I wanted to point those interested here to a discussion thread from late January, after an article about this very likelihood (New Mac Pro may not support PCI-E GPUs). The discussion really gets interesting toward the end, starting with a comment by user Programmer, mentioning TBR (tile-based rendering) and tamping down expectations for the M2 Mac Pro, much of which has now come to pass: https://forums.appleinsider.com/discussion/comment/3396853/#Comment_3396853
Programmer follows up with two final comments about how possibilities open up going forward, as Apple builds on the advantages of TBR architecture. Scroll down to see those.
Bottom line is this Mac Pro is a starting point. Apple Silicon PCI-E GPUs are coming, and they will be in a class of their own (or shared with Imagination Technologies), with the advantages and disadvantages that come with that.
The comments here and elsewhere with people gnashing teeth and rending garments in disappointment and dismay are absurdly shortsighted.
It's good that that they shipped something, and something that a subset of the Mac Pro market would buy, instead of waiting. They really couldn't wait much longer. They weren't very proud of it. Just enough for about 2 minutes of the Keynote, but shipping demonstrates commitment. They need to rev it to a M3 Ultra model and an Extreme model in 18 months, as repetition and continuing improvements is what really demonstrates commitment,
The pricing, if it wasn't obvious since 2019, is pure "workstation". You really have to need something like it to buy it. Small market. It's basically their content creation market with large storage and networking needs.
The original plan was to bridge 4 M1 Max chips together, and presumably 4 M2 Max chips in the current cycle. Would be curious to know if it was a technical issue, like they couldn't get the performance scaling right, or an economic one. For the RAM, they could have done 384 GB by putting RAM packages on the other side of the SOC substrate, but perhaps the expense was too much?
I am interested in how many PCIe 4 lanes are coming out of the SoC.
The original plan was to bridge 4 M1 Max chips together, and presumably 4 M2 Max chips in the current cycle. Would be curious to know if it was a technical issue, like they couldn't get the performance scaling right, or an economic one. For the RAM, they could have done 384 GB by putting RAM packages on the other side of the SOC substrate, but perhaps the expense was too much?
Do you have actual concrete evidence of that? Geometrically, it just doesn't work. A chip only has 4 edges, and the M1/M2 Max use 2 of those for their memory interface and one for the Ultra-connector. If the fourth is unused then perhaps another edge connector, but signal propagation like that along a series of end-to-end chips would be troublesome and physically these are massive chips so 3+ in a row is pushing packaging. Plus they already have non-linear GPU scaling, so going to more with the above problems would really stumble on diminishing returns.
And while going to even more radical topologies (stacking memory, for example) is theoretically possible, it ignores the reality that that isn't what the base/Max/Pro chips do, and the whole point here is to get a higher performing machine by using what they already have. Apple is not going to design a chip just for their highest end products.
My guess is that the M3 generation will likely follow the same model because that will be a 3nm process shrink, which is enough risk/expense for one generation. I expect new/updated core designs across the CPU/GPU/NPU (i.e. improvements within each die), but the packaging/interconnect will look pretty much the same as M1/M2. More exotic solutions will wait until M4 (or beyond) when the rest of the industry has matured the solutions, so Apple doesn't need to carry the burden of initial research/experimentation/implementation they just need to carry their architecture to it.
That doesn't mean that M3 generation (or even the box they just announced this week) couldn't do some interesting things (see discussions kindly linked above by TenThou), but is it worth it for Apple to do those things? Not clear, despite the whining.
Another observation though... why are people so stuck on the idea that the case with the expansion slots must have a higher powered processor than a form factor without them? Especially with power/heat efficient SoCs like Apple Silicon, if you don't need the big case and internal slots, why should you have to pay for them? I have always wanted the most powerful processor, but don't really care much about the expansion space... and I don't want a monster case anymore. So I expect Apple from 2022 on will likely always have a form factor that is relatively small and compact with an option to hold their best silicon. The huge case doesn't need a monopoly on the fastest processor (although it might be able to hold a lot more of them... see discussions linked above).
The original plan was to bridge 4 M1 Max chips together, and presumably 4 M2 Max chips in the current cycle. Would be curious to know if it was a technical issue, like they couldn't get the performance scaling right, or an economic one. For the RAM, they could have done 384 GB by putting RAM packages on the other side of the SOC substrate, but perhaps the expense was too much?
