References to Intel's Latest Ice Lake SP Processors in Xcode 13 Beta Hint at Forthcoming Mac Pro Update

[deconstruct60]

I don't think this is true. Ice Lake W (if released) will almost certainly use socket 4189. I'm not sure there are any historical examples of memory channels increasing without pin counts. If there was a 3647 variant it was probably a very early engineering sample, I think.

Just because the Memory channels are there on the die doesn't mean Intel has to hook them up to pins.
For the max sized die ( XCC ) with 38-40 cores they probably also need the larger package size for better thermals. However, for the 'low core count" ( LCC) dies they could just skip the 2 memory controllers being hooked up and just more to higher base clocks on the smaller boards space.

They didn't do it with memory channel pins before but I think they did this with UPI I/O in the past. Using the same die for multiple products. Downsiding the I/O to "lower" die into a more affordable price segment.

At the point where in the same 8-20 core count as the W3200 is the two extra memory controllers going to make a huge difference to their implementation given a bump in RAM clock speeds ?

Moving from 6 to 8 channels is so the additional cores don't 'starve' waiting on data when under max memory bisection bandwidth loads. If have 12+ more less cores that "max pressure" is going to be lower.

The problem with cranking up to the bigger package is that soak up more board space. ATX boards are only so big. 3497 is already heavily on the "large" size. More DIMMs also soak up more board space. At some point likely to end up with less general purpose PCI-e slots . In the "enthusiast" market that can be 'bad' becauase more slots is "more power".

 

[csdesigns]

Still. To spend this kinda money when you know in the future Intel support will end.

I dont understand? Why push expensive HIGH POWER workstations when we all know the future is CLOUD COMPUTING and running applications from a web browser on say a fast 6G internet connection and monthly rental of software.

We could be looking easily at 8 years of security support updates. It may make no sense for you, but then again, you may not be the intended market for this machines.

If the PowerMac G5, machines limited to Mac OS 9, and the "granddaddy architecture transition of them all" to PowerPC from 68k are any indication, then there is no way in hell Apple has any intention to support Intel Macs, even with just security updates, 8 years into the future. That is a foolish bet to make given Apple's history. (And yes, I still feel burnt by my Quad G5 purchase.)

Regardless of future support (at most, I can envision 3 more years of OS support and maybe 5 years of security updates for Intel systems once a Mx-based Mac Pro is released), for the clients who need the most powerful Macs (i.e. not your typical customer and around just the 1%), then updating the Mac Pro to the latest Ice Lake chips makes a lot of sense. The real-world performance increase on Ice Lake is HUGE. It's really hard to state just how much of a performance increase we're seeing on Ice Lake chips under Linux at least, and if Apple doesn't update the CPUs in the Mac Pro, then it will be the trash can situation all over again – but it's also not like Apple always seems to care too much about that either, so who knows...

 

[deconstruct60]

No way. They'd need to completely redesign it anyway for the new internals, so why reuse on old design? It assumed 3 heats sources - CPU + 2 GPUs - which was part of its undoing. Any new machine would have a single SoC.

It assumed 3 roughly equal heat sources.

If Apple only going to support one, unified memory GPU then the CPU and GPU are even more thermally coupled than on the MP 2013. ( can't get much closer than on the same multiple chip module. ) Apple is probably not moving away from that "corner" that they like.

That isn't the major "paint ourselves into a corner" issue that would case a major problems for a MP 2013 volume sized product.

The other substantive issues MP 2013 was the whole "one and only one" internal drive problem and leaning too hard on Thunderbolt. With most data moving to SSDs that has a high degree of overlap.

If they tried to use the GPU slots in the MP 2013 basic design for storage ( e..g., one for Apple SSD NAND modules , one for some M.2 SSD slots ) then that would be the thermal mismatch problem.

Even if tapped danced around that, it wouldn't solve the additional non storage add-in-cards issues ( multiple A/V capature cards , > 10GbE , etc. ). The "Just buy an external PCI-e card holder" didn't fly with many hard core Mac Pro users.

