I have a feeling that Apple will introduce an M-Series Mac Pro, but keep the Intel version around.
If they were to keep the Intel Mac Pro around, then I don't see the point of an Apple silicon Mac Pro that lacks upgradeable RAM. Might as well keep the Mac Pro Intel (or AMD) only and use the Mac Studio as the highest-end Apple Silicon machine.
Apple Silicon offers better performance-per-dollar than Intel workstation CPUs. The following shows the $4k Ultra Mac Studio is roughly the same performance as the highest CPU $13k Xeon Mac Pro:
Following your link, the $4k Ultra Mac Studio scores 41,033. The AMD 7950x scores 63,535 and costs $569 (for the CPU). You can get a premium 7950x linux system for about $2800 (https://system76.com/desktops/thelio-mira-r3/configure)
So... it's not clear to me that Apple silicon offers better price/performance using your metric.
having said that, I think there are scenarios in which Apple Silicon might be competitive in price/performance, and then when you take into account heat, noise, and various advantages of Macs (for people who like Macs), the Mac makes sense. That's where I am -- my workloads do better on Apple Silicon than what your link would imply and I care about the other stuff that makes Macs nice.
Why does Apple have to keep bending over its customers?
I don't think Apple is doing this because they want to. I'm thinking it's either they release what they know they can safely release, or delay it again. I wouldn't expect any new Apple Silicon MacPro to offer up to 1.5TB of RAM. That's an insane amount of RAM. I wouldn't expect any Apple Silicon MacPro to offer anywhere near that for a a couple of years from now. I'm sure the new MacPro will be pretty great though even with the lack of upgradable RAM.
It also goes back to the people who buy a high end MacPro. They buy what they can afford, use it until it's not useful to them anymore and then get something newer. They just wanna take something out of the box, install their software and get back to work. They don't wanna spend time screwing around with buying and upgrading RAM and other things. Where I work we do this too. We have a specific spec of Mac and PC's we buy. We use them until they're EOL (3yrs) and then we swap them out. We don't screw around with RAM, extra HD, upgrading video cards, etc.
I have a feeling that Apple will introduce an M-Series Mac Pro, but keep the Intel version around.
Apple historically doesn’t really do the hang on to the past thing.
Also consider that getting rid of Intel would vastly simplify software development efforts.
True, but there’s Mac Pro users that need massive amounts of RAM. Amounts that Apple Silicon just doesn’t support.
That said, I’m not a computer engineer so it may be a crazy idea; Would it be feasible to use two tiers of RAM? The high speed RAM built into the chip, and then a TB or more of comparatively slow conventional RAM in sticks on the MB like it has now? It would have to keep track of what needed to be kept in the extra high speed on chip space, and what could be parked on the sticks. It would be like virtual RAM does now but not to the SSD. Not sure how feasible this is but it was a crazy idea that just crossed my mind.
That’s interesting throughput, but SSD (unless using Optane) has a rather finite number of write cycles, so that’s bonkers for a solution. I’ve purposely provisioned all my machines with enough RAM keeping that in proper perspective and context, because I don’t replace machines super often. Swap files in a normal usage scenario are one thing, but treating SSD with finite write cycles as RAM is a whole other story.
I have a feeling that Apple will introduce an M-Series Mac Pro, but keep the Intel version around.
Apple historically doesn’t really do the hang on to the past thing.
Also consider that getting rid of Intel would vastly simplify software development efforts.
True, but there’s Mac Pro users that need massive amounts of RAM. Amounts that Apple Silicon just doesn’t support.
That said, I’m not a computer engineer so it may be a crazy idea; Would it be feasible to use two tiers of RAM? The high speed RAM built into the chip, and then a TB or more of comparatively slow conventional RAM in sticks on the MB like it has now? It would have to keep track of what needed to be kept in the extra high speed on chip space, and what could be parked on the sticks. It would be like virtual RAM does now but not to the SSD. Not sure how feasible this is but it was a crazy idea that just crossed my mind.
That’s interesting throughput, but SSD (unless using Optane) has a rather finite number of write cycles, so that’s bonkers for a solution. I’ve purposely provisioned all my machines with enough RAM keeping that in proper perspective and context, because I don’t replace machines super often. Swap files in a normal usage scenario are one thing, but treating SSD with finite write cycles as RAM is a whole other story.
