MP 7,1 What does Apple's switch to ARM based processors mean for the future of the Mac Pro

[deconstruct60]

....
Apple however went the opposite direction. They didn't make their chips efficient by making them simple. They made their chips efficient by making them complex. Their CPUs can process more instructions per single clock than any x86 chip and they can do it by consuming very little power. The result is a mobile phone CPU core running at sub 3Ghz while consuming 5 watts of power that can trade blows (in synthetic benchmarks) with a desktop x86 core running at 4.5+Ghz consuming who knows how much.

Errrr.. More instructions per clock than the baseline ARM design but the full set of x86 chips? Got any backup for that.

Apple's chips are in part just bigger than many of the other ones in their classe.

The A13 is

".. The new die is 98.48mm² which is 18.3% larger than the A12 of last year. ..."

There are a few here that have cellular modems in the m that are smaller (4th column from left in table(s) below)

"...

Snapdragon 8cx	7nm TSMC	8.3 x 13.5	112*	> 5.3b < 10.6b	> 56.4 < 94.6
Snapdragon 845/850	10LPP Samsung		94	5.3 b	56.4
Snapdragon 835	10LPE Samsung		72.3	3.0 b	41.5
Kirin 980	7nm TSMC		74.13	6.9 b	93.1

8-core Ryzen	14nm GloFo	22.06 x 9.66	192	4.8 b	25.0
Skylake 4+2	14nm Intel	13.31 x 9.19	122	1.75 b	14.3

"

Spotted: Qualcomm Snapdragon 8cx Wafer on 7nm

www.anandtech.com

There are two ways to increase effective IPC. One way is to wider. Anohter way is crank up the cache hits so that issuing less "no-op". if the clock is running fast but if all doing is issuing "no ops" 10-40% of the time then effective 'wall clock' isn't going to necessarily represent how 'wide' the core function units are.

Somewhat questionable whether Apple's system memory (cache) and layerin will do as well on 40GB or 400GB working set footprints as it does on 4GB working set footprints. iOS wouldn't even let anything even allocate over

"...On iOS, 429.mcf was a problem case as the kernel memory allocator generally refuses to allocate the single large 1.8GB chunk that the program requires (even on the new 4GB iPhones). I’ve modified the benchmark to use only half the amount of arcs, thus roughly reducing the memory footprint to ~1GB. The reduction in runtime has been measured on several platforms and I’ve applied a similar scaling factor to the iOS score – which I estimate to being +-5% accurate. The remaining workloads were manually verified and validated for correct execution. ..."

The iPhone XS & XS Max Review: Unveiling the Silicon Secrets

www.anandtech.com

Apple has made progress here since 2018 but one (or two) orders of footprint magnitude progress. Small working set , hot drag racing benchmarks are nice. it isn't that hard to tune to a subset of those to make yourself look good.

What if you have more room to spare than just 5 watts however? What if you can give the CPU 30 watts? 50 watts? 150 watts? You see what I am getting at? Apple claims very confidently that their CPUs scale up. Personally, I doubt that we will see an Apple CPU running at 4Ghz anytime soon. But they most likely can run at 3.5Ghz — which will give it a healthy performance boost over anything in the x86 world while still consuming very little performance in the relative terms.

Scaling up on cores is gimimicky is the workloads aren't really scaling up all that much. 10 cores at a 3GB working set will probably do a bit worse than 15 cores at a 3GB working set if the cache striding manages not to overwhelem the memory channel(s).

So let's have a look at a practical example. Intel Xeon W-3275M, which is the largest Mac Pro CPU currently packs 28 cores running nominal 2.5Ghz into a 205Watt TDP package. Well, Apple could pack 40 A13 cores running at 2.5Ghz into the same package. And an A13@2.5Gz core is considerably faster than a Xeon@2.5Ghz core.

packing the cores isn't the issue. At some point the cores can't all sit close to the singular system cache. At some poin will need a ring or mesh and the latencies will go up and/or get more irregular. And then what happens to Apple's design. Apple has simply just avoided issues. They have grow core count as fab modes have gone down. What happens when can't wait for fab mode to go down to doube , quad cores and caches, GPU , memory controllers all get more spread out ?