That said they seems to have swapped the "2 GPUs for everyone" pain-into-corner problem for a "1 GPU for everybody" problem. macOS 12 on M-series apparently still only support just soley Apple GPU is a not a good sign. It is is a very similar "paint into corner" syndrome. It is has very large use case cover of the overall Mac users base, but it falls apart at the upper edges. Like the MP 2013, I suspect it will take 3-4 years for Apple to unwind from that iGPUs for everyone stance. So a stop-gap , W-3300 at the high end will give them a longer runway to adjust.

[ Some chance that will end up Unified Memory over CXL/PCi-e v5(or higher) still homogenized on an Apple implementation. ]

The old one was a bit of a failure, not selling well and getting universally nicknamed 'the trash can'. I think Apple will want to move on.

The MP 2013 sold well enough to make money. It didn't "print money" but it very likely didn't loose it either.
The iMac Pro sold reasonably well also up until Apple let it go relatively stale.

Apple probably does have some obsessive compulsive disorder on limits on the footprint of a desktop Mac.
MP 2013 fit inside the footprint of the Mac Mini. The pedestal of the iMac Pro did also. That could end up the real "paint themselves into a corner" if they don't ease up on that. Where the "smaller" Mac Pro has to take cuts in multiple dimensions ( not just height).

I suspect they will just cut the height though on the Mac Pro 2019 design for Apple Slicon variant. Fewer slots because leaning on the iGPU more heavily but not taking it all the way down the zero ( so can add-in m.2 , A/V capture , network cards ... even if leave out dGPUs for a very long time. )

 

[Freeangel1]

We could be looking easily at 8 years of security support updates. It may make no sense for you, but then again, you may not be the intended market for this machines.

Lets see when they come out, The pricing.

Willing to bet they will be even higher priced than the 2019 Mac Pro model.

I bet you could build one for almost half the cost. A Hackintosh. using the same components.

 

[deconstruct60]

Great news - means years more Mac OS support for Intel macs too.

Apple hasn't replaced the iMac 27" . It is a bit buried by they are still selling the non-Retina , anceint MBA processor powered 21.5" iMac version. The future year support has been rolling forward now anyway.

If Apple does a M-series , "half sized" Mac Pro overlapped with a refresh W-3300 Mac Pro then that will go a bit.
But if Apple stops the W-3300 version 1-2 years later it won't be much of an extension. Depends upon how that last Intel Mac Pro sales go. If sales drop a lot, Apple may turn it off not that far afterwards.

The probably going to be in 2023-2024 both AMD and Intel are going to quickly move past Ice-Lake SP. The non Mac competition in the workstation space is going to get much harder. On the Mac side, any following on to M1=larger-die iteration will also be peeling off user base also. ( Same with how Apple prioritized iMac Pro over doing a Mac Pro tower. Apple is going to spend more money and effort enticing people out of this system class than keeping up with it )

Apple has some very slow moving client blocks. That "edu" iMac and the MBP 2012 are substantive examples. The inertia of folks who are holding onto Macs longer life cycles is going to protect macOS Intel getting upgrades much more than an increasingly smaller , niche at the top. Super small number user base likely means they get cut off "earlier than Vintage age out" at some point. ( Apple would strip features and basically would shift more to just security updates. )

Shifting from W3200 to W3300 does have to be very costly tweak. Straightforward board updates. ( PCI-e v4 to slots 1 and 3 , RAM trace tweaks, firmware tweaks , new socket mount, minor tweak to chipset implementation). That will mean they probably won't need years make their money back on that effort. ( Intel also isn't going to want to keep making these CPUs very long time either. This chip isn't "SuperFin 10nm" process. It is the harder to deal with one. ) [ And if Apple put some W-Ice-Lake working into a iMac Pro prototype then even more "work" that is already paid for or reuse on base firmware work and platform feature coverage. So any abanoned iMac Pro work could shorten this Mac Pro update also. W-Ice-Lake was really late so there is probably some prep work that was in flight before completely finalized on exact M-series transition timeline. ]

 

[deconstruct60]

Lets see when they come out, The pricing.

