Mac Pro 2023 = Another Mac Pro 2013 crisis and failure.

[VivienM]

Being faster than a 2019 Xeon isn't really much to write home about - I mean the original basic M1 had faster single core than the Mac Pro's Xeon W-3223 and similar multi-core. M2 Ultra seems about the same as a top-end consumer Intel CPU, which is fine but also not all that exciting.

Main issue is that the Intel CPU doesn't have the same RAM and GPU constraints as Mac Pro. At the workstation level there are also x64 CPUs that have way more cores than M2 Ultra.

Are Intel's Sapphire Rapids Xeons delivering a big performance boost over the 2019 Xeons? I got the impression... they didn't, but there aren't a ton of benchmarks out there for the workstation-flavoured chips.

Fundamentally, and I say this after buying a 2010 Mac Pro over the weekend and looking at how it benchmarks against newer things, large RAM capacity, PCI-E flexibility, and to some extent many-core performance are basically what defines a workstation, certainly not single-core performance or even multi-core performances.

And even, what is large RAM capacity? In 2012 when someone ordered my Mac Pro with 64GB of Apple RAM (which I presume was the top of the line option), the maximum RAM you could get in an Apple laptop was 16GB. In 2023, you can order an M2 Max MacBook Pro with 96GB. In 2012 the Mac Pro had 4X the RAM capacity of the laptop; in 2023, the Mac Pro has... twice... the RAM capacity of the laptop. If you look at it that way, then sure, the Mac Pro gives you half the benefit over a laptop that the 2012 models did.

And interestingly, if they had managed to build an M2 Extreme with 384GB of RAM... that would have maintained the same 4X ratio over the laptop as in 2012.

 

[VivienM]

And even, what is large RAM capacity? In 2012 when someone ordered my Mac Pro with 64GB of Apple RAM (which I presume was the top of the line option), the maximum RAM you could get in an Apple laptop was 16GB. In 2023, you can order an M2 Max MacBook Pro with 96GB. In 2012 the Mac Pro had 4X the RAM capacity of the laptop; in 2023, the Mac Pro has... twice... the RAM capacity of the laptop. If you look at it that way, then sure, the Mac Pro gives you half the benefit over a laptop that the 2012 models did.

And interestingly, if they had managed to build an M2 Extreme with 384GB of RAM... that would have maintained the same 4X ratio over the laptop as in 2012.

Just thinking more after posting the above.

I think these numbers prove the Mac Pro pessimists' point. Somebody, 10 years ago, bought a Mac Pro with 64GB of RAM and a Geekbench 6 score of 500 single core, 3800 multi-core.

Today, you can get a M2 Max studio or an M2 Max laptop with 64GB of RAM, a Geekbench 6 score of 2600 single-core and 14400 multi-core.

I don't have a good GPU benchmark but I'm sure the dual 5770s in my Mac Pro are crushed by the M2 Max as well.

So, every workload that doesn't require PCI Express expansion, basically, that the original purchaser of my Mac Pro needed 64GB of RAM for can be done on a laptop or Mac studio... 4-5X faster... a decade later.

Focusing on RAM capacity, the Mac studio or new Mac Pro will give you 192 gigs, i.e. three times what you could get 10 years ago. Assuming that you need about the same amount of RAM to do the same thing as ten years ago, that means, effectively, that the only workloads that need more are workloads that were impossible ten years ago because they needed more than 4X the maximum RAM on a 2010 Mac Pro.

I'm sure there are some workloads like that, but how many are there, particularly in Apple's traditional areas of strength (i.e. not datacenters crunching AI models)? And is it worth engineering a different chip/memory controller/etc for those folks?

What actually you conclude, looking at this, is that the main workloads from 2012 that can't be done with a laptop or Mac studio are the workloads that require PCI-E expansion rather than monstrous capacities of RAM or processing power. And... interestingly, that leads to the 2023 Mac Pro - effectively a Mac studio with PCI-E expansion.

I would finally add that the 2019 Mac Pro was the only Mac Pro with widely higher RAM capacity (the 2013 was limited to 128 gigs). And Apple Silicon was announced in 2020. Did anybody in 2019 really start saying "gee, Apple finally replaced a system they neglected for four years and finally has a system that can handle more than 128 gigs of RAM, time to port/develop our new super-cool RAM guzzling technology to macOS?" and now they're stuck like idiots because they can't get more than 192GB on a Mac anymore?

That's perhaps the bigger point - the 2023 Mac Pro runs circles around the 2013 trash can. Did anybody between 2019 and, oh, 2020 when the ARM shadow started looming over the Mac laptop really develop a ton of macOS workflows that require the additional capabilities of the 2019 that the 2023 doesn't have?

 

[Stevenyo]

We can argue about the MP separately (and I did acknowledge that "there are limits to the development dollars they can put into it"). But it was also clear your statement was about the HEDT market as a whole, not merely Apple's HEDT market. And it was based on statements like "Threadripper hasn't had a new release in years and the TR 7000 series keeps slipping." But that discounts the fact that AMD has been working on TR 7000, and *is* releasing it (https://www.tomshardware.com/news/r...0-storm-peak-cpu-surfaces-with-64-zen-4-cores). Release dates often slip. That hardly means a market is "dead".

