Rumored pro chip

[theorist9]

Apple will need more cache and the die space is precious. In fact, cache is more important to Apple than to anyone else given their focus on heterogeneous computing and the fact that their systems have lower aggregate memory bandwidth than traditional systems with similar performance.

Contrary to the popular belief, the only significant advantage of Apple integrating the RAM on package is reduced power consumption (and probably material cost). There is no latency or bandwidth advantages to what Apple is doing.

Interesting. So this means there's no functional reason for Apple not to offer user-upgradeable RAM slots on the furture AS large iMac and AS MacPro, right? They may also offer a more powerful Mac mini that can support more than two external monitors; would be nice to see user-upgradeable RAM there as well.

And does this also apply to the GPU? I.e., is there no significant advantage to Apple integrating the GPU on-package (other than reduced power consumption/materials cost). Or is the GPU qualitatively different from RAM in this regard?

 

[leman]

Interesting. So this means there's no functional reason for Apple not to offer user-upgradeable RAM slots on the furture AS large iMac and AS MacPro, right? They may also offer a more powerful Mac mini that can support more than two external monitors; would be nice to see user-upgradeable RAM there as well.

I will try to reply to this to the best of my ability, but please keep in mind that I am not en engineer and the complexities of these things completely elude me. I hope others with more knowledge can chime in.

The problems I see with user-replaceable RAM is that Apple is likely use wider memory interfaces than is common in the industry (meaning more memory channels) and that supporting standard DIMM modules is a different cup of tea than specializing on some hand-picked RAM chips with well known parameters. To be honest, I am skeptical whether it is feasible to build a user-upgradeable system like that that will work as intended. Also, since iMac will likely reuse the same chips as the Mac laptops will use, manufacturing a separate user-replaceable RAM system is a lot of hassle (and additional expense) for Apple with very little benefit.

Mac Pro is a different beast and the problem is that it traditionally could be configured with very different RAM capacities to cater for different use cases. How Apple will handle that remains to be seen.

And does this also apply to the GPU? I.e., is there no significant advantage to Apple integrating the GPU on-package (other than reduced power consumption/materials cost). Or is the GPU qualitatively different from RAM in this regard?

I don't think it matters much. GPU could be on-package or off-package, as long as it is connected to the same memory hierarchy it still has all the UMA properties.

 

[theorist9]

The problems I see with user-replaceable RAM is that Apple is likely use wider memory interfaces than is common in the industry (meaning more memory channels) and that supporting standard DIMM modules is a different cup of tea than specializing on some hand-picked RAM chips with well known parameters. To be honest, I am skeptical whether it is feasible to build a user-upgradeable system like that that will work as intended. Also, since iMac will likely reuse the same chips as the Mac laptops will use, manufacturing a separate user-replaceable RAM system is a lot of hassle (and additional expense) for Apple with very little benefit.

Well, there is a work-around for the first part: Apple could configure its large iMac and MacPro with custom memory slots, and make the custom RAM available from Apple only. I don't know what the economies of scale would be for Apple doing this for the iMac. Or maybe they will produce a new iMac Pro, that uses whatever upgradeable RAM system they develop for the Mac Pro

Are they using the wider memory interfaces now, or is this something that you think we're likely to see with the next-gen AS?

 

[Joelist]

This would be a good cmaier question to be honest. IIRC there are latency advantages to having the RAM in package and thus that much closer to everything else.

 

[leman]

Are they using the wider memory interfaces now, or is this something that you think we're likely to see with the next-gen AS?

M1 uses the usual 128-bit memory bus, but the faster Macs absolutely will need more bandwidth. Apple mentioned multiple times that they intend to go wide here.

As to custom RAM modules, maybe. User-replaceable modules usually have limited pinout, for easier handling. I don’t know how a 512 or 1024 bit RAM module would look like ?