I don't think it will completely stump Apple but also think they have tip toe around some issues also because not chasing core counts and pushing more optimization tasks on the software developers (e.g. smaller, limited RAM working sets , fewer ( than high double digit or triple digit ) threads , hand optimizing GPU low level code.and cache loading to minimize bandwidth, etc. etc. )

Of course, making large chips like that is not trivial at all, and Apple probably still has a lot of work to do before they can deliver it. But this is where their efficiency can be turned into some serious power.

Doesn't talk to dGPUs. Doesn't do Thunderbolt speeds. Or even 10GbE I/O. RAM capacities one or two levels higher. Apple systems are barely over the 32-bit , 4GB limit. There are a couple more digits to go to be in the current Mac Pro capacity range. The current A-series has a MMU but is it seriously tasked with anything substantial ? Apple has thrown some advanced security tasks at it but capacity?

 

[ZombiePhysicist]

Then I fail to see why you are stating that its (even now) out of date? The latest Xeons possible only support PCIe 3.0. I fail to see how that means its an out of date system when ANY Intel based computer is on PCIe 3.0. Therefore, PCIe 4 and 5 is still rolling out.

Again, the 10th generation Intel chips - i9-10900k is still on PCIe 3.0. So what if AMD has it? That does not make the Mac Pro immediately outdated if even the latest Intel chips released in 2020 are on PCIe 3.0.

Actually it does. It's one version behind. By definition, dated technology that runs at half the bandwidth. Hand waving that "it doesn't matter that it's one version behind" is something you can try to advance, but that it's out of date technology isn't reality.

 

[pasamio]

You can think of business reasons why it was not convenient for apple. Fair. Professionals dont care. And they can think of many core/performance reasons to get what is current and faster and better. Also fair.

When your 'convenient for your business' reasons clash with the reality of the competition, in a fair system, you tend to lose.

To write off a multiyear partnership with a chip vendor with whom you likely have solid established ties and a need to meet supply for existing devices to switch to a brand new vendor for two years or less as "convenient for your business" belittles the complexity of such arrangements, the amount of forward planning that generally goes into these partnerships and the planning for ensuring supply of product to meet the expected demand.

My point was that there are plenty of business reasons why the business that is Apple might choose to make the decisions it made for the Mac Pro 7,1. There is enough evidence that folk bought the hardware which would indicate that some amount of professionals bought the device and felt it fit for purpose. I've commented elsewhere on this forum that there is a point when what makes sense for the consumer and what makes sense for the business diverge sufficiently that people stop buying the product but given that there has been reasonable evidence that the Mac Pro 7,1 is selling that there exists a market that is buying the computer so those interests haven't sufficiently diverged.

I think the transition will be the test of that interest between the company and it's consumers ultimately diverging too far but until an ARM powered Mac Pro appears we won't know about that.

 

[Ethosik]

Actually it does. It's one version behind. By definition, dated technology that runs at half the bandwidth. Hand waving that "it doesn't matter that it's one version behind" is something you can try to advance, but that it's out of date technology isn't reality.

But it is not one version behind on Intel processors. It is the most up to date version available.

 

[leman]

Errrr.. More instructions per clock than the baseline ARM design but the full set of x86 chips? Got any backup for

[...]

There are two ways to increase effective IPC. One way is to wider. Anohter way is crank up the cache hits so that issuing less "no-op". if the clock is running fast but if all doing is issuing "no ops" 10-40% of the time then effective 'wall clock' isn't going to necessarily represent how 'wide' the core function units are.

Apple did both. Check out the in-depth Anandtech article. A13 has 6 integer ALUs and 3 128 bit fp/vector units, 50% more than Intel Ice Lake. Also, it can decode and issue 7 instructions per cycle. A13 also has a very large L2 cache and estimated 16MB of SoC-level cache. Check out the SPEC benchmark results in the article, they are rather impressive for a CPU core running at 5watts of power.