So again, how do you explain AMD's substantial investment in the TR 7000 if the market is "too small to develop specific hardware for"?

I'm reminded of Mark Twain's famous letter to the New York Journal in response to an inquiry about his possible death: "...I have even heard on good authority that I was dead....The report of my death was an exaggeration."

The entire HEDT market is too small to develop silicon for. Threadripper is built out of Zen and Epyc chiplets. Xeon has long been datacenter first with some small HEDT applications. Apple creating a separate M2-based die for just the MacPro CPU, as well as all the other hardware and software infrastructure to support DIMMs and GPUs, would likely cost Billions, costs that could only be recouped by using those parts in millions of systems. That would require selling more Mac Pros than MacBook Pros, something that cannot happen in 2023. The other pathway to sales is to sell to datacenter customers, but for Apple that would require investing massively in a long term play to build out an entirely different business model, all for far less profit than farting out another iPhone with a slightly nicer camera.

By the way, the reason Threadripper releases keeps slipping or disappearing (TR5000 outside of OEM TR Pro, for example) is because they can sell the parts for more money to better customers as EPYC than they can as Threadripper. Because HEDT is dead. There isn't enough market between the 16-24 core top level "Consumer" CPUs and the totally different market for commercial datacenter. Again, in 2003 and even 2013, I had a tower as well as a laptop, and the tower was so much more powerful, for less money. The laptop was for light, on the go tasks or doing real work in a pinch. Now, in the 2020s? I rarely turn on my desktop machines anymore. My MacPro (admittedly an older model) collects dust. My Mac Mini acts mainly as a glorified NAS controller. My Ryzen 5900x Box gets turned on once or twice a month. But my iPad Pro, my MacBook Pro and my thin 14" PC laptop with Ryzen and Radeon 6000 parts get used all the time. Maybe I could have spent the Macbook Pro + mac mini money on a Studio+ macbook air with slightly better overall specs, but then I'd be stuck doing only lightweight work whenever I move even 2 feet from my desk. I could have invested in Ryzen 7000 or a 40 series GPU instead of a laptop, but then I could never just play a game from my bed or when I travel.

Portability won. The cloud won. HEDT and towers died. It's sad. 20 years from now I won't be able to pick up cheap old towers to play with like I can with PowerPC, and most Apple Silicon macs will likely be struggling from dead or dying soldered SSDs so won't even be fun for nostalgia hits. I wish someone had come up with a way to make portability, and cloud computing as powerful and convinenet to the average consumer as towers/modularity/on device computing/etc. But no one did. We're into a different era of tech now, and the couple million towers (desktop and HEDT) still sold each year are among the last of their kind.

 

[Stevenyo]

Always? I'd like to introduce you to my Power Mac G4 MDD, which is the loudest computer I've ever owned, Windows or Mac, in over 30 years. The second closest would probably have been a home-built Preshot Deleron around 2005-8 whose CPU fan would ramp up to 5000RPM if the room air was on the warm side, and I think the poor thermals in that machine probably have more to do with me having picked a less than optimal case, etc. - a professionally-engineered system would have run the same CPU just fine.

With the MDD, what happened is relatively clear - the G4's power consumption and heat output, especially if you had two of them, went up much more than a case design dating back to a single 400MHz G3 could handle, and the solution to just run the fans faster made for some serious noise. Something they corrected for with their next enclosure design.

Yes and no on the corrected for the next design, high end G5s are wind tunnels as well. Infinitely better thermal design than any G4 tower, sure, but quad 2.5 or dual 2.7 G5s put out a lot of heat, and cooling tech was nothing like today.

 

[theorist9]

The entire HEDT market is too small to develop silicon for. Threadripper is built out of Zen and Epyc chiplets. Xeon has long been datacenter first with some small HEDT applications. Apple creating a separate M2-based die for just the MacPro CPU, as well as all the other hardware and software infrastructure to support DIMMs and GPUs, would likely cost Billions, costs that could only be recouped by using those parts in millions of systems. That would require selling more Mac Pros than MacBook Pros, something that cannot happen in 2023. The other pathway to sales is to sell to datacenter customers, but for Apple that would require investing massively in a long term play to build out an entirely different business model, all for far less profit than farting out another iPhone with a slightly nicer camera.

By the way, the reason Threadripper releases keeps slipping or disappearing (TR5000 outside of OEM TR Pro, for example) is because they can sell the parts for more money to better customers as EPYC than they can as Threadripper. Because HEDT is dead. There isn't enough market between the 16-24 core top level "Consumer" CPUs and the totally different market for commercial datacenter. Again, in 2003 and even 2013, I had a tower as well as a laptop, and the tower was so much more powerful, for less money. The laptop was for light, on the go tasks or doing real work in a pinch. Now, in the 2020s? I rarely turn on my desktop machines anymore. My MacPro (admittedly an older model) collects dust. My Mac Mini acts mainly as a glorified NAS controller. My Ryzen 5900x Box gets turned on once or twice a month. But my iPad Pro, my MacBook Pro and my thin 14" PC laptop with Ryzen and Radeon 6000 parts get used all the time. Maybe I could have spent the Macbook Pro + mac mini money on a Studio+ macbook air with slightly better overall specs, but then I'd be stuck doing only lightweight work whenever I move even 2 feet from my desk. I could have invested in Ryzen 7000 or a 40 series GPU instead of a laptop, but then I could never just play a game from my bed or when I travel.