 

[Sydde]

For upgradable RAM, Apple would probably go with some sort of PoP-type configuration – not necessarily full stacking but maybe offset module sockets – that could accommodate either a RAM block package or a specialized GPoC (initially only available from Apple) or some other compact package-complete device (which would call for more technical skill to install than would a PCI card).

 

[theorist9]

M1 uses the usual 128-bit memory bus, but the faster Macs absolutely will need more bandwidth. Apple mentioned multiple times that they intend to go wide here.

**thicc** ?

 

[cmaier]

This would be a good cmaier question to be honest. IIRC there are latency advantages to having the RAM in package and thus that much closer to everything else.

There are two main advantages. Latency is reduced. On-chip, flight time is 6ps per mm. Probably not too far off from the flight time if you stay inside the package. Once you start going out of the package, the latency obviously increases because the distance is further. Also, the picoseconds per millimeter goes up a bit, because the wires need to be bigger and the dielectric isn’t as good. You also likely need to increase the voltage to combat external noise sources, which means it takes longer to switch the signals. And you need bigger drivers to drive all the extra capacitance, which means higher power consumption. You also have to start worrying about inductance and cross-capacitance, unless you shield things very well. All of that increases latency and power consumption. Whether that’s a big concern really depends. If your cache hit rate is high enough, and you have a wide bus between the cache and the memory, extra latency may rarely really be felt by the user.

 

[CWallace]

I suppose Apple could do with the ASi Mac Pro that they did with the Intel Mac Pro - give it actual PCIe slots so you can plug in cards that handle functions not (yet) built into the SoC as well as allow massive amounts of off-package memory with sufficient bandwidth and latency and such.

 

[neinjohn]

M1 uses the usual 128-bit memory bus, but the faster Macs absolutely will need more bandwidth. Apple mentioned multiple times that they intend to go wide here.

As to custom RAM modules, maybe. User-replaceable modules usually have limited pinout, for easier handling. I don’t know how a 512 or 1024 bit RAM module would look like ?

I have no idea if it makes sense but I was thinking if Apple couldn't re-draw (if any is needed, maybe it's already there) their cores to be very memory sensitive (or more attuned to Mac Pro kinda of workloads) as if changed the clock, quantity, wide, type (DDR5/HBM) and how you mixed it (their patent on mixed type memory if I ain't wrong) you could get very different performance levels on a myriad of workloads.

I know Intel and AMD CPUs are already sensitive to differences on memory as DDR4-2400, 3200 or LPDDR4X will get you different performance but further than that and mixed.

I came to this idea about RAM for two reasons. It could serve as a simpler mechanism to segment the market providing different levels of performance on the Pro machines, enticing the buyers to pay 3000/4000/5000$, serious money, as upgraded specs Pro16''/iMac 27'' are a more significant share, and from a post you did some weeks ago, about Apple improving their thermal management to support more watts when they did the redesign for the Pro 15'' 2016.

If Apple was to use the same chip throughout their Pro machines, balancing of form and function is at its hardest on the Pro 16''. It would make sense to push its thermal management the further they can to get a unique SOC competitive on multiple platform.

To give a example to explain myself better I was think of the following products:

> MacBook Pro 16'' with 16 big cores and 32/40 GPU cores as the base and unique option to a combined TDP of let's say 150W. Then you have so defined mix and match for RAM to up 128gb total or more;

> iMac 27''/30'' with 16 big cores and 32/40 GPU cores as the base. Completely similar to the Pro 16'';

> the rumoured Mac mini Pro with 16 big cores and 32/40 GPU, etc;

> the Mac Pro, using Apple version of chiplets, with 32 and 64 big cores CPU options and GPU all the way to 128 cores;

> the Pro 14'' is two M1 glued together and still on the common side of things.

Main idea here would be upgrade the normal professional stuff to further close to workstation (as the Xeon laptop that Apple never launched) as the 1799$ Pro 14'' would probably still be fastest than any Ryzen 9 or i9 on a normal laptop and fit on the low-power GPU focus.