Regarding everything else you said, I completely agree. So far, the only thing we know is that Apple has designed a very efficient CPU/GPU architecture that excels in IPC. We have no idea how it will integrate into an actual desktop system and how well it will scale. Still, I am optimistic. I don’t know about the pro workstation application and its challenges (my above post was more about discussing the relationship between performance and efficiency), but in the laptop space Apple chips have good chances to do well.

 

[Ryan P]

Good thoughts all.

I have a 7,1 myself, so hope it is not abandoned. if Apple, does one major update PCIe 4.0 etc...and a couple minor refreshes over the next 5 years, it may be that their consumer ARM level chips of 5 years from now do outperform the, by then, very stale Mac Pro over almost any workload. Apple can then just declare the Mac Pro form factor unnecessary.

Will be a very easy task if in 2025, Intel is on 14nm++++++++++ while TSMC is shipping 3nm in bulk.

 

[ZombiePhysicist]

But it is not one version behind on Intel processors. It is the most up to date version available.

yea and Intel are behind in cores, heat, power, and performance. Outright obsoleted by amd. There is a reason Apple is abandoning intel. They suck and are behind.
[automerge]1593907450[/automerge]

To write off a multiyear partnership with a chip vendor with whom you likely have solid established ties and a need to meet supply for existing devices to switch to a brand new vendor for two years or less as "convenient for your business" belittles the complexity of such arrangements, the amount of forward planning that generally goes into these partnerships and the planning for ensuring supply of product to meet the expected demand.

My point was that there are plenty of business reasons why the business that is Apple might choose to make the decisions it made for the Mac Pro 7,1. There is enough evidence that folk bought the hardware which would indicate that some amount of professionals bought the device and felt it fit for purpose. I've commented elsewhere on this forum that there is a point when what makes sense for the consumer and what makes sense for the business diverge sufficiently that people stop buying the product but given that there has been reasonable evidence that the Mac Pro 7,1 is selling that there exists a market that is buying the computer so those interests haven't sufficiently diverged.

I think the transition will be the test of that interest between the company and it's consumers ultimately diverging too far but until an ARM powered Mac Pro appears we won't know about that.

ummm they have quite ACTUALLY written intel off, abandoning them wholesale And jumping to another platform, their own processors.
[automerge]1593907529[/automerge]

Good thoughts all.

I have a 7,1 myself, so hope it is not abandoned. if Apple, does one major update PCIe 4.0 etc...and a couple minor refreshes over the next 5 years, it may be that their consumer ARM level chips of 5 years from now do outperform the, by then, very stale Mac Pro over almost any workload. Apple can then just declare the Mac Pro form factor unnecessary.

Will be a very easy task if in 2025, Intel is on 14nm++++++++++ while TSMC is shipping 3nm in bulk.

personally I hope they skip pci4 and go straight to pci5.

 

[pasamio]

ummm they have quite ACTUALLY written intel off, abandoning them wholesale And jumping to another platform, their own processors.

I was referring to you writing off such a relationship as merely "convenient for your business" when there is more depth to such negotiations. I apologise for that paragraph being confusingly written.

 

[ssgbryan]

There are still people that use the 2010 Mac Pro. In fact, I was using it until 2018 and I am not using it now only because it won't produce HEVC content. Not because its on PCIe 2.0. My GTX 1080 worked fine on it as PCIe 3 is backwards compatible.

I was using mine (see sig) until a few months ago. A Ryzen based system was much faster, and I no longer had to worry about Does this peripheral work; what is the work around for that; what software do I now need to replace because it no longer runs.

Now, I install software, and it just works....

I install hardware, and it just works....

It is easy to get used to install and get cranking on my stories.

 

[t90]

I was using mine (see sig) until a few months ago. A Ryzen based system was much faster, and I no longer had to worry about Does this peripheral work; what is the work around for that; what software do I now need to replace because it no longer runs.

Now, I install software, and it just works....

I install hardware, and it just works....

It is easy to get used to install and get cranking on my stories.

Windows 10 - It just works! Oh how the tables have turned.

I use both OSes daily and I’m leaning more towards windows nowadays. Apples quality control is going out of the window it seems. it’s such a shame.