Portability won. The cloud won. HEDT and towers died. It's sad. 20 years from now I won't be able to pick up cheap old towers to play with like I can with PowerPC, and most Apple Silicon macs will likely be struggling from dead or dying soldered SSDs so won't even be fun for nostalgia hits. I wish someone had come up with a way to make portability, and cloud computing as powerful and convinenet to the average consumer as towers/modularity/on device computing/etc. But no one did. We're into a different era of tech now, and the couple million towers (desktop and HEDT) still sold each year are among the last of their kind.

It's interesting that you and I have traded places. I used to be a desktop guy, but starting in 2008 I switched to MacBook Pro's because they combined decent power and mobility. I owned 2008, 2011, and 2014 versions. However, while I don't need the high core counts, I did get frustrated with their TDP and RAM limitations, and have thus moved back to a desktop....while you were (I gather) using desktops then, but have recently switched to mobiles.

I was running my 2014 MBP with three external displays, including a 4k, and that puts a lot of stress on a mobile GPU, forcing it to run at high-fan pretty much continuously, which made the thing noisy. And if my room got too warm I'd need to turn on a button fan I set up behind it. If I didn't, it would essentially shut down due to thermal throttling.

And some of my calculations need more RAM than one can get in a laptop, so it was inconvenient when I enountered them.

I solved both of those problems with a 2019 iMac. It runs three large displays (including it's built-in 5k) quietly, and I was easily able to install 128 GB RAM. If I upgrade it would be nice to run a central 6k and two 5k's. I wonder how quiet an AS MBP would be when driving those--maybe with AS, that's handled. But it still doesn't solve the RAM issue.

And, finally, there's procesing speed. Most of my programs are single-threaded, including the ones with the biggest wait times, and the AS desktops provide no benefit over the AS laptops in that regard. But I think they potentially could (and should).

As I mentioned in an earlier post, many of us desktop users were hoping Apple would take advantage of the high TDP's (and constant wall power) available in both the Studio and the MP, to boost the clocks in their M2 desktops (both CPU and GPU) over what's available in the mobile devices. Even if power consumption increases quadratically or cubically with clock speed, the TDP's are currently so low for AS that they have the headroom to improve perfomance for desktop devices. Plus since many programs are single-threaded, they could provide a substantial performance improvment for those, with just a modest increase in TDP, by just allowing a "turbo" boost for a a few cores

Alas, they didn't do that in the M2 Ultra Studio (it's 3.7 GHz, same as the 16" M2 Max). Not sure what they've done with the MP. It's possible that the M2 can't be boosted (it may not be designed to handle the higher voltage). If so, we will need to wait for the M3.

 

[Stevenyo]

It's interesting that you and I have traded places. I used to be a desktop guy, but starting in 2008 I switched to MacBook Pro's because they combined decent power and mobility. I owned 2008, 2011, and 2014 versions. However, while I don't need the high core counts, I did get frustrated with their TDP and RAM limitations, and have thus moved back to a desktop.

I was running my 2014 MBP with three external displays, including a 4k, and that puts a lot of stress on a mobile GPU, forcing it to run at high-fan pretty much continuously, which made the thing noisy. And if my room got too warm I'd need to turn on a button fan I set up behind it. If I didn't, it would essentially shut down due to thermal throttling.

And some of my calculations need more RAM than one can get in a laptop, so it was inconvenient when I enountered them.

I solved both of those problems with a 2019 iMac. It runs three large displays (including it's built-in 5k) quietly, and I was easily able to install 128 GB RAM. If I upgrade it would be nice to run a central 6k and two 5k's. I wonder how quiet an AS MBP would be when driving those--maybe with AS, that's handled. But it still doesn't solve the RAM issue.

And, finally, there's procesing speed. Most of my programs are single-threaded, including the ones with the biggest wait times, and the AS desktops provide no benefit over the AS laptops in that regard. But I think they potentially could (and should).

As I mentioned in an earlier post, many of us desktop users were hoping Apple would take advantage of the high TDP's available in both the Studio and the MP to boost the clocks in their M2 desktops (both CPU and GPU), to improve performance over the mobile devices. Even if power consumption increases quadratically or cubically with clock speed, the TDP's are currently so low for AS that they have the headroom to improve perfomance for desktop devices. Plus since many programs are single-threaded, they could provide a substantial performance improvment for those, with just a modest increase in TDP, by just allowing a "turbo" boost for a a few cores

Alas, they didn't do that in the M2 Ultra Studio (it's 3.7 GHz, same as the 16" M2 Max). Not sure what they've done with the MP. It's possible that the M2 can't be boosted (it may not be designed to handle the higher voltage). If so, we will need to wait for the M3.

2008, 2011, and 2014 MacBook Pros were still machines I always considered secondary to whatever tower or iMac I could get my hands on at the time. Really, the first potentially desktop-replacement MacBook Pro was the 2019 16" but since the M1max was such a step up I don't mind saying that MacBook Pros only became true workstation replacement machines in terms of processing power in 2021.

