need to be on the cutting edge.

Agreed. No company dropping tens of thousands on a Mac Pro would still be using that machine as a daily driver 7-10 years down the line. Computers will be turned over somewhere between every 1-3 release cycles to ensure they can offer the performance they need and so machines are in warranty etc.

Unless the company massively changes its entire business, how could it go from needing top of the line equipment in 2019 to needing iMac level speed in 2024 (by that I’m making the admittedly gratuitous implication that a 2019 Mac Pro is as fast as a 2024 iMac).

This really will affect prosumers/enthusiasts who are buying something they don’t really need (which is absolutely their right to do so). However since the Mac Pro has doubled in price in the last ten years, this isn’t a sizeable portion of the userbase.

As long as the Intel Mac Pro still runs the latest MacOS for 3 years after it’s discontinued, as was the case with the G5, I’d expect most companies would be happy with this. Third parties may drop software support earlier but that’s not Apple’s fault.

 

[Waragainstsleep]

Agreed. No company dropping tens of thousands on a Mac Pro would still be using that machine as a daily driver 7-10 years down the line. Computers will be turned over somewhere between every 1-3 release cycles to ensure they can offer the performance they need and so machines are in warranty etc.

Unless the company massively changes its entire business, how could it go from needing top of the line equipment in 2019 to needing iMac level speed in 2024 (by that I’m making the admittedly gratuitous implication that a 2019 Mac Pro is as fast as a 2024 iMac).

I know why you say this but actually there was a market segment where this already happened. A lot of publishing and design houses shifted away from G5/Mac Pro towers to using 27" iMacs because they got good enough for the task and with the built in screen that was enough for a double page spread they could save a lot of cash per seat over a Mac Pro with a 27" ACD.

It really depends on the specific use case but there are certainly some that need a Mac Pro now that might not in 5 years or so.

 

[Pressure]

Where ARM is a bit lacking still is in single-core performance compared to the best Intel & AMD, but it's not too far off - there are 3.5GHz ARM cores, even though most run at 2.5-3GHz for power reasons.

Speed ≠ performance.

You can go wide and slow or narrow and fast.

 

[Emanuel Rodriguez]

Speed ≠ performance.

You can go wide and slow or narrow and fast.

Yep. It has been shown that Apple's mobile processors perform far better than most of their competition. The GHz might not beat everyone, but that really doesn't matter. If Apple can scale up the performance to desktop-levels, Intel is definitely in trouble. Other OEMs will want to be able to compete, and transitioning to ARM themselves might be the only solution. They'll have to deal with the fact that Microsoft still hasn't put out a usable version of Windows on ARM, though. Apple may have a winning combination here.

 

[val3rie]

Hi, this discution is quite interesting. i'd like to have your thoughts, if you think that a future Macbook pro 16' ARM would be faster (for video export use Premiere & After Effect) than a Macpro 2019, 16 core, 96 ram, Radeon Pro Vega II duo? someone told me that it's not comparable, because a graphic card in a desktop computer is more powerfull that the one in the laptop, and also you can add more ram inside..
But this ARM compressor seems to be really the next best thing...so shall i wait and buy for cheaper a laptop, or get a 18000€ big computer... ?

 

[dburkhanaev]

Not necessarily, assuming you are referring to the capacity of the current Intel MacPro.

A future ARM-based MacPro doesn't need to set a precedent for powerful ARM CPUs.

Here are the high-level specs for the Ampere 80-core ARM CPU released this March:

Ampere Altra Features
• Up to 80 single-threaded cores in a 1P and 160 cores in a 2P platform
• 7nm process technology
• 8 channels of DDR4-3200 at 2 DPC, supporting up to 4 TB memory per socket
• 128 PCIe Gen4 lanes in 1P and 192 PCIe Gen4 lanes in 2P platforms
• CCIX for coherent accelerator attach
• Two 128 bit SIMD units
• AI inference acceleration using int8 and fp16 instructions
• Server class RAS
• Arm v8.2+, SBSA Level 4

Current Mac Pro tops out at 28-cores and 1.5TB RAM. Just adorable!


[automerge]1593577996[/automerge]


AFAIK, the computer will still continue to function perfectly even without OS updates...so definitely not a paperweight.

There are plenty of 20 year-old-plus computers running just fine all over the world. Sure, there isn't any significant software development for them going on, but they still work and do their job.

Yes and they work perfectly well doing the same thing today that they did 20 years ago. Like my fleet of vintage Apple PowerMacs and PowerBooks. Love em and they do run, but it’s always a neat surprise to find new software to run on them. And you hope the tenfourfox people keep their browser project alive because that’s your last tie to the working web

while I appreciate where you’re coming from, this is the second time I’ve bought a relatively expensive laptop from Apple just prior to a processor transition announcement. Believe me you won’t get six years of OS updates on that Intel Mac. It doesn’t matter how large apples install base is on Intel. You’ll get like two major OS releases after Big Sur and two years of patches beyond that. I’m looking for a buyer ASAP for my nearly new MacBook Pro that I can’t return.

 

[Kostask]

Nvidia abd Apple don't get along because they are both control freaks. Nvidia won't allow Apple to write direct to the hardware for Metal and MacOS by not giving Apple the low level information needed. Apple refuses to let third parties write drivers for what they term essential software drivers for fear of instability and being held up by Nvidia when new software releases are sent out. There are also pricing/cost considerations, without a doubt.

 

[dmccloud]

Yes and they work perfectly well doing the same thing today that they did 20 years ago. Like my fleet of vintage Apple PowerMacs and PowerBooks. Love em and they do run, but it’s always a neat surprise to find new software to run on them. And you hope the tenfourfox people keep their browser project alive because that’s your last tie to the working web

while I appreciate where you’re coming from, this is the second time I’ve bought a relatively expensive laptop from Apple just prior to a processor transition announcement. Believe me you won’t get six years of OS updates on that Intel Mac. It doesn’t matter how large apples install base is on Intel. You’ll get like two major OS releases after Big Sur and two years of patches beyond that. I’m looking for a buyer ASAP for my nearly new MacBook Pro that I can’t return.

Intel Macs (at least 2017 and newer) will get more than two OS releases. When the PPC-Intel transition occurred, OS releases were roughly every two years. Now Apple has moved to an annual release cycle, and given that ~2012 is the cutoff for Big Sur, it stands to season that a current Mac should easily have four OS updates before support is killed off.

 

[JMacHack]

I know why you say this but actually there was a market segment where this already happened. A lot of publishing and design houses shifted away from G5/Mac Pro towers to using 27" iMacs because they got good enough for the task and with the built in screen that was enough for a double page spread they could save a lot of cash per seat over a Mac Pro with a 27" ACD.

It really depends on the specific use case but there are certainly some that need a Mac Pro now that might not in 5 years or so.

The small magazine company I work at now is doing just that (Facilitated by yours truly, if I can brag). I'm sitting next to a Mac Pro 4,1 right now that's gonna get relegated to legacy duty and we're gonna move to iMacs because they're enough performance to put together a magazine. It's happened gradually too, if the G3/G4 towers in the back indicate. We already have one older 27-inch iMac that we use.

 

[Erehy Dobon]

Speed ≠ performance.

You can go wide and slow or narrow and fast.

Correct.

The gigahertz wars ended about twenty years ago but some people here apparently didn't get the memo.