Windows has its issues too, don’t get me wrong but overall it’s absolutely more performant and stable than macOS these days.

 

[thunng8]

Just my 2 cents from using Apple ARM dev machine. I've been comparing between Mac Pro 7,1. Epyc 7502p and Apple silicon. It's single thread performance is very good and multi-threading not so bad either.

I took some single thread & multi-threaded c++ code compiled with all the bells & whistles on each machine. I capped the # of cores on all machines to 8 to match the ARM cpu. AMD smoked everything. But my Mac Pro 7,1 got beat by ARM, not really surprised.

Keep in mind ARM has no SMT and I doubt it will any time soon. It only has 4 high performance cores. Other 4 are low performance. With that said. I am still skeptical how well it will actually scale up to higher core counts. Apple would have to remove the low performance cores to stand a chance. Then you have to contend with possibly using chiplets and communication between those. All of which AMD has figured out. Apple still has a long ways to go.

Also not surprisingly, the compiled binary on the ARM machine is double the size of the 7,1.

I think that's a very promising results from Apple. A13 is already 20-25% faster than A12 in single threads and presumably A14 would also be another similar bump in performance. I'd say A14 should be approx 1.5X faster than A12 in single threads.

The first macs are rumored to have 8 high performance cores and all fit within about 15W. That would be a game changer in thin laptops and a good basis for future scaling in desktops.

 

[Boil]

...it’s quite possible that Apple could produce a 65-95w version of the chip that absolutely demolishes anything Intel or AMD have at 135-200w.

But Apple is not making a CPU to go against Intel & AMD CPUs; they are making an APU to go against everything...

I could see a high-wattage APU from Apple:

32 P cores / 16 E cores / 64 GPU cores / 32GB HBM2e UMA / 420W

More DDR5 RAM & NVMe storage on the logic board...

Perfect for a Mac Pro Cube (entry-level workstation)

Just my 2 cents from using Apple ARM dev machine. I've been comparing between Mac Pro 7,1. Epyc 7502p and Apple silicon. It's single thread performance is very good and multi-threading not so bad either.

I took some single thread & multi-threaded c++ code compiled with all the bells & whistles on each machine. I capped the # of cores on all machines to 8 to match the ARM cpu. AMD smoked everything. But my Mac Pro 7,1 got beat by ARM, not really surprised.

Keep in mind ARM has no SMT and I doubt it will any time soon. It only has 4 high performance cores. Other 4 are low performance. With that said. I am still skeptical how well it will actually scale up to higher core counts. Apple would have to remove the low performance cores to stand a chance. Then you have to contend with possibly using chiplets and communication between those. All of which AMD has figured out. Apple still has a long ways to go.

But did you turn off SMT on the Intel & AMD chips? You know, to match the ARM cpu (which is actually an APU)...?!?

 

[leman]

Just my 2 cents from using Apple ARM dev machine. I've been comparing between Mac Pro 7,1. Epyc 7502p and Apple silicon. It's single thread performance is very good and multi-threading not so bad either.

I took some single thread & multi-threaded c++ code compiled with all the bells & whistles on each machine. I capped the # of cores on all machines to 8 to match the ARM cpu. AMD smoked everything. But my Mac Pro 7,1 got beat by ARM, not really surprised.

Keep in mind that the ARM DTK is running sub 10 watt. And it’s essentially a quad-core machine (the efficiency cores barely count for benchmarks). So if you want to be “fair” you should probably compare it to quad core Intel and AMD with SMT. And of course, A12 is two years old by now, A13 leads it by a healthy margin.

Apple already made it clear that they are keeping asymmetric design for the desktop, which makes sense for battery life.

Regarding binary size: did you per chance compiled it as fat binary? Which optimization level did you use?
[automerge]1593937965[/automerge]

I think that's a very promising results from Apple. A13 is already 20-25% faster than A12 in single threads and presumably A14 would also be another similar bump in performance. I'd say A14 should be approx 1.5X faster than A12 in single threads.