 

[JouniS]

Well, there is a work-around for the first part: Apple could configure its large iMac and MacPro with custom memory slots, and make the custom RAM available from Apple only.

That would remove half of the advantage of having user-upgradeable memory. High-memory Macs tend to be poor value for money if you have to pay Apple prices for RAM. If you can't buy a low-memory configuration and replace the installed memory with standard modules, it's better to just buy the high-memory configuration.

 

[theorist9]

That would remove half of the advantage of having user-upgradeable memory. High-memory Macs tend to be poor value for money if you have to pay Apple prices for RAM. If you can't buy a low-memory configuration and replace the installed memory with standard modules, it's better to just buy the high-memory configuration.

True, but it leaves the other half: You buy an iMac with enough RAM to suit your current & future expected needs, but a year later you need to work on a new project that requires more RAM than you have. You don't want to have to buy a new machine just because of that.

 

[deconstruct60]

True, but it leaves the other half: You buy an iMac with enough RAM to suit your current & future expected needs, but a year later you need to work on a new project that requires more RAM than you have. You don't want to have to buy a new machine just because of that.

So businesses that accurately and prudently predict their expected needs succeed better than the businesses that don't. Not sure why that is Apple's problem. There are 1 , 2 , 3 year leasing options if the business's future is abnormally cloudy. ( Similarly, the resale market for high end Macs typically holds a substantive value of the original price. Sell and then "right size" the equipment in 2-3 won't cost much is projects are lowering return on investment break even point at a steady rate. )

This really isn't the primary driver these "memory slot requirements". Most buyers can see the future pretty clearly by just looking at their own historical track record with rational , quantitative measures. But they also see that in the future what will be this "old" style of DRAM DIMMS ( 3-4 years out ) will be cheaper because it is old. It really primarily not about ease of predicting future needs . A substantive driver of the fill slots later is about backward looking products as much as it is about forward looking ones. RAM will be cheaper in the future and they'll do a "just in time" buy.

That "Just in time" is a trade-off. Higher very long term latencies for cheaper lifecycle costs. Apple is out for maximum shorter term performance and about 3-6 year churn rate. They are also probably willing to throw away "Maximum Capacity" levels too in the very high range over "normal" ( e.g. cap out at around 128-256GB range for next couple of iterations. Longer term will wait for increasing RAM die data densities to move up higher. Same on single primary storage drive (and in most Macs only internal drive.).

 

[deconstruct60]

There are two main advantages. Latency is reduced. ... If your cache hit rate is high enough, and you have a wide bus between the cache and the memory, extra latency may rarely really be felt by the user.

One way of driving a higher cache hit rate is just cover less capacity. A 256MB is about 0.2% of a 128GB main RAM store. 0.4% of a 64GB and 0.8% of 32GB. Wherea if has it is even smaller of a 0.03% of a TB. For an active data set of 2 to 3GB range the relative percentage is in 13% to 9% range ( which is decent).

Once enable the ability to deal with high triple and quad digit GB data working sets then this "keep everything" shortest as possible tends to fall down on the negative outcome trail. Also constrained by the number of RAM packages can place around the immediate perimeter of the primary die(s). If the highly active data working set has to go from RAM to storage then internal caches of the system ( and super wide, single thread at a time function units ) aren't going to save them.

The bulk of Apple's mac units sold are laptops which haven't yet cracked the triple digit GB RAM capacities. Mostly likely their SoCs will make that trade-off. Apple will look to swap out very large data working set customers for another set of customers that don't.
( the bulk of the iPhone and iPad line up is even lower max capacities).

And you need bigger drivers to drive all the extra capacitance, which means higher power consumption.