Flash write durability hasn’t been a serious limit on SSD lifespan in close to two decades. Controller firmware bugs take out multiple orders of magnitude more drives than the flash wearing out does.
This is literally talking about using some high-performance SSDs as dedicated swap devices. No more, no less.
The headache would be random access, just like with swap. NVMe SSDs can get DDR2 levels of data throughput, but much worse latency for random operations. If you mix reads and writes, performance gets ridiculously bad. Give a 70/30 R/W workload to a flash-based SSD and its performance drops to around 15% of high-queue-depth peak read or peak write.
Why does Apple have to keep bending over its customers?
I don't think Apple is doing this because they want to. I'm thinking it's either they release what they know they can safely release, or delay it again. I wouldn't expect any new Apple Silicon MacPro to offer up to 1.5TB of RAM. That's an insane amount of RAM. I wouldn't expect any Apple Silicon MacPro to offer anywhere near that for a a couple of years from now. I'm sure the new MacPro will be pretty great though even with the lack of upgradable RAM.
It isn’t an insane amount of RAM if you are doing 3D engineering and physics problems. I actually tried doing a simulation of a uniquely shaped optical waveguide a few years ago and the software (Lumerical) required 1.5TB of RAM or no go. So I was never able to run the simulation. I’m actually currently looking to purchase a Z8 since it can access up to 3TB.
Apple’s inability so far to make different Mac’s in different form factors (full tower, half tower, rack) along with the classic iMac all in one with different screen sizes 27”, and 32” is ridiculous up until this point, all should be on the market now, absolute speed isn’t the problem. It’s the lack of different form factors to choose from, a laptop is fine for many users but not all.
The 32” Pro Display XDR monitor enclosure with the M1 Mac Studio series computer built in should be a no-brainer, the all in one form factor is superior to the (Leader/Max Headroom style enclosure computer).
I have a feeling that Apple will introduce an M-Series Mac Pro, but keep the Intel version around.
Apple historically doesn’t really do the hang on to the past thing.
Also consider that getting rid of Intel would vastly simplify software development efforts.
True, but there’s Mac Pro users that need massive amounts of RAM. Amounts that Apple Silicon just doesn’t support.
That said, I’m not a computer engineer so it may be a crazy idea; Would it be feasible to use two tiers of RAM? The high speed RAM built into the chip, and then a TB or more of comparatively slow conventional RAM in sticks on the MB like it has now? It would have to keep track of what needed to be kept in the extra high speed on chip space, and what could be parked on the sticks. It would be like virtual RAM does now but not to the SSD. Not sure how feasible this is but it was a crazy idea that just crossed my mind.
That’s interesting throughput, but SSD (unless using Optane) has a rather finite number of write cycles, so that’s bonkers for a solution. I’ve purposely provisioned all my machines with enough RAM keeping that in proper perspective and context, because I don’t replace machines super often. Swap files in a normal usage scenario are one thing, but treating SSD with finite write cycles as RAM is a whole other story.
Flash write durability hasn’t been a serious limit on SSD lifespan in close to two decades. Controller firmware bugs take out multiple orders of magnitude more drives than the flash wearing out does.
This is literally talking about using some high-performance SSDs as dedicated swap devices. No more, no less.
The headache would be random access, just like with swap. NVMe SSDs can get DDR2 levels of data throughput, but much worse latency for random operations. If you mix reads and writes, performance gets ridiculously bad. Give a 70/30 R/W workload to a flash-based SSD and its performance drops to around 15% of high-queue-depth peak read or peak write.
The only time I recall being in a situation where I needed more than 256GB of RAM, I doubt that a lot of small random writes to swap were needed. Instead, it really was a case of swapping big chunks, and I could see the SSD read/write statistics in glances go through the roof (Linux threadripper system).
To need more than 1TB of RAM is way outside my experience. For folks who have such experience— does your work involve a lot of random reads and writes to that 1TB or do you think you’re really swapping in and out multi-GB chunks?
I've been digging in to more of the rumors around how Apple plans to fill in some of the gaps for pro users.