Willing to bet they will be even higher priced than the 2019 Mac Pro model.

For the same core count? Probably not.

Cut off offering 8-12 core options? Probably yes. If they deploy a M-series iMac large screen and/or M-series "Half sized " Mac Pro then even more probably yes.

However, there is pretty good chance the 28 core W-3300 will be priced much lower than the W-2300M version was. In 2020, Intel cut the prices on the W-2200 series that Apple never picked up due to pricing competition from AMD Ryzen and Threadripper. That competition is even more problematical for Intel now ( Fall arrival of Threadripper update is going to put them into a deeper hole. ) A super-tax on > 1TB RAM isn't going to fly even more in 2021-22 than it did in 2019-20. Intel is loosing share here. Apple's "tax" on top of the Intel RAM tax was painful. Intel isn't going to have problems rounding up high volume customers for W-3300 also. ( Probably going to shift to a W-3400 (off the Sapphire dies ) within a year or so ( W-3300 isn't gonig to have an unusually long run)

Now when walk up the a 38 core W-3300 ... yeah that will creep back to the nosebleed range . Relatively high clocked versions of those are likely binned super thin coming out of Intel fabs at the "plain" 10nm process.

I bet you could build one for almost half the cost. A Hackintosh. using the same components.

Matching PCI-e slot numbers, DIMM slot count , and power supply ? The Xeon W chips are they major drivers of costs for the first 1-2 entries in the line up. At the top end? Yeah... just simply not selection the ">1TB" taxed CPU versions is a $2-4K savings after both the Intel and Apple tax.

 

[csdesigns]

Lets see when they come out, The pricing.

Willing to bet they will be even higher priced than the 2019 Mac Pro model.

I bet you could build one for almost half the cost. A Hackintosh. using the same components.

The pricing will indeed be interesting. Word is that Intel is being quite competitive with the pricing of chips to head off AMD's charge. So prices for Ice Lake are most likely to be less than they were similar Cascade Lake processor part numbers that are considerably less performing. Whether Apple feels like they need to pass these savings on to customers is anyone's guess, but that decision will probably be somewhat based on what the PC competition does.

 

[deconstruct60]

macOS 12 now has features only for the M1 chips

Are you sure about this?

Several macOS Monterey Features Unavailable on Intel-Based Macs

While there are many great new features in macOS Monterey, several of them are not available on Intel-based Macs, according to Apple. On the macOS...

www.macrumors.com

Places where Apple is leaning pretty heavily on the AI/ML and the NPU cores are going to produce gaps. T2 chips don't have NPU cores so they can't really backstop that kind of computation and the older Intel chips don't do ML inferencing as well (as even the latest Intel CPUS do let alone Apple's M-series).

While T2 does have an Apple image processing subsystem it is relatively ancient compared to the latest versions. ( more advanced Portrait mode over time. )

Over time there will probably be more quirks that are Apple GPU only optimized. ( although still small in number, some Metal features only work on newer Apple GPU).


P.S. Apple is doing a pretty poor job of explicilty documenting this with the usual marketing and technical documents they provide. [ but "glossy" tech support isn't really new for Apple. "feel good" support as opposed to answer my detailed question or solve my problem with tech insights support. ]

 

[farewelwilliams]

No way. They'd need to completely redesign it anyway for the new internals, so why reuse on old design? It assumed 3 heats sources - CPU + 2 GPUs - which was part of its undoing. Any new machine would have a single SoC.

The old one was a bit of a failure, not selling well and getting universally nicknamed 'the trash can'. I think Apple will want to move on.

I meant the general concept. One giant fan, components placed surrounding a thermal core. CPU on one side, GPU on another, and SSD/RAM on the third side.

 

[farewelwilliams]

Yeah, that's not great.