Same on the PC side, in 2019 I built my Ryzen box (with a super cheap 2700x, knowing I'd get the zen 3 12 core when it came out), and there was no PC laptop in the sub $2k, sub 7lb range that gave performance anywhere close to the desktop. Fast forward a couple years and skyrocketing GPU prices, plus 6nm zen 3 8 core chips running with little throttling in even small chassis, plus high resolution, high refresh rate monitors coming standard, etc. all means suddenly i'd rather spend ~$1000 every couple years on a new laptop than the same money on a Desktop or upgrade parts for it. (which isn't to say cheap gaming laptops can compete with desktops performance wise, just that they're "good enough" that the cost/benefit skews their way. Gaming laptops used to be a joke, bad at gaming and bad at portability. Now I have console level performance in approximately a 14" MacBook Pro sized package all for ~$1k when on sale. Are 4 more cores with more headroom and 2.5x the GPU power Radeon 6700s vs 6800XT big benefits? Absolutely, but so is throwing my gaming machine in my backpack and knowing I can "good enough" anywhere)

Apple's RAM issue is a real one though. M2 supporting 24GB isn't terrible for consumer devices these days, but it should be 16 with a 24GB upcharge. M2 Max going to 96GB in laptops is a nice bump, but really in 2023, there's a good argument that they needed to offer 128GB (a number even a dual CPU 2009 MacPro can hit, as well as almost any 4+ core DDR4 PC system with the right motherboard). Same with at least 256 in the Ultra. Especially if they keep wanting to harp on the bandwidth train, 400/800GPBS for CPU+ GPU is nice, but I hope M3 will get an even wider bus along with faster RAM to push that closer to 600/1200 for the Max/Ultra.

The flipside of course is that PCIe 4.0 x4 NVMe drives transfer data around DDR2 speeds, so while a decade ago everything you needed to access at more than ~50MB/s HDD speed was likely loaded into RAM, now it's only things you need faster than ~5 GB/s. SSDs are the real reason RAM use topped off since around 2010, there are so few workflows or use cases where that extra RAM is needed.

Personally I find my 32GB M1Max to be far more limited by SSD space and the frustration of running out of TB ports when I start attaching drives, docks, card readers/cameras, etc (seriously SD card and HDMI are nice, but I'd take 2 more TB ports even on 3 total busses, or an expansion bay or two so I could have internal CFexpress as well). than by the RAM. For $200 or less you can buy/build a 2TB thunderbolt drive that will sustain ~8 4k60 ProRES streams for multicam editing, or likely a couple 8k60 (I've never used 8k footage, I feel like editing it on a machine without internal NVMe RAID storage would be like trying to work with HDV on a sawtooth G4 tower (been there, tried that. Even with a 1.4Ghz upgrade it was not worth the trouble and quickly leaped to the dual 2.5 G5)

As for the M2Max/ultra clock speeds, I think you're right, it's highly likely the architecture just can't run any faster/higher voltage. Even with unlimited cooling and power, there is just a speed at which a system can't be made stable. Apple designed the M2 series with the 3.5-3. ghz range in mind. Stability wise, an M2ultra at 3.7ghz is likely more like a 13900KS at 6+ ghz already. I wouldn't hold my breath for dramatically faster cores anytime soon, the whole conceit of Apple Silicon is scaling up the iPhone parts, so even Apple's performance cores are going to stay optimized for the most work that can be done in up to 5watts per core for a long time to come.

 

[Minghold]

...Clearly, Apple Silicon Mac is doomed with its performance and workstation. They better bring a real Mac Pro or this is a huge mess forever just like Mac Pro 2013 did.

Never assume that Apple is "failing" -- *always* invert Hanlon's Razor.

Like most of "Big Tech" these days, they are an amalgamated intelligence-gathering entity wearing a constellation of accumulated hardware/software-manufacturers as a skinsuit. What the retail customer wants, and even making a profit off them, is no longer their primary goal, and hasn't been for decades. (Now go be a good peasant, and put your whole life on the cloud.)

 

[theorist9]

The flipside of course is that PCIe 4.0 x4 NVMe drives transfer data around DDR2 speeds, so while a decade ago everything you needed to access at more than ~50MB/s HDD speed was likely loaded into RAM, now it's only things you need faster than ~5 GB/s. SSDs are the real reason RAM use topped off since around 2010, there are so few workflows or use cases where that extra RAM is needed.

I'm assuming my calculations would slow down significantly if I had to switch from RAM to swap, even if I had a current AS machine. With my iMac (gen 4 WD SN850 SSD in a PCIe 3.0 slot), it swaps at 260 MB/s when it exceeds my 128 GB RAM. Considering that's probably a combination of sequential and random tasks, I doubt swap for a current AS machine would be that much faster.

The good news is that the next gen of Macs should switch from LPDDR to LPDDR5x RAM, which offers significantly higher densities (and thus much larger RAM chips). If Apple does as they should, and takes advantage of this, the next-gen AS Ultra chips should offer a max of ~1 TB RAM. And you should be able to get 128+ GB on the Max, thus obviating the need to waste money on the Ultra if all you need is that level of RAM.

Personally I find my 32GB M1Max to be far more limited by SSD space and the frustration of running out of TB ports when I start attaching drives, docks, card readers/cameras, etc (seriously SD card and HDMI are nice, but I'd take 2 more TB ports even on 3 total busses, or an expansion bay or two so I could have internal CFexpress as well). than by the RAM.