The consensus opinion seems to be that Apple will add AMD GPU support to Apple Silicon for some workflows (AI, Scientific Users, Mining), but may add specialty GPUs of their own design like they did with ProRes acceleration years ago.
There are rumors that Imagination Technology is helping Apple to design a custom GPU with raytracing support. It could be Imagination Technologies is helping Apple design a chip dedicated to raytracing for pro users in addition to integrating with embedded GPUs. It would make a lot of sense if it results in a significant speed increase over an nVidia card. Using multiple GPUs is complex and Apple was burned by that in the trash can Mac Pro that could only use half of it's GPU for most apps, so it would make a lot of senses that any discrete GPU could be targeted for just raytracing acceleration and the built-in GPU would be used for most other tasks.
Apple was rumored to be building a discrete GPU codenamed Lifuka, but it was likely scrapped since it wasn't good enough at any particular task. They could either make a card good at AI, numerical processing, or raytracing but not everything. The Imagination partnership seems to have started around the time it was scrapped.
I have a feeling that Apple will introduce an M-Series Mac Pro, but keep the Intel version around.
Apple historically doesn’t really do the hang on to the past thing.
Also consider that getting rid of Intel would vastly simplify software development efforts.
True, but there’s Mac Pro users that need massive amounts of RAM. Amounts that Apple Silicon just doesn’t support.
That said, I’m not a computer engineer so it may be a crazy idea; Would it be feasible to use two tiers of RAM? The high speed RAM built into the chip, and then a TB or more of comparatively slow conventional RAM in sticks on the MB like it has now? It would have to keep track of what needed to be kept in the extra high speed on chip space, and what could be parked on the sticks. It would be like virtual RAM does now but not to the SSD. Not sure how feasible this is but it was a crazy idea that just crossed my mind.
That’s interesting throughput, but SSD (unless using Optane) has a rather finite number of write cycles, so that’s bonkers for a solution. I’ve purposely provisioned all my machines with enough RAM keeping that in proper perspective and context, because I don’t replace machines super often. Swap files in a normal usage scenario are one thing, but treating SSD with finite write cycles as RAM is a whole other story.
Flash write durability hasn’t been a serious limit on SSD lifespan in close to two decades. Controller firmware bugs take out multiple orders of magnitude more drives than the flash wearing out does.
This is literally talking about using some high-performance SSDs as dedicated swap devices. No more, no less.
The headache would be random access, just like with swap. NVMe SSDs can get DDR2 levels of data throughput, but much worse latency for random operations. If you mix reads and writes, performance gets ridiculously bad. Give a 70/30 R/W workload to a flash-based SSD and its performance drops to around 15% of high-queue-depth peak read or peak write.
The only time I recall being in a situation where I needed more than 256GB of RAM, I doubt that a lot of small random writes to swap were needed. Instead, it really was a case of swapping big chunks, and I could see the SSD read/write statistics in glances go through the roof (Linux threadripper system).
To need more than 1TB of RAM is way outside my experience. For folks who have such experience— does your work involve a lot of random reads and writes to that 1TB or do you think you’re really swapping in and out multi-GB chunks?
I mostly need lots of RAM for running lots of virtual machines. When you have a parent scheduler swapping stuff out, and multiple guest schedulers working with what they think is RAM and swapping stuff out on their own, you can wind up with extremely unpredictable memory access latencies in software in the guests.
The consensus opinion seems to be that Apple will add AMD GPU support to Apple Silicon since they need some sort of GPU expansion. I'm having some trouble believing that though. If anything, at this point a shift back to nVidia would make more sense because those cards are more preferred by scientific and high-end rendering users...
Not happening. Nvidia won't give anybody else sufficient documentation and access to write drivers for their GPUs, and Apple won't give anybody else the ability to write kernel code for their platforms. This is what caused the split in the first place. Apple wants control of all of the code which can cause a kernel panic, because the users who see one will always blame them. Remember that presentation where they said something like 90% of all Safari crashes were caused by "plugins" and everybody knew they meant "Flash"? This is that, but for kernel code.
Apple’s inability so far to make different Mac’s in different form factors (full tower, half tower, rack) along with the classic iMac all in one with different screen sizes 27”, and 32” is ridiculous up until this point, all should be on the market now, absolute speed isn’t the problem. It’s the lack of different form factors to choose from, a laptop is fine for many users but not all.