Well, that's what I'm saying. Lots of people who might have bought the 2013 Mac Pro if it had been more of a general-purpose high-end machine.

I'm saying you can't just convert every single piece of software to use GPGPUs. But there are plenty of subroutines that involve SIMD. Something as basic as sorting can be accelerated from GPGPUs which is some of the stuff that Safari does today with Metal.

 

[chucker23n1]

I'm saying you can't just convert every single piece of software to use GPGPUs. But there are plenty of subroutines that involve SIMD. Something as basic as sorting can be accelerated from GPGPUs which is some of the stuff that Safari does today with Metal.

I know. But there are many fields where you might be interested in a high-end Mac, but most of your workloads don't involve GPGPUs. So, again, I'm surprised they ever thought that was a good bet, even leaving aside how OpenCL panned out or how GPUs ultimately demanded more heat.

 

[Stephen.R]

I know. But there are many fields where you might be interested in a high-end Mac, but most of your workloads don't involve GPGPUs. So, again, I'm surprised they ever thought that was a good bet, even leaving aside how OpenCL panned out or how GPUs ultimately demanded more heat.

Dual gpu is definitely overkill for anyone who just wants globs of cpu cores, memory and storage options.

 

[farewelwilliams]

I know. But there are many fields where you might be interested in a high-end Mac, but most of your workloads don't involve GPGPUs. So, again, I'm surprised they ever thought that was a good bet, even leaving aside how OpenCL panned out or how GPUs ultimately demanded more heat.

i'm saying most workloads can involve GPGPUs which is why Apple placed a bet on multi gpus.

 

[mtneer]

I thought that the M1 already beat this in all single and multi thread bench tests. What's the point?

 

[deconstruct60]

I know. But there are many fields where you might be interested in a high-end Mac, but most of your workloads don't involve GPGPUs. So, again, I'm surprised they ever thought that was a good bet, even leaving aside how OpenCL panned out or how GPUs ultimately demanded more heat.

Errr, GPUs demanded more heat because GPGPUs largely work over a substantially large number of workloads. Apple's bet was way off in how fast that would accelerate and how quickly would either speed up the main backhaul (PCI-e bus) or move to affordable backhaul that was more focused on flatter memory space support than odd-ball goosing of framerates.

And on how much cutting off Nvidia would "cost" if AMD stumbled (which they did).

OpenCL Apple lost control ( or at least control over direction they most wanted to go and Nvidia's "embrace-extend-extinguish" kneecapping of that standard was pretty effective. ). OpenCL also got caught up in the morass of OpenGL evolution. Those are more predictable with 20/20 hindsight.

Those last two is probably major reason they went proprietary and fewer partners over time.

To a large extent the MP 2013 --> iMac Pro --> M-series transition. Megawatt GPUs were not inevitable back then . All the 2nd GPU support that deployed over time also substantively help a sizable faction of the Mac laptops with eGPUs. Frankly the Vega II Duo does cover more workload than the Vega II solo at launch ( the Vega II Duo is covered by a single fan. )

Apple tying the MP 2013 to a smaller desktop footprint than the Mini was a bigger problem. One 240mm fan would have worked better than an approximately one 140mm. And still would be in the 11 inch square footprint. Still needed another "container box" option , but could have had easier time on that 6 year run when they didn't.

 

[hatchettjack]

I thought that the M1 already beat this in all single and multi thread bench tests. What's the point?

This is the new intel chip! M1 could only dream of beating this!

 

[hatchettjack]

Does power consumption really matter on machines with no battery?

 

[chucker23n1]

i'm saying most workloads can involve GPGPUs which is why Apple placed a bet on multi gpus.

Maybe Apple believed that fallacy as well.

This is a bit like Intel betting on Itanium because surely, a sufficiently advanced compiler is just around the corner.

 

[deconstruct60]

I thought that the M1 already beat this in all single and multi thread bench tests. What's the point?

Multi Thread? No.