Is this during desktop or mobile use? if the former, wouldn't that be addressed by a TB dock? Or is the issue the I/O bandwidth rather than the number of ports?

And I assume that the M1 Max MBP's 8 TB SSD option would provide enough space, but either your space needs changed from when you bought it (the big downside of non-upgradeabilty), or you (reasonably) found the price exorbitant.

As for the M2Max/ultra clock speeds, I think you're right, it's highly likely the architecture just can't run any faster/higher voltage. Even with unlimited cooling and power, there is just a speed at which a system can't be made stable. Apple designed the M2 series with the 3.5-3. ghz range in mind. Stability wise, an M2ultra at 3.7ghz is likely more like a 13900KS at 6+ ghz already. I wouldn't hold my breath for dramatically faster cores anytime soon, the whole conceit of Apple Silicon is scaling up the iPhone parts, so even Apple's performance cores are going to stay optimized for the most work that can be done in up to 5watts per core for a long time to come.

A former chip designer for AMD that used to be active on this site said there could be room to increase clocks on the M3. We shall see....

 

[Stevenyo]

I'm assuming my calculations would slow down significantly if I had to switch from RAM to swap, even if I had a current AS machine. With my iMac (gen 4 WD SN850 SSD in a PCIe 3.0 slot), it swaps at 260 MB/s when it exceeds my 128 GB RAM. Considering that's probably a combination of sequential and random tasks, I doubt swap for a current AS machine would be that much faster. f

Even with only 2 4k60 streams, I regularlry pull about 500MB/s off my external drives consistently for an hour+ straight. Enough to make a 10Gbps USB drive stumble, but not even stress a TB3 enclosure or PCIe 3.0 drive. But of course, that's much more of a sequential workflow than a lot of swap uses. Before TB3, I used multi-drive RAID and could pull more than that 260MB/s off just 4 disks in RAID 5, so I'm not sure if you're actually drive limited in your workflow.

The good news is that the next gen of Macs should switch from LPDDR to LPDDR5x RAM, which offers significantly higher densities (and thus much larger RAM chips). If Apple does as they should, and takes advantage of this, the next-gen AS Ultra chips should offer a max of ~1 TB RAM. And you should be able to get 128+ GB on the Max, thus obviating the need to waste money on the Ultra if all you need is that level of RAM.

Is this during desktop or mobile use? if the former, wouldn't that be addressed by a TB dock? Or is the issue the I/O bandwidth rather than the number of ports?

Desktop and mobile kinda blend together for me. My dock is only used in one place, but it only offers one extra TB port, and that's usually used for a display rather than anything else.

Bandwidth is rarely my limitation, but I regularly need to offload from 2 non-SD camera media cards to a drive for editing/backup (ideally taking from 2 places to 2 other places at the same time, but I often settle for offload then backup due to only 3 TB ports. Not really a huge headache, just frustrating since I rarely use the SD card or HDMI port.

And I assume that the M1 Max MBP's 8 TB SSD option would provide enough space, but either your space needs changed from when you bought it (the big downside of non-upgradeabilty), or you (reasonably) found the price exorbitant.

I have 1TB in my MBP. No way in heck I'm paying Apple for internal storage. ProRES 422 HQ 4k60 uses up almost 1TB/hr. Add in usually shooting shows or events on at least 2 cameras, plus Final Cut's insane "cache" files and a 1 hour event needs 4TB that sustains more than 10Gb/s. Not expensive or hard externally, but I don't want to waste the money with apple and then put like 4TBW/week on my soldered drive.

A former chip designer for AMD that used to be active on this site said there could be room to increase clocks on the M3. We shall see....

 

[Boil]

The rumor was for an ASi Mac Pro with a M2 Ultra SoC with the CPU cores clocked at 4.2GHz, don't know what the GPU clocks would be...?

 

[leman]

Are Intel's Sapphire Rapids Xeons delivering a big performance boost over the 2019 Xeons? I got the impression... they didn't, but there aren't a ton of benchmarks out there for the workstation-flavoured chips.

Yes, it's a very substantial boost. They doubled the number of cores and significantly increased the compute per core.

Fundamentally, and I say this after buying a 2010 Mac Pro over the weekend and looking at how it benchmarks against newer things, large RAM capacity, PCI-E flexibility, and to some extent many-core performance are basically what defines a workstation, certainly not single-core performance or even multi-core performances.

What defines a workstation depends on your needs. A decade or so ago, workstations were all about maximising total compute, often using multiprocessor systems. That was just something consumer hardware couldn't do. Recently attention has shifted more on the GPU side. Historically, Xeons sacrificed single core performance for multi-core throughput, which could hamper some hybrid workloads. A few years ago Apple has introduced a workstation model that retained high single-core performance — this was the Intel Xeon W series which first were Apple exclusive and now became a fairly successful Xeon family.

 

[daneoni]

3 certainties in life. Death, taxes, and unhappy Mac Pro customers every generation.

 

[ectoplasmosis]

Turns out that ALL of the Mac Pro 2023's PCIe slots are piped through a PCIe switch with a single 16-lane connection to the SoC: https://social.treehouse.systems/@marcan/110493886026958843

Very disappointing.