Do you have actual concrete evidence of that? Geometrically, it just doesn't work. A chip only has 4 edges, and the M1/M2 Max use 2 of those for their memory interface and one for the Ultra-connector. If the fourth is unused then perhaps another edge connector, but signal propagation like that along a series of end-to-end chips would be troublesome and physically these are massive chips so 3+ in a row is pushing packaging. Plus they already have non-linear GPU scaling, so going to more with the above problems would really stumble on diminishing returns.
I should have said the original plan as rumored by Mark Gurman. 2 years ago, he was saying the CPU perf + CPU eff + GPU core configurations for Macs would be 8+2+16, 16+4+32, 32+8+64, with the higher end to have 64 and 128 GPU core options. It wasn't hard to notice that the core counts were going up by factors of 2. People didn't know what Apple was going to do to be like those rumors, but once they revealed the M1 Ultra (a 16+4+64 config), it's not much of a leap to conclude the M1 "Extreme" was going to be 4 M1 Max chips bridged together. That was really the rumor rage of what the "Extreme" was going to be.
Yes, the chip only has 4 sides, but there are ways to do it. One way they could have done it was use both sides of the SoC substrate. So, take the M2 Ultra, and put another M2 Ultra on the opposite/other side of the SoC substrate. A one piece silicon bridge that under-laps the SoC chips would have 4 ports, 2 on the same side like it is today with the UltraFusion bridge, and for the "Extreme", it would have another 2 ports on the other side to make 4 ports, and thusly 4 Max chips. The PCIe lanes would be routed across the SoC substrate to an edge connector, like a 40 lane PCIe 4 connector. Heat sinks on both sides of the board.
There would be 16 LPDDR5 RAM pads, 8 per side, to double the RAM from the Ultra models. Apple could double or quadruple RAM capacity further by stacking the LPDDR5 RAM. This is a lot of expense for a very limited amount of unit sales. So perhaps they simply could not accept the costs of doing something like this because of the economics. A lot of people were thinking they would have to make a server box and use it at least in their own data centers to amortize costs.
(This is one way to do it. They could have just made a double long silicon bridge, and eaten the increased latency).
Another observation though... why are people so stuck on the idea that the case with the expansion slots must have a higher powered processor than a form factor without them? Especially with power/heat efficient SoCs like Apple Silicon, if you don't need the big case and internal slots, why should you have to pay for them? I have always wanted the most powerful processor, but don't really care much about the expansion space... and I don't want a monster case anymore. So I expect Apple from 2022 on will likely always have a form factor that is relatively small and compact with an option to hold their best silicon. The huge case doesn't need a monopoly on the fastest processor (although it might be able to hold a lot more of them... see discussions linked above).
Everyone would have no qualms about an Max or an Ultra or an "Extreme" in a Mac Pro box. The issue always remains the price/perf of the box. 7k for the Mac Pro with an Ultra is 3k over the Studio model. Tough 3k pill to swallow for getting internal expansion. The value isn't there but for the niche of users who really need the PCIe slots and internal storage.
Yes, as I said those alternatives are theoretically possible…. But they ignore the reality that Apple isn’t going to do a chip design for just a low volume product. It has to make economic sense.
IMO, the most likely explanation for the original rumour is that Apple was evaluating what would happen if using a multi-socket design. Such a design analysis at an early stage may never have left paper design, and been dismissed for all sorts of reasons very early on. Easy for such a back of the envelope sort of list of possibilities to be leaked.
Apple always charges a huge markup on their big boxes. They sell so few because really very few people need expansion anymore, Small productions runs are very expensive. So the people who really need it will pay the premium.
The rack mounted unit is the most interesting here because such a mounting is usually used headless, and if you’re going to use it headless then why not have a PCIe card with an M2 Ultra on it as well? The Apple Silicon efficiency makes it a good candidate for such a role. The main stumbling block is at the OS because it’s not a market Apple has catered to for quite a while. Of course, the more probable reality is that the rack mount is just for studio builders who have a rack under their mixer board and want to easily slot their Mac in there as well. But we can dream!