If Apple gets to solely pick the memory on the other side they can build a memory controller/driver that only works with memory made to their specifications. So if they pick an incremental variation on smartphone RAM packages that lower noise and channel power driving requirements they get a bigger power budget to spend on keeping the computational cores clocking higher for incrementally longer. ( not a Maximum burst clock contest but a higher sustained average clock issue. )

And again, this all happens to line up to be better also for iPhones and iPads too. ( custom RAM but unit volume buys which support paying for custom RAM for Apple. Not that they are passing along huge savings in system prices. )

 

[cmaier]

One way of driving a higher cache hit rate is just cover less capacity. A 256MB is about 0.2% of a 128GB main RAM store. 0.4% of a 64GB and 0.8% of 32GB. Wherea if has it is even smaller of a 0.03% of a TB. For an active data set of 2 to 3GB range the relative percentage is in 13% to 9% range ( which is decent).

Once enable the ability to deal with high triple and quad digit GB data working sets then this "keep everything" shortest as possible tends to fall down on the negative outcome trail. Also constrained by the number of RAM packages can place around the immediate perimeter of the primary die(s). If the highly active data working set has to go from RAM to storage then internal caches of the system ( and super wide, single thread at a time function units ) aren't going to save them.

The bulk of Apple's mac units sold are laptops which haven't yet cracked the triple digit GB RAM capacities. Mostly likely their SoCs will make that trade-off. Apple will look to swap out very large data working set customers for another set of customers that don't.
( the bulk of the iPhone and iPad line up is even lower max capacities).




If Apple gets to solely pick the memory on the other side they can build a memory controller/driver that only works with memory made to their specifications. So if they pick an incremental variation on smartphone RAM packages that lower noise and channel power driving requirements they get a bigger power budget to spend on keeping the computational cores clocking higher for incrementally longer. ( not a Maximum burst clock contest but a higher sustained average clock issue. )

And again, this all happens to line up to be better also for iPhones and iPads too. ( custom RAM but unit volume buys which support paying for custom RAM for Apple. Not that they are passing along huge savings in system prices. )

I’m the guy who wrote a 300 page book on cache (https://www.ecse.rpi.edu/frisc/theses/MaierThesis/), but i have no idea what you’re saying.

I think there are missiNg words in your first section? If you want a higher cache hit rate, then the cache size needs to be a bigger percentage of the address space, or you need a more sophisticated cache fill algorithm (something better than least-recently-used, for example).

I think somewhere in there you are pointing out that paging takes longer than cache fill - that’s true. So if paging is dominating your process, it is true that nobody would notice cache fill latency. But that’s sort of irrelevant to my point, and it’s not the case that most workloads are thrashing the page file.

As for your second point, I don’t care how fancy the RAM packages are - if you have longer wires between cache and RAM, you have more impedence, and you need bigger drivers to compensate. If you have longer wires you also have more noise, regardless of what magic you do with the RAM package. Noise is caused by the wires acting as antennas and cross-capacitors to neighboring wires. Even on-chip this is a major issue - the first tape out of the PowerPC x704 ran 60mhz slower than designed because, to save power, on-chip drivers were reduced in size. This resulted in slower rise times which created cross-coupling problems we hadn’t accounted for. Long wires = noise problems.

 

[deconstruct60]

Interesting. So this means there's no functional reason for Apple not to offer user-upgradeable RAM slots on the furture AS large iMac and AS MacPro, right? They may also offer a more powerful Mac mini that can support more than two external monitors; would be nice to see user-upgradeable RAM there as well.

Strict Functional reason? partially. As much as Apple takes the reduced RAM single driving power on die and applies it to more on die power applied to computational cores , it is still in play for the desktop. Apple "thinned out" the iMac 24" so the max TDP the new chassis can handle went substantially down.

Apple may not "thin out" the Mini but they didn't have loads of room with the current chassis to make it more quiet. So again, they are pretty likely to keep the power savings if can cover the expected max RAM capacity they want.


iMac 27" ... the already did the iMac Pro 2017 that didn't have any access to DIMM slots. And if the iMac 27" SoC is sharing major SoC die components from the MBP 16" SoC (e.g., RAM controller design) then that's a pretyy narrow slippery slope for generic DIMMs.