 

[leman]

Turns out that ALL of the Mac Pro 2023's PCIe slots are piped through a PCIe switch with a single 16-lane connection to the SoC: https://social.treehouse.systems/@marcan/110493886026958843

Very disappointing.

Yeah, one can say that loudly. It will still work out fine for the intended use, but one can’t really lose the feeling that at least some aspects of this Mac Pro are a slap in a face for prospective users…

Anyway, all of this suggests that they had bigger plans but either had to scrape or significantly delay them, and this machine is an emergency effort to bring something out. I still believe that they wanted to ship a 3nm product, maybe even last year, but couldn’t do so because of delays. Or maybe I am being delusional and it’s just Apple counting beans...

 

[Numa_Numa_eh]

Yeah, one can say that loudly. It will still work out fine for the intended use, but one can’t really lose the feeling that at least some aspects of this Mac Pro are a slap in a face for prospective users…

Anyway, all of this suggests that they had bigger plans but either had to scrape or significantly delay them, and this machine is an emergency effort to bring something out. I still believe that they wanted to ship a 3nm product, maybe even last year, but couldn’t do so because of delays. Or maybe I am being delusional and it’s just Apple counting beans...

I wonder how much bandwidth is required for the pci cards people will use (audio, networking etc) given no gpus can be added?

 

[joevt]

Turns out that ALL of the Mac Pro 2023's PCIe slots are piped through a PCIe switch with a single 16-lane connection to the SoC: https://social.treehouse.systems/@marcan/110493886026958843

So it has 24 gen 4 PCIe lanes. That's 75% the bandwidth of the 64 gen 3 PCIe lanes of the 2019 Mac Pro? Not too bad.

 

[falainber]

So it has 24 gen 4 PCIe lanes. That's 75% the bandwidth of the 64 gen 3 PCIe lanes of the 2019 Mac Pro? Not too bad.

25% degradation in 4 years? Perhaps Mac Pro will survive for 4..8 more years?

 

[k27]

Hector Martin (@marcan@treehouse.systems)

Yeah okay so they are totally cheating. They have two big PCIe switches and are hanging everything off of them. From the device tree: # 8x Gen4 lanes on die 0, switched: - x1 to USB internal - x2 to SATA internal - x4 to I/O card (USB, probably single lane connected) - x8 to slot6 - x1 to...

social.treehouse.systems

 

[maikerukun]

View attachment 2213755
If you can remember back in 2013, a trash can Mac announced and a lot of Mac Pro users were so furious about how Apple create a workstation. As a result, a lot of pro users left Mac system and Apple left the pro markets for several year until iMac Pro or Mac Pro 2019 release. This time, it's happening again.

View attachment 2213756
Mac Pro 2023 lack so many things and it does NOT promise anything. When I saw it when Mac Pro 2023 announced, it just reminded Mac Pro 2013.

CPU
Up to 24 core is not good. Mac Pro 2019 already supported up to 28 cores and currently, it can go as high as possible up to 128 cores for workstation computers. Beside, there are reasons to support dual CPU in order to support more PCIe lanes.

GPU
Apple GPU's performance is still fundamentally bad. Yes, it's bad. Both M1,2 series aren't even close to RTX 30 series which is based on Samsung 8nm, not TSMC 5nm. RTX 40 series are TSMC 5nm based and it's clearly not comparable. Since M1 Max/Ultra, Apple stopped comparing their own chips to Nvidia instead of Intel Mac because they know their GPU performance is just bad. Beside, the power consumption is too limited as well. Since M1 Ultra has similar performance to RTX 3060, I highly doubt that M2 Ultra is close to RTX 3090. RTX 4090's bandwidth is already way beyond M2 Ultra's 800GB/s and those workstation GPUs are way beyond that. Whoever defends Apple GPU's performacne, you never used RTX 40 series and workstation series. This is why Nvidia is dominating GPU performance for a while.

On the other hand, Mac Pro 2023 supports only one M2 Ultra which is a joke. Mac Pro 2019 supports up to 4x highend workstation GPU and others can go beyond that. Apple did not make M2 Extreme or something better instead of re-using M2 Ultra for so called workstation. No, M3,4 can not save Mac Pro as long as it's SOC and not expandible.

View attachment 2214594
Apple GPU is powerful? Since M1 Ultra cant even close to RTX 3090 but 3060, which is a hard fact, I dont think M2 Ultra is still close to RTX 3090.

RAM
Nope, you cant even upgrade it and expand beyond 192GB of RAM which is WAY less than what Mac Pro 2019 can provide which is 1.5TB of RAM. Unified memory is not a magic and the RAM size still matters. Yes, that's a lot of VRAM but the truth is, Apple GPU itself has a poor performance, the bandwidth speed is way slower than both highend and workstation GPU, and PC can also expand VRAM with more GPU as well. This is a huge limitation and disappointment since Mac Pro users ever since heard the first rumor that Mac Pro will not have upgradable RAM.

Price
Really? $1000 more for less features? Apple justified their transition from Intel to Apple Silicon by reducing the price dramatically as they dont need to purchase both CPU and GPU components which can save a lot of money but in reality, Apple increased way more than before. For example, upgrading RAM is extremly expesnive and yet the memory chip itself is really cheaper than you think. Beside, Apple Silicon uses less memory chips than normal RAM. Dont forget that Mac Pro series started from around $3000 price range and Apple increased the price up to $7000.