The 32” Pro Display XDR monitor enclosure with the M1 Mac Studio series computer built in should be a no-brainer, the all in one form factor is superior to the (Leader/Max Headroom style enclosure computer).
You misunderstand. There is no inability from Apple to make towers and rack computers. It’s simply a choice. They got out of the rack server business long ago, and also don’t make Dell-style towers. Laptops, iMac AIO, the Pro, and the mini/Studio. That’s plenty for them.
Apple’s inability so far to make different Mac’s in different form factors (full tower, half tower, rack) along with the classic iMac all in one with different screen sizes 27”, and 32” is ridiculous up until this point, all should be on the market now, absolute speed isn’t the problem. It’s the lack of different form factors to choose from, a laptop is fine for many users but not all.
The 32” Pro Display XDR monitor enclosure with the M1 Mac Studio series computer built in should be a no-brainer, the all in one form factor is superior to the (Leader/Max Headroom style enclosure computer).
You misunderstand. There is no inability from Apple to make towers and rack computers. It’s simply a choice. They got out of the rack server business long ago, and also don’t make Dell-style towers. Laptops, iMac AIO, the Pro, and the mini/Studio. That’s plenty for them.
The inability is defined as a lack of will, and bad decision making similar to Intel saying no to building cpu’s for the iPhone when they had the chance, it’s not that Apple can’t do it, it’s just a sign of the rot and lack of focus two steps forward one step back within Apple.
Another example would be Apple deciding not to make curated monitors for a period of time (many on this very site thought it was a good idea at the time), turns out that none of the third parties could design a monitor that had color accuracy, correct scaling or that wasn’t made out of cheap plastic, the choices were so bad that Apple had to get back into designing competent and curated monitors.
Apple will end up getting back into making (some new version of, AirPort Express) routers for many of the same reasons lack of good third-party support for Apple devices, two steps forward, one step back.
Apple’s inability so far to make different Mac’s in different form factors (full tower, half tower, rack) along with the classic iMac all in one with different screen sizes 27”, and 32” is ridiculous up until this point, all should be on the market now, absolute speed isn’t the problem. It’s the lack of different form factors to choose from, a laptop is fine for many users but not all.
The 32” Pro Display XDR monitor enclosure with the M1 Mac Studio series computer built in should be a no-brainer, the all in one form factor is superior to the (Leader/Max Headroom style enclosure computer).
You misunderstand. There is no inability from Apple to make towers and rack computers. It’s simply a choice. They got out of the rack server business long ago, and also don’t make Dell-style towers. Laptops, iMac AIO, the Pro, and the mini/Studio. That’s plenty for them.
The inability is defined as a lack of will, and bad decision making similar to Intel saying no to building cpu’s for the iPhone when they had the chance, it’s not that Apple can’t do it, it’s just a sign of the rot and lack of focus two steps forward one step back within Apple.
Apple would surely say, and many would agree, that those decisions are focus decisions, or, they are exercising their "no" decisions. They are allowed to make mistakes, like not having a branded monitor for awhile, or only focusing on consumer facing or desk facing hardware. Arguably, it takes them too long to course correct now. The ship is too big.
Why do people think the only coprocessors Apple will ever build into the M series have to do with image processing? Slap in a Blender designed GPU ray tracing coprocessor, toss on 96 Gb of shared RAM, launch the 3nm process and WOW!
With an M series Max or Ultra, the Mac Pro has no reason to exist.
It’s the most powerful and MOST EXPANDABLE Mac. That’s it’s DNA. we have the Mac Studio now. So the Max Pro needs to hit way harder, especially when you have (admittedly power hungry) PC options that no Mac can touch on the desktop side.
The Mac Pro is the Mac daddy that slaps all the other wannabes around and puts them in their place.
So either:
1) Apple has been developing an X series of chips that are dedicated desktop SOCs that are freed up to gulp down power as needed and have the GPU updates missing from M2. This would solve the lack of horsepower from putting M series into what is supposed to be the most beastly computer on the planet. Expandability remains a question.