Mac Pro 28 core Multithread 19552 https://browser.geekbench.com/macs/249
Mac Mini M1 Multhread 7412

that is 2.6x as fast advantage for Intel. ( -62% drop from the 28 core score ).

Threadripper 3990X 25016 https://browser.geekbench.com/processors/amd-ryzen-threadripper-3990x

An even bigger beat down (3.4x) . Couple more months and will have a next gen of that from both Intel and AMD.

There is lots of "it will scale perfectly linearly" from folks projecting that a "Much bigger M1 die " will catch the 28 core. We'll see if that pans out. But probably for at least another year or so it is substantially behind the Intel/AMD high core count offerings if willing to plug into the wall and add a large capacity power supply. (and don't anchor yourself in a 2018-2019 time machine to do comparisons with).


P.S. if restrict comparisons to the bottom half of Intel's line up then yeah... M1 has more than a few wins against those. However, the Mac Pro doesn't use those, so not particularly relevant.

 

[chucker23n1]

Does power consumption really matter on machines with no battery?

Yes. Power consumption means heat dissipation.

 

[elptdbi3lYI]

I thought that the M1 already beat this in all single and multi thread bench tests. What's the point?

Can m1 have more than 16gb of memory, can it support more than 1 external display, can it have more than 8 cores?

 

[deconstruct60]

Yes. Power consumption means heat dissipation.

which matters less for a larger container than a smaller one. If baselining on the Power G5 and 2006-2012 chassis history ... what is the lower "buying" a whole lot of ?

To push into a MP 2013 container at around 1/6 the volume? Sure there is a helpful gap there.

Lower Power is going to help Apple move more into the same "go smaller volume" path they are on with the rest of the Mac line up. But that leads to fewer "add in" choices for the users also.

 

[mode11]

I meant the general concept. One giant fan, components placed surrounding a thermal core. CPU on one side, GPU on another, and SSD/RAM on the third side.

Sure, but that's not really how an SoC works. It would be CPU + GPU + RAM in one package, with the SSD perhaps requiring a small passive heatsink. Essentially, more like a scaled up Mac mini.

I'm sure an SoC could be made to work in a chassis that looks just like a 2013 MP. But Apple already put that design behind them when they reverted to a big aluminium tower.

 

[mode11]

The MP 2013 sold well enough to make money. It didn't "print money" but it very likely didn't loose it either.

Is that much of an achievement though, by Apple's standards?

Breaking even was helped by not updating it for 6 years. No way Dell would get away with that, but those are the spoils of a captive market I guess.

I find it hard to believe it would have been impossible to update the 2013 MP in that time, even given the restrictions of the chassis. Sure, it would never have taken a Radeon VII, but Apple could have still moved to newer Intel CPUs, faster RAM / SSDs, Thunderbolt 3, a pair of RX580s etc., had they been so inclined.

I suspect Apple was seduced by the concept of the compact cylinder, which in turn relied on spreading out the heat sources in order to work. Contrary to their later spin, I don't think it was lead by a firm conviction that GPGPU was the future. That was a bit of a stretch - and an unnecessary gamble if their priority was making a Pro workstation, not a design icon that showed they 'still had it' after Steve passed.

 

[farewelwilliams]

Sure, but that's not really how an SoC works. It would be CPU + GPU + RAM in one package, with the SSD perhaps requiring a small passive heatsink. Essentially, more like a scaled up Mac mini.

I'm sure an SoC could be made to work in a chassis that looks just like a 2013 MP. But Apple already put that design behind them when they reverted to a big aluminium tower.

For high powered desktop graphics, GPU would likely not be on the same chip as the thermal envelope would be too limiting.

M1 uses DDR4X and is shared among CPU/GPU/etc...likely a Mac Pro graphics by Apple would not tap into the unified memory, but instead have a separate DDR6 pool of memory.

Or rather, there will still be a GPU on the SOC, but that would be "low powered" GPU and will spin up dedicated GPUs for higher end workflows.