PCIe slots
You cant even use either AMD or Nvidia GPU. PCIe slot is only gen 4 while others are using gen 5 and MPX module is gone! What are we suppose to do with PCIe slots?

What a mess
Mac Pro 2023 proves that Apple can NOT make a powerful AS chip for Mac Pro, they seriosuly dont care about Mac Pro and Pro markets, and they just ruined it. It just reminds me Mac Pro back in 2013 when Apple proudly announced it and it turned out it was a failure. At this point, because of Mac Pro 2023, the 3D and AI software which requires high GPU performance will either not support Mac or ditch Mac system. Dont say this is not for you, that's the worst excuse to make and we know what Apple did with Mac Pro 2013. Quite a lot of youtubers already disappointed about Apple's move toward Mac Pro 2023 so I'm not the only one complaining about this.

Clearly, Apple Silicon Mac is doomed with its performance and workstation. They better bring a real Mac Pro or this is a huge mess forever just like Mac Pro 2013 did.

I don't know man. I of all people am EXTREMELY DISAPPOINTED. I have a 2019 Mac Pro with well over 320+ gigs of ram, 2 w6800x DUO's, and 8tb of storage. it's on a 28 core CPU and sadly, it is STILL the most powerful Mac that Apple makes...and sadly it's the most powerful BY FAR.

The M2 Ultra Mac Studio is actually FAR SUPERIOR in every way...EXCEPT, GPU. 2 w6800x Duo's keeps my 2019 Mac Pro clocking in on GPU performance in Octane and Redshift at about 33% faster than 2 RTX 3090's so it's a fast system...but it can't keep up with what's happening with the RTX 4090...not even close. So patiently, I waited, because I was just sure that those GPU cards were gonna be announced...and they weren't, I was certain they wouldn't show off a less powerful Mac Pro than the previous generation...and they did. And I was furious.

Furious enough that I ordered a Puget system with 2 RTX 4090's and a 64 core threadripper. EVERY OTHER PRODUCT IN MY HOUSE IS APPLE. I mean all of it, watches, phones, laptops, tablets, desktops, monitors...everything, there's a good $200k worth of Apple stuff here...and they forced my hand. They forced me into a corner...they did that to their power users that needed a professional Mac Pro. Users that dropped $40 - $60k on their previous generation and who were prepared to do it again...they spit directly in our faces.

And yet...I still hope. I still dream. I still wonder if this Mac Pro that they showed off was just a placeholder. A stop gap of sorts, very specifically meant to complete the transition...but honestly just there to keep face while they work.

Because quite frankly, we know those GPU cards exist...and they didn't leave that ridiculously overpowered power source in the machine for nothing. Apple doesn't do anything without reason. And I have to assume that power brick is meant to power something that we haven't seen yet. Something they know that we don't.

Or maybe the Apple fanatic in me keeps dreaming...

 

[Numa_Numa_eh]

I don't know man. I of all people am EXTREMELY DISAPPOINTED. I have a 2019 Mac Pro with well over 320+ gigs of ram, 2 w6800x DUO's, and 8tb of storage. it's on a 28 core CPU and sadly, it is STILL the most powerful Mac that Apple makes...and sadly it's the most powerful BY FAR.

The M2 Ultra Mac Studio is actually FAR SUPERIOR in every way...EXCEPT, GPU. 2 w6800x Duo's keeps my 2019 Mac Pro clocking in on GPU performance in Octane and Redshift at about 33% faster than 2 RTX 3090's so it's a fast system...but it can't keep up with what's happening with the RTX 4090...not even close. So patiently, I waited, because I was just sure that those GPU cards were gonna be announced...and they weren't, I was certain they wouldn't show off a less powerful Mac Pro than the previous generation...and they did. And I was furious.

Furious enough that I ordered a Puget system with 2 RTX 4090's and a 64 core threadripper. EVERY OTHER PRODUCT IN MY HOUSE IS APPLE. I mean all of it, watches, phones, laptops, tablets, desktops, monitors...everything, there's a good $200k worth of Apple stuff here...and they forced my hand. They forced me into a corner...they did that to their power users that needed a professional Mac Pro. Users that dropped $40 - $60k on their previous generation and who were prepared to do it again...they spit directly in our faces.

And yet...I still hope. I still dream. I still wonder if this Mac Pro that they showed off was just a placeholder. A stop gap of sorts, very specifically meant to complete the transition...but honestly just there to keep face while they work.

Because quite frankly, we know those GPU cards exist...and they didn't leave that ridiculously overpowered power source in the machine for nothing. Apple doesn't do anything without reason. And I have to assume that power brick is meant to power something that we haven't seen yet. Something they know that we don't.

Or maybe the Apple fanatic in me keeps dreaming...

Thoughts and prayers.

 

[deconstruct60]

Turns out that ALL of the Mac Pro 2023's PCIe slots are piped through a PCIe switch with a single 16-lane connection to the SoC: https://social.treehouse.systems/@marcan/110493886026958843

Very disappointing.

Why is this surprisingly disappointing ? The Mac Pro 2019 is doing the basically the same thing for six of its slots

"...
So basically 5 of the slots including both x16 ones are bottlenecked into a single x16 gen4 channel. Boo. ..."