Also economically, since Apple is using custom RAM packages it only helps lower the overal product line bill of material costs if most of the Mac line up is one the same stuff from the same highly limited set of vendors Apple is buying from. Apple will need economies of scale to drive down the additional cost they add in by going custom.

If Apple makes a "high end" desktop memory controller then they'll have to validated against the DIMM market. That is more time and effort on their part. ( the rest of the Mac and iPhone/iPad line up don't do).

in port above I go through why if Apple puts a 128GB or so maximum capacity on RAM that even more so tracks them into this narrow controller design constraint being "acceptable".

And does this also apply to the GPU? I.e., is there no significant advantage to Apple integrating the GPU on-package (other than reduced power consumption/materials cost). Or is the GPU qualitatively different from RAM in this regard?

if Apple wants to run native iPhone apps at full speed then the integrated Apple GPU is basically a requirement since assumptions about that are pervasive in most iPhone apps. That is a significant "paint yourself into a corner" constraint. Since those Apple make more money than Macs do , it is unlikely that Apple is going to look at that as being a "bad' corner to be in.

iGPU are extremely likely going to be present across all Mac SoCs. The only real question is whether the 2nd (or 3rd) 3rd party GPU is going to be an option down stream in 1-2 years.


Apple could go to non homogenous RAM, but still unified RAM. That would take some work they haven't shown yet. The memory subsystem would have to get more uniform timing access tolerant, but that can conceptually be done.

There are major segments of GPU active memory workload that are substantively different from average user CPU active workload metrics. 100% all of it shouldn't go through a common cache pool with a single cache replacement policy.

But again the more you make those different workload patterns share the same memory system the farther going to take it from generic open market DIMMs slots performance constraints.


if Apple isn't trying to kill off the highest of high end GPU 3rd party GPUs then they have less of a need for the workload access metric gaps to grow so large to cause a bring from their "unified everything memory" approach.

 

[Appletoni]

I have no idea if it makes sense but I was thinking if Apple couldn't re-draw (if any is needed, maybe it's already there) their cores to be very memory sensitive (or more attuned to Mac Pro kinda of workloads) as if changed the clock, quantity, wide, type (DDR5/HBM) and how you mixed it (their patent on mixed type memory if I ain't wrong) you could get very different performance levels on a myriad of workloads.

I know Intel and AMD CPUs are already sensitive to differences on memory as DDR4-2400, 3200 or LPDDR4X will get you different performance but further than that and mixed.

I came to this idea about RAM for two reasons. It could serve as a simpler mechanism to segment the market providing different levels of performance on the Pro machines, enticing the buyers to pay 3000/4000/5000$, serious money, as upgraded specs Pro16''/iMac 27'' are a more significant share, and from a post you did some weeks ago, about Apple improving their thermal management to support more watts when they did the redesign for the Pro 15'' 2016.

If Apple was to use the same chip throughout their Pro machines, balancing of form and function is at its hardest on the Pro 16''. It would make sense to push its thermal management the further they can to get a unique SOC competitive on multiple platform.

To give a example to explain myself better I was think of the following products:

> MacBook Pro 16'' with 16 big cores and 32/40 GPU cores as the base and unique option to a combined TDP of let's say 150W. Then you have so defined mix and match for RAM to up 128gb total or more;

> iMac 27''/30'' with 16 big cores and 32/40 GPU cores as the base. Completely similar to the Pro 16'';

> the rumoured Mac mini Pro with 16 big cores and 32/40 GPU, etc;

> the Mac Pro, using Apple version of chiplets, with 32 and 64 big cores CPU options and GPU all the way to 128 cores;

> the Pro 14'' is two M1 glued together and still on the common side of things.