I’d be surprised if A14 were that much faster. They would either need higher clocks or an even wider core for that...

 

[thunng8]

Keep in mind that the ARM DTK is running sub 10 watt. And it’s essentially a quad-core machine (the efficiency cores barely count for benchmarks). So if you want to be “fair” you should probably compare it to quad core Intel and AMD with SMT. And of course, A12 is two years old by now, A13 leads it by a healthy margin.

Apple already made it clear that they are keeping asymmetric design for the desktop, which makes sense for battery life.

Regarding binary size: did you per chance compiled it as fat binary? Which optimization level did you use?
[automerge]1593937965[/automerge]


I’d be surprised if A14 were that much faster. They would either need higher clocks or an even wider core for that...

Apple are moving to TSMC's 5nm process. That by itself should get 15%. Put in some further optimisations in the chip, I think a 20% increase going from A13-> a14 is the bare minimum.

 

[leman]

Apple are moving to TSMC's 5nm process. That by itself should get 15%. Put in some further optimisations in the chip, I think a 20% increase going from A13-> a14 is the bare minimum.

Right, I forgot about the reported move to 5nm. In that case, the first Apple Silicon Mac is Definitely going to fly.

 

[deconstruct60]

....
personally I hope they skip pci4 and go straight to pci5.

Intel? Not going to happen. The PCI-e v4 processors are already built and in beta testing.

Apple? Quite unlikely. They don't even have a x16 lane processor yet. What exactly does Apple have "ants in its pants" desire to hook to that has PCI-e v5 ? Hooking to the AMD GPUs with PCI-e v4 that would be a target over the next two years, but beyond that? If Apple was doing any on chip "inter chip die" communiation they'd probably build their own propriatary comm ( that could be layered on top of PCI-e v5 PHYS, but probably no pins off chips. )

The folks in a rush to get to PCI-e v5 typically have a tie-in with CXL or CCIX . Apple has shown no explicit interest in those at all. In constrast Intel is looking to CXL to do CPU to (GP)GPU interconnection. They'll move in 2021-22 timeframe. IBM Power will. IBM Power had PCI-e v4 for years before it trickled down to the mainstream. PCI-e v5 will likely have a silimar trickle down path.

Mainstream GPUs probably aren't going to PCI-e v5 any time over the next couple of years. Removing the "price premium" for PCI-e v4 is probably going to be higher on folks list than moving on to something else.

 

[Voyageur]

But which cards require today the PCI-e 4 connection standard and no less? It seems that even in 2019-2020 the same usual top-end GPU cards still did not fully fill the potential of PCI-e 3. Or am I mistaken?

 

[Ethosik]

But which cards require today the PCI-e 4 connection standard and no less? It seems that even in 2019-2020 the same usual top-end GPU cards still did not fully fill the potential of PCI-e 3. Or am I mistaken?

Yes this is what I have a problem with the whole PCIe 4.0 and the Mac Pro having no future or being obsolete day one because its on PCIe 3.0.

 

[teagls]

But did you turn off SMT on the Intel & AMD chips? You know, to match the ARM cpu (which is actually an APU)...?!?

Doesn't matter. I only created 8 pthreads and assigned them directly to cores. Results are valid.

Keep in mind that the ARM DTK is running sub 10 watt. And it’s essentially a quad-core machine (the efficiency cores barely count for benchmarks).

Are you referring to just the chip? Or the total system. The highest I could pull was 27W fully maxing out CPU&GPU from the wall. The GPU is incredibly weak btw. Just dragging around windows the usage goes over 30%.

 

[ssgbryan]

But which cards require today the PCI-e 4 connection standard and no less? It seems that even in 2019-2020 the same usual top-end GPU cards still did not fully fill the potential of PCI-e 3. Or am I mistaken?

We don't have single video cards that need 16 PCIe 4.0 - yet. A 2080ti uses 8 PCIe 4.0 lanes. But the 2080ti isn't going to be the top dog GPU wise much longer.

BUT, once you start using multiple GPUs, the calculus changes - my motherboard (ASUS WS-570-ACE) can take 3 video cards that need 16 PCIe 3 lanes. Lots of use cases for it in machine learning.