Or
2) The Mac Pro will use Ultra chips, but have a logic board “fabric” that enables slotted interconnect expsndability. Sort of Like an Xsan within one box, but much more efficient. This would solve the horsepower embarrassment as well as the expandability problem.
Or
3) apple launches X series dedicated desktop SOCs AND ties them together with a logic board “fabric” and comes out with the most advanced system on the planet. This would be the “above and beyond” machine that Apple has been known for selling nice the G5 with the exception of the trash can Mac Pro.
Otherwise, Apple could have come out with the Mac Pro last year and forgone the Studio. No need for two desktops with the same specs. So I think a lot of the recent “Mac Pro doom and gloom” rumors are bologne. For Apple to create so much pent up demand only to also everyone with the letdown of the century by releasing a lame duck Mac Pro would seriously tarnish the Apple brand. The Mac Pro is Apple’s performance halo and adds shine to all the desktops, by virtue of its existence - at least it did when Apple actually invested in the thing. It makes sense that they pulled back when they knew they’d be making their own SOCs. But not after.
Now is the time to show the world what the pinnacle of desktop computing can be. Not what easy street cruise control hardware development looks like. If the Mac Pro isn’t ready to wow the world, just wait go back to the drawing board if you have to. But whatever you do, don’t embarass yourselves and disrespect your fans snd customers by releasing a Mac Studio in a different form factor.
And I don’t think Apple will do that. The Mac Pro will be special again.
That said, I’m not a computer engineer so it may be a crazy idea; Would it be feasible to use two tiers of RAM? The high speed RAM built into the chip, and then a TB or more of comparatively slow conventional RAM in sticks on the MB like it has now? It would have to keep track of what needed to be kept in the extra high speed on chip space, and what could be parked on the sticks. It would be like virtual RAM does now but not to the SSD. Not sure how feasible this is but it was a crazy idea that just crossed my mind.
That’s interesting throughput, but SSD (unless using Optane) has a rather finite number of write cycles, so that’s bonkers for a solution. I’ve purposely provisioned all my machines with enough RAM keeping that in proper perspective and context, because I don’t replace machines super often. Swap files in a normal usage scenario are one thing, but treating SSD with finite write cycles as RAM is a whole other story.
If fast virtual memory swapping over PCIe5 is what you want, why use flash at all? Just put a DRAM controller and a lot of DIMM slots on a PCIe card, and make it look like a swap "drive". These things have existed for ages. This avoids the flash wear problem and is full PCIe bandwidth. With 96-192GB of fast in-package memory in the SoC, all the transfers across PCIe are page-sized bursts so you get maximum PCIe bandwidth. The 96-192GB fast memory is an enormous working set (dwarfing the on-chip caches), which allows the PCIe transfer bottleneck (v4 x16 is 31.5GB/s and v5 is 63GB/s) to be greatly amortized. Compression is even used sometimes to squeeze more data across the bus.
I've been saying for a few years that Apple could do this to deliver expandable memory to those who need it (any how many really need it?), either by building it into the motherboard or by offering/supporting an actual PCIe card. RAM is expensive, but if users really need it, they'll pay.
Comments
The AMD 7950x scores 63,535 and costs $569 (for the CPU).
You can get a premium 7950x linux system for about $2800 (https://system76.com/desktops/thelio-mira-r3/configure)
So... it's not clear to me that Apple silicon offers better price/performance using your metric.
having said that, I think there are scenarios in which Apple Silicon might be competitive in price/performance, and then when you take into account heat, noise, and various advantages of Macs (for people who like Macs), the Mac makes sense. That's where I am -- my workloads do better on Apple Silicon than what your link would imply and I care about the other stuff that makes Macs nice.
It also goes back to the people who buy a high end MacPro. They buy what they can afford, use it until it's not useful to them anymore and then get something newer. They just wanna take something out of the box, install their software and get back to work. They don't wanna spend time screwing around with buying and upgrading RAM and other things. Where I work we do this too. We have a specific spec of Mac and PC's we buy. We use them until they're EOL (3yrs) and then we swap them out. We don't screw around with RAM, extra HD, upgrading video cards, etc.
This is literally talking about using some high-performance SSDs as dedicated swap devices. No more, no less.