Main idea here would be upgrade the normal professional stuff to further close to workstation (as the Xeon laptop that Apple never launched) as the 1799$ Pro 14'' would probably still be fastest than any Ryzen 9 or i9 on a normal laptop and fit on the low-power GPU focus.

MacBook Pro 16-inch with 16 performance cores and 40 GPU cores sounds good for the beginning.
128 GB RAM is possible when going from one 32 GB DDR4 module to one 128 GB DDR5 module.

But I would prefer the MacBook Pro 16-inch with 64 performance cores for my work, while the GPU will sleep most of the time. Means much lower heat = no problem.

 

[theorist9]

Apple "thinned out" the iMac 24" so the max TDP the new chassis can handle went substantially down.

Yeah, they really went to extremes with the 24" iMac. And even within that series, they further cut the TDP of the 7-core GPU models by removing one fan! [The ones with 8-core GPU's have 2 fans; the 7-core GPU models have one.] Hence the latter show a bit more thermal throttling. Given that the number of fans is different between the two, Apple should really indicate that in their specs.

 

[theorist9]

So businesses that accurately and prudently predict their expected needs succeed better than the businesses that don't. Not sure why that is Apple's problem. There are 1 , 2 , 3 year leasing options if the business's future is abnormally cloudy. ( Similarly, the resale market for high end Macs typically holds a substantive value of the original price. Sell and then "right size" the equipment in 2-3 won't cost much is projects are lowering return on investment break even point at a steady rate. )

Oh, please. It sounds like you're more interested in being argumentative than reasonable. It's not about it being or not being "Apple's problem". Nor is it about businesses that are 'accurate and prudent' vs. those that aren't.

Instead, responsible businesses *understand* that many changes *can't* be predicted, and thus they plan for the unexpected by purchasing, when applicable, products that are modular/upgradeable. The difference, I would say, is between those who understand the limits of their knowledge, and those who don't.

Apple understands this is a business need. That's why memory and graphics are upgradable, post-purchase, on its new Mac Pro. But, by your thinking, it would have been fine for Apple to offer non-upgradable configurations only, since, hey, if their customers are responsible and prudent they should know what model to get from the start, and if not they can lease, or just sell them and buy new ones.

 

[Jorbanead]

Yeah, they really went to extremes with the 24" iMac. And even within that series, they further cut the TDP of the 7-core GPU models by removing one fan! [The ones with 8-core GPU's have 2 fans; the 7-core GPU models have one.] Hence the latter show a bit more thermal throttling.

We know the M1 can run without a fan, albeit throttled, and the two-fan variant could have a TDP that’s capable of much more than what the M1 requires. The cooling in the Mac mini is substantially better than the cooling in the 13” MacBook Pro which only has a single fan, yet they get similar performance. The Mac mini is only marginally better in certain scenarios - so two fans for the iMac doesn’t seem all that necessary for many people. I’m glad they did it though as it will allow them more thermal headroom for future iMacs if needed.

Aside from that, the baseline iMac is not meant for power users. If you are someone who needed that last 5-10% of juice out of the cpu, you would likely purchase the mid/high range iMac (or more likely you wouldn’t even get a 24” iMac because most power users get a MBP, 27” iMac, or Mac Pro).

 

[Bug-Creator]

And even within that series, they further cut the TDP of the 7-core GPU models by removing one fan! [The ones with 8-core GPU's have 2 fans; the 7-core GPU models have one.] Hence the latter show a bit more thermal throttling.

I'm 99% sure that the 2-fan iMac were supposed to get an M1x which for some reason didn't make the cut. Would also have turned the 2 non TB USB-C into TB.

 

[theorist9]

I'm 99% sure that the 2-fan iMac were supposed to get an M1x which for some reason didn't make the cut. Would also have turned the 2 non TB USB-C into TB.

That would be one possible explanation. It is curious that they offered double the fan capacity for the models that have just one extra GPU core.