Right now, PCIe 4.0 is about fast(er) storage - that will change in the future.

 

[AidenShaw]

they weren't about to jump ship to AMD for the final year or two of x86.

All Intel MAC's are x64, not x86.

 

[Blair Paulsen]

Most folks who bought 7,1 MPs really wanted to stay on OSX and avoid Windows/Linux. We can debate whether prioritizing that preference when choosing hardware is smart or not, but it clearly exists. Otherwise, the demographic for an expensive workstation would have decamped years ago. It may be pricey, but if it offers the desired user experience - including looking good in the home, nearly silent, robust build and able to host industry/task specific PCIe cards - then there is a cohort willing to pay the Apple Tax.

Whether the move to ARM decimates the resale value of the Intel 7,1 may depend greatly on something nobody knows for sure - how long it will take before an ARM solution appropriate for workstations is ready for prime time. Most people on this thread, including me, believe moving laptops to ARM is a viable near term strategy. IMO, until the more difficult task of creating a bomber SoC variant is accomplished, the Intel based 7,1 will still be worth a decent dollar for the OSX user who needs it for immediate projects.

I repeat my prediction that an ARM based 8,1 with significantly superior performance to the 7,1 won't be available until 2023. Even then, some subset of 7,1 users will have no need to upgrade for their use case. Bottom line: all computers lose value quickly - especially expensive ones - but IMO the 7,1 will suffer that only marginally more than other workstations.

 

[mode11]

All Intel MAC's are x64, not x86.

Well, if we're being pedantic, Mac is not an acronym, it's an abbreviation.

In any case, x64 is just shorthand for x86-64. Given this discussion is vs. ARM, I don't think I needed to be more specific.

 

[leman]

In any case, x64 is just shorthand for x86-64. Given this discussion is vs. ARM, I don't think I needed to be more specific.

Its Aarch64, not ARM

(just kidding)

 

[mode11]

Most folks who bought 7,1 MPs really wanted to stay on OSX and avoid Windows/Linux. We can debate whether prioritizing that preference when choosing hardware is smart or not, but it clearly exists. Otherwise, the demographic for an expensive workstation would have decamped years ago. It may be pricey, but if it offers the desired user experience - including looking good in the home, nearly silent, robust build and able to host industry/task specific PCIe cards - then there is a cohort willing to pay the Apple Tax.

Whether the move to ARM decimates the resale value of the Intel 7,1 may depend greatly on something nobody knows for sure - how long it will take before an ARM solution appropriate for workstations is ready for prime time. Most people on this thread, including me, believe moving laptops to ARM is a viable near term strategy. IMO, until the more difficult task of creating a bomber SoC variant is accomplished, the Intel based 7,1 will still be worth a decent dollar for the OSX user who needs it for immediate projects.

I repeat my prediction that an ARM based 8,1 with significantly superior performance to the 7,1 won't be available until 2023. Even then, some subset of 7,1 users will have no need to upgrade for their use case. Bottom line: all computers lose value quickly - especially expensive ones - but IMO the 7,1 will suffer that only marginally more than other workstations.

At the end of the day, if an individual or company buys a 7,1 today and is putting it to use, it's earning its keep. It will likely be a couple of years before the Mac Pro goes to ARM and even when it does, probably won't be much faster. The action will likely be at the lower end of the range, with the MP getting a 'good enough' CPU every few years. How much this affects 7,1 resale value depends on when you're selling. The value will hold up fine for a few years, due to those who need a powerful, dependable machine that avoids all the hassle of software, plugins and drivers that need to be rewritten for ARM. In five years time, though, x86 macOS will be well on the way to obsolescence, so a 7,1 will likely be worth about as much as a G5 Quad in 2010 i.e. not much.

The other variable is if the move to ARM (and the general iOSification of macOS) causes further high end users (e.g. of Adobe CC, or 3D packages) to throw in the towel and just move to Windows 10 PCs. There are only so many Logic and FCP users, which are bound to macOS. This would obviously reduce the market for Mac Pros in general.