Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Upgradable memory in future Apple Silicon chips
- Thread starter dogslobber
- Start date
- Sort by reaction score
That's even more incorrect
It's on the same package, but not the same silicon die, which you can see here on the M1 Max:
The "SoC" package is the whole rectangle showing the 4 RAM modules, and the the actual silicon die containing CPU & GPU cores, memory controllers, neural engine and all the other good stuff.
The memory is on the package but not in the primary silicon die.
I'm not convinced that Apple will support user-replaceable RAM again, even on the Mac Pro.
I expect the large iMac will offer single or double M1 Pro / Max SoCs, with up to 128GB RAM
The Mac Pro may offer 4 x M1 Max at up to 256GB, with a possibility of doubling the RAM modules capacity at the some point to give 512GB.
If they did offer additional memory, I think they would need to extend the Apple Silicon architecture to support "local" and "extended" RAM, which would probably need a new version of MacOS to support tiered memory (similar to tiered storage)
I expect the large iMac will offer single or double M1 Pro / Max SoCs, with up to 128GB RAM
The Mac Pro may offer 4 x M1 Max at up to 256GB, with a possibility of doubling the RAM modules capacity at the some point to give 512GB.
If they did offer additional memory, I think they would need to extend the Apple Silicon architecture to support "local" and "extended" RAM, which would probably need a new version of MacOS to support tiered memory (similar to tiered storage)
You say that as a joke, but I can tell you if you have an IBM Power series rack in your data center thats probably how you are paying for hardware.
I don’t see Apple going THAT route, but it is in place today.
I largely agree with this.
One thing that I would add though is it is feasible to add additional memory in the form of caches to feed the memory on package.
It’s been a bit since I looked at Unified Memory setups, or similar configurations like HBM2 or whatever it’s called on AMDs side of the fence, but it is something that could happen.
I don’t know of a commercial implementation, let alone what the trade off for performance vs. power but it is an option.
Unless the transfer rates from the SSD is suffering I don’t know if it’s worth the trade offs.
Like the good old days of extended ram in your 286, or chip memory and fast memory in your amiga.
How is that supposed to work in practical terms? It seems that you are suggesting one of two things: a) that all Macs will