Same thing for slots 4 and 6 (two of the x16s ) on the MP 2019.

".. I said the Mac Pro was going to use a new chip with a lot more PCIe. ..."

It is more than what the M1 Max had. The backhaul on the MP 2019 was two x16 PCI-e v3 . Which is equivalent to x16 PCIe v4. So the 'other' x8 PCI-e v4 is all uplift over the last generation backhaul.

If there are exact twin chips, The storage stuff could be elsewhere and Apple is just going bandwidth reservation QoS by turning off x8 of that PCI-e controllers. The SSD controller doesn't use standard PCI-e to talk to modules ( a subset. ).

Apple went through internal mesh gyrations to uncork the GPU cores for M2 generations. There wasn't likely going to be copious extra bandwidth and QoS latency available for bucketloads of x16 PCI-e v4 controllers. Apple was trying to crank up CPU clusters , NPU clusters , video en/decode clusters ... they had lots of hogs at the trough .

Few folks were screaming "Boo" at the MP back in 2019. Not much of a difference now. (old Slot 1 and 3 are gone along with the three other 8-pin AUX power supplies. and (at this point) 'solution in search of a problem " MPX connectors. )

 

[joevt]

The backhaul on the MP 2019 was two x16 PCI-e v3

It was four x16 PCIe gen 3 = 64 lanes.
32 lanes when went to slot 1 and 3.
16 lanes went to one upstream port of the PCIe bridge and 16 lanes went to a second upstream port of the PCIe bridge.
See page 11 of https://www.apple.com/ua/mac-pro/pdf/Mac_Pro_White_Paper_Aug_2021.pdf

 

[deconstruct60]

It was four x16 PCIe gen 3 = 64 lanes.
32 lanes when went to slot 1 and 3.
16 lanes went to one upstream port of the PCIe bridge and 16 lanes went to a second upstream port of the PCIe bridge.
See page 11 of https://www.apple.com/ua/mac-pro/pdf/Mac_Pro_White_Paper_Aug_2021.pdf

Through the PCI-s switch that is being 'boo'ed' here it was only two x16 PCI-e v3. There is a large switch on page 11 there in the MP 2019. There is a large switch now. So that part it is up.

Slots 1 and 3 have basically disappeared here. But if there is next to nothing to power something in slot 3 it isn't really going to be a big loss. And the 'extra' x8 PCI-e v4 covers that 3rd x16 PCI-e v3. Really left with the nominal MPX model in slot 1. The main display GPU, which is no inside the primary SoC. If want to leap into holistic whole system analysis the bandwidth to the SoC GPU is WAY higher than than even x16 PCI-e v5. That is why is works far more efficiently. That is more so crying about "you took away the vastly slower x16 PCI-e v3 solution". Go there if you wish, but it is basically a "Form over function" position. The backhaul of the internal mesh inside of a Max/Utra completely crushes x16 PCI-e v3 (even the whole x64 PCI-e v3).

I was purposely leaving 'old' slot 1 out of the backhaul discussion because the uplift that Apple did there just completely skews the math and the issue.

 

[theorist9]

Apple creating a separate M2-based die for just the MacPro CPU, as well as all the other hardware and software infrastructure to support DIMMs and GPUs, would likely cost Billions, costs that could only be recouped by using those parts in millions of systems.

Returning to the Mac Pro: This would be the case if a, say, "M3 Extreme" (4x Max) required an entirely separate chip, but it doesn't. I asked a former AMD chip designer (cmaier) who posts on another site about this, and he said if Apple did this, they would probably join two M3 Ultras using something akin to the interposer that is currently used to join two M2 Max's into an M2 Ultra, and the development costs for that would be relatively modest (they'd still be using the same M3 Max chip in the Max, Ultra, and Extreme SoC's).

He said there would be strips of the Max that would be used only for connecting to other Max's, but that would only slightly increase the chip size, and thus the cost, over a Max that didn't have these.

What could be high would be the manufacturing costs (e.g., if the yield for the interposer were low), but these would be reflected in the selling price.

He added using the Max in this way would necessitate some performance upgrades not required if it were used alone (high off-chip bandwidth); but while not required, these would benefit all Max chips.

 

[Longplays]

Returning to the Mac Pro: This would be the case if a, say, "M3 Extreme" (4x Max) required an entirely separate chip, but it doesn't. I asked a former AMD chip designer (cmaier) who posts on another site about this, and he said if Apple did this, they would probably join two M3 Ultras using the same type of interposer that is currently used to join two M2 Max's into an M2 Ultra, and the development costs for that would be relatively modest (they'd still be using the same M3 Max chip in the Max, Ultra, and Extreme SoC's).

He said there would be edge strips of the Max that would be used only for connecting to the interposer, but that would only slightly increase the chip size, and thus the cost, over a Max that didn't have these.

What could be high would be the manufacturing costs (e.g., if the yield for the interposer were low), but these would be reflected in the selling price.

He added using the Max in this way would necessitate some performance upgrades not required if it were used alone (high off-chip bandwidth); but while not required, these would benefit all Max chips.

M3 Ultra/Extreme would probably be out Q1 2025. Enough time for Apple to figure it out. The Extreme would likely have more than 384GB unified memory, Thunderbolt 5 80Gb/s & allow for PCIe 5.0 slots.