Comments
Then user Marvin added links to good information about differences in GPU architectures: https://forums.appleinsider.com/discussion/comment/3397397/#Comment_3397397
Programmer follows up with two final comments about how possibilities open up going forward, as Apple builds on the advantages of TBR architecture. Scroll down to see those.
Bottom line is this Mac Pro is a starting point. Apple Silicon PCI-E GPUs are coming, and they will be in a class of their own (or shared with Imagination Technologies), with the advantages and disadvantages that come with that.
The pricing, if it wasn't obvious since 2019, is pure "workstation". You really have to need something like it to buy it. Small market. It's basically their content creation market with large storage and networking needs.
The original plan was to bridge 4 M1 Max chips together, and presumably 4 M2 Max chips in the current cycle. Would be curious to know if it was a technical issue, like they couldn't get the performance scaling right, or an economic one. For the RAM, they could have done 384 GB by putting RAM packages on the other side of the SOC substrate, but perhaps the expense was too much?
I am interested in how many PCIe 4 lanes are coming out of the SoC.
And while going to even more radical topologies (stacking memory, for example) is theoretically possible, it ignores the reality that that isn't what the base/Max/Pro chips do, and the whole point here is to get a higher performing machine by using what they already have. Apple is not going to design a chip just for their highest end products.
My guess is that the M3 generation will likely follow the same model because that will be a 3nm process shrink, which is enough risk/expense for one generation. I expect new/updated core designs across the CPU/GPU/NPU (i.e. improvements within each die), but the packaging/interconnect will look pretty much the same as M1/M2. More exotic solutions will wait until M4 (or beyond) when the rest of the industry has matured the solutions, so Apple doesn't need to carry the burden of initial research/experimentation/implementation they just need to carry their architecture to it.
That doesn't mean that M3 generation (or even the box they just announced this week) couldn't do some interesting things (see discussions kindly linked above by TenThou), but is it worth it for Apple to do those things? Not clear, despite the whining.
Another observation though... why are people so stuck on the idea that the case with the expansion slots must have a higher powered processor than a form factor without them? Especially with power/heat efficient SoCs like Apple Silicon, if you don't need the big case and internal slots, why should you have to pay for them? I have always wanted the most powerful processor, but don't really care much about the expansion space... and I don't want a monster case anymore. So I expect Apple from 2022 on will likely always have a form factor that is relatively small and compact with an option to hold their best silicon. The huge case doesn't need a monopoly on the fastest processor (although it might be able to hold a lot more of them... see discussions linked above).
Yes, the chip only has 4 sides, but there are ways to do it. One way they could have done it was use both sides of the SoC substrate. So, take the M2 Ultra, and put another M2 Ultra on the opposite/other side of the SoC substrate. A one piece silicon bridge that under-laps the SoC chips would have 4 ports, 2 on the same side like it is today with the UltraFusion bridge, and for the "Extreme", it would have another 2 ports on the other side to make 4 ports, and thusly 4 Max chips. The PCIe lanes would be routed across the SoC substrate to an edge connector, like a 40 lane PCIe 4 connector. Heat sinks on both sides of the board.
There would be 16 LPDDR5 RAM pads, 8 per side, to double the RAM from the Ultra models. Apple could double or quadruple RAM capacity further by stacking the LPDDR5 RAM. This is a lot of expense for a very limited amount of unit sales. So perhaps they simply could not accept the costs of doing something like this because of the economics. A lot of people were thinking they would have to make a server box and use it at least in their own data centers to amortize costs.
(This is one way to do it. They could have just made a double long silicon bridge, and eaten the increased latency).
Everyone would have no qualms about an Max or an Ultra or an "Extreme" in a Mac Pro box. The issue always remains the price/perf of the box. 7k for the Mac Pro with an Ultra is 3k over the Studio model. Tough 3k pill to swallow for getting internal expansion. The value isn't there but for the niche of users who really need the PCIe slots and internal storage.
IMO, the most likely explanation for the original rumour is that Apple was evaluating what would happen if using a multi-socket design. Such a design analysis at an early stage may never have left paper design, and been dismissed for all sorts of reasons very early on. Easy for such a back of the envelope sort of list of possibilities to be leaked.
Apple always charges a huge markup on their big boxes. They sell so few because really very few people need expansion anymore, Small productions runs are very expensive. So the people who really need it will pay the premium.