 

[EntropyQ3]

Quit making excuses for the single high speed fan of the 24" iMac. It makes for a noisier and throttling computer, neither of which is necessary even in that enclosure as demonstrated by its sibling. It’s just really, really cheap.

When it comes to RAM, apart from the power and speed costs of socketed DIMMs, there is also the aspect of granularity. DIMMs present a 64-bit wide (total) path to DRAM, typical consumer PCs including Apples offering using two sets of DIMMs in parallel, for a (in total) 128-bit wide interface. More professionally oriented products use more channels, the widest workstation systems today being AMD:s Threadripper Pro line that can operate 8 DIMMs in parallel for 512-bits worth of DDR4-speed data, which offers a bandwidth of 200GB/s, nominal.
Going wider than that, though possible, is impractical even for dedicated workstations. It’s also worth noting the toral number of DRAM dies it allows.

Scaling up the M1 in terms of processing capabilities quickly runs into the issue of not being able to scale up the memory subsystem correspondingly, and aside from going completely proprietrary, there isn’t a do-it-all solution around. Which, I guess, is why the question keeps popping up in these threads. Going from 8 to 128 GPU units is a factor of 16. No socketed memory system will offer 1100GB/s in a Mac Pro. And this is about the Mac Pro, no portables or anorectic iMacs are likely at the power levels these solutions will probably end up, which also means that we don’t need to constrain ourselves to mobile packaging. So, picking options from the tried and true we can:
1. Use a wide "normal" RAM system. This gives access to really large amounts of RAM in a straightforward manner. Unfortunately it means feeding the system with data through an (in this context) itty bitty straw. This can to some extent be alleviated by beefing up the cache hierarchy. But only to some extent.
2. Use industry standard HBM. This can, now that Hynix has recently announced their first HBM3 modules, provide 665GB/s per stack. Going with four stacks of HBM, such as in the old AMD Fiji graphics cards, could provide in excess of 2TB/s. Which would be nice, and appropriate. The compromise of course being total memory capacity, which might initially be constrained to 128GB or so. This could conceivably be extended with regular DIMMed RAM, but as noted earlier, that would require some work on the software side of things to maintain the illusion of a homogenous RAM system while maintaining the performance benefits.

Personally, I’m strongly in favour of speed vs. extendability. In computational chemistry, there are a fair number of codes that are ”GPU accellerated", but that actually scales with GPU memory bandwidth, not TFLOPs. (Even published as such which I wouldn’t let pass if I were doing the reviewing). Anything that can be GPU accelerated today would benefit from enhancements in that area, and the problem set would be extended by virtue of the larger memory footprints allowed. Not to mention the benefits to a bunch of parallell CPU codes that use memory beyond LLC cache capacities. And 3D-graphics, and… Comparatively, I just don’t see much of a case for prioritising TB+ RAM areas over speed.

 

[Bug-Creator]

Quit making excuses for the single high speed fan of the 24" iMac.

Who is making "excuses"? Apple did what they did for reasons we can only guess at. Saving 2$ per unit might have been one of them maybe there were others..... *shrug*

What is true is that most people who buy the base iMac and use it in the way it is intended to will never notice it throttling or the fan being noisy.

I do have a 2017 "i7" rMB and yes I am aware that this thing throttles like hell for any workload longer then 10s. Do I notice it in my real life usage? Nope.....

 

[leman]

I'm 99% sure that the 2-fan iMac were supposed to get an M1x which for some reason didn't make the cut. Would also have turned the 2 non TB USB-C into TB.

That would be one possible explanation. It is curious that they offered double the fan capacity for the models that have just one extra GPU core.

It is possible that Apple plans to put a bigger chip (or that the future entry-level chips will scale better with available thermal headroom) in the two-fan iMac in the future. It is also very much possible possible that this is just the usual Apple marketing fanfics to make people choose the second-tier model.

Instead, responsible businesses *understand* that many changes *can't* be predicted, and thus they plan for the unexpected by purchasing, when applicable, products that are modular/upgradeable. The difference, I would say, is between those who understand the limits of their knowledge, and those who don't.