The headache would be random access, just like with swap. NVMe SSDs can get DDR2 levels of data throughput, but much worse latency for random operations. If you mix reads and writes, performance gets ridiculously bad. Give a 70/30 R/W workload to a flash-based SSD and its performance drops to around 15% of high-queue-depth peak read or peak write.
The 32” Pro Display XDR monitor enclosure with the M1 Mac Studio series computer built in should be a no-brainer, the all in one form factor is superior to the
(Leader/Max Headroom style enclosure computer).
The consensus opinion seems to be that Apple will add AMD GPU support to Apple Silicon for some workflows (AI, Scientific Users, Mining), but may add specialty GPUs of their own design like they did with ProRes acceleration years ago.
There are rumors that Imagination Technology is helping Apple to design a custom GPU with raytracing support. It could be Imagination Technologies is helping Apple design a chip dedicated to raytracing for pro users in addition to integrating with embedded GPUs. It would make a lot of sense if it results in a significant speed increase over an nVidia card. Using multiple GPUs is complex and Apple was burned by that in the trash can Mac Pro that could only use half of it's GPU for most apps, so it would make a lot of senses that any discrete GPU could be targeted for just raytracing acceleration and the built-in GPU would be used for most other tasks.
Apple was rumored to be building a discrete GPU codenamed Lifuka, but it was likely scrapped since it wasn't good enough at any particular task. They could either make a card good at AI, numerical processing, or raytracing but not everything. The Imagination partnership seems to have started around the time it was scrapped.
See #62 for wrap-up on some of the Apple GPU rumors out of Japan.
https://egg-5ch-net.translate.goog/test/read.cgi/mac/1606059468/l50?_x_tr_sl=auto&_x_tr_tl=en&_x_tr_hl=en&_x_tr_pto=wapp
Not happening. Nvidia won't give anybody else sufficient documentation and access to write drivers for their GPUs, and Apple won't give anybody else the ability to write kernel code for their platforms. This is what caused the split in the first place. Apple wants control of all of the code which can cause a kernel panic, because the users who see one will always blame them. Remember that presentation where they said something like 90% of all Safari crashes were caused by "plugins" and everybody knew they meant "Flash"? This is that, but for kernel code.
Another example would be Apple deciding not to make curated monitors for a period of time (many on this very site thought it was a good idea at the time), turns out that none of the third parties could design a monitor that had color accuracy, correct scaling or that wasn’t made out of cheap plastic, the choices were so bad that Apple had to get back into designing competent and curated monitors.
Apple will end up getting back into making (some new version of, AirPort Express) routers for many of the same reasons lack of good third-party support for Apple devices, two steps forward, one step back.
1) Apple has been developing an X series of chips that are dedicated desktop SOCs that are freed up to gulp down power as needed and have the GPU updates missing from M2. This would solve the lack of horsepower from putting M series into what is supposed to be the most beastly computer on the planet. Expandability remains a question.
Or
2) The Mac Pro will use Ultra chips, but have a logic board “fabric” that enables slotted interconnect expsndability. Sort of Like an Xsan within one box, but much more efficient. This would solve the horsepower embarrassment as well as the expandability problem.
3) apple launches X series dedicated desktop SOCs AND ties them together with a logic board “fabric” and comes out with the most advanced system on the planet. This would be the “above and beyond” machine that Apple has been known for selling nice the G5 with the exception of the trash can Mac Pro.
If fast virtual memory swapping over PCIe5 is what you want, why use flash at all? Just put a DRAM controller and a lot of DIMM slots on a PCIe card, and make it look like a swap "drive". These things have existed for ages. This avoids the flash wear problem and is full PCIe bandwidth. With 96-192GB of fast in-package memory in the SoC, all the transfers across PCIe are page-sized bursts so you get maximum PCIe bandwidth. The 96-192GB fast memory is an enormous working set (dwarfing the on-chip caches), which allows the PCIe transfer bottleneck (v4 x16 is 31.5GB/s and v5 is 63GB/s) to be greatly amortized. Compression is even used sometimes to squeeze more data across the bus.
I've been saying for a few years that Apple could do this to deliver expandable memory to those who need it (any how many really need it?), either by building it into the motherboard or by offering/supporting an actual PCIe card. RAM is expensive, but if users really need it, they'll pay.