I am not sure how accurate this is. I have spent around a decade in a function as a manager for IT resources of a mid-size research department and I can assure you that concerns over modularity of upgradeability never even entered the equation. For the regular work machines, you buy what you need every couple of years — they are cheap. Even our supercomputer department never upgraded anything: they had a financial plan and wold just buy new hardware every five years or so. For them, it was more important that the systems were extensible, e.g. that they could buy a new server blade or a new storage unit to extend the existing system.

Upgradeability in a business environment is rarely meaningful:

- It puts a lot of pressure on the usually strained resources of IT departments that really have better things to do instead of tinkering with components
- You are putting yourself at a financial risk by investing into hardware outside of it's warranty or service window
- By the time you need to upgrade, computers likely became much faster anyway, so replacing the system is often the better choice
- Upgrading rarely save any meaningful amount of money anyway (if the cost of the new GPU is 60% of the computer cost, you are not saving anything, those few $$$$ are just peanuts)
- Should your demands change in an unpredictable manner, upgradeability does not help. It is never the case that you go "oh, I have misjudged how much RAM I need, I should have bought 64GB instead of 16GB". If you find yourself in such a situation, it's not just more RAM that you need. You likely need a bigger system overall.
- Equipment is cheap, labor is expensive

My impression from all this is that upgradability is mostly a thing of a home enthusiast PC builder, who likes tinkering with the components, upgrade every time a new gaming GPU comes out and has a limited budget.

When it comes to RAM, apart from the power and speed costs of socketed DIMMs, there is also the aspect of granularity. DIMMs present a 64-bit wide (total) path to DRAM, typical consumer PCs including Apples offering using two sets of DIMMs in parallel, for a (in total) 128-bit wide interface. More professionally oriented products use more channels, the widest workstation systems today being AMD:s Threadripper Pro line that can operate 8 DIMMs in parallel for 512-bits worth of DDR4-speed data, which offers a bandwidth of 200GB/s, nominal.
Going wider than that, though possible, is impractical even for dedicated workstations. It’s also worth noting the toral number of DRAM dies it allows.

Scaling up the M1 in terms of processing capabilities quickly runs into the issue of not being able to scale up the memory subsystem correspondingly, and aside from going completely proprietrary, there isn’t a do-it-all solution around. Which, I guess, is why the question keeps popping up in these threads.

Yep, that's pretty much this and that's what folks have difficulty getting. We are so used to the "traditional" modular PC industry that we tend to forget that it comes with it's own set of issues and tradeoffs. PCs are modular because it is a market consisting of many players (both manufacturers and users), and modularity is what allows this industry to throve and to compete, while giving the users the flexibility to mix and match. But modularity limits performance, and Apple build integrated systems, not mx-and-match systems for the general market.

Going from 8 to 128 GPU units is a factor of 16. No socketed memory system will offer 1100GB/s in a Mac Pro.

It's not impossible, but I don't know whether I would call a system like that user-upgradeable. This kind of bandwidth is achievable with 16 DDR5 SO-DIMM slots, all of which have to be filled with identical modules. You can't just decide to add another stick or two if you feel like it, you have to replace everything at once. Any change in RAM configuration in a system like hat would cost thousands of dollars, and that's on top of the already extremely expensive mainboard (wiring and powering 16 RAM channels does not come for free). I mean, already something trivial like 64GB or RAM would be prohibitively expensive (how much would 16x 4GB DDR5 modules cost?)

Personally, I’m strongly in favour of speed vs. extendability.

Exactly. In the consumer market, one is used for things being modular and extensible. In the professional market, nobody really cares. HPC hardware is optimized for speed, not upgradeability. Modularity is interesting at the node level (so that you can replace failed nodes or add new nodes), not at the sub-node level. And Apple Silicon is basically supercomputer HPC design shrinked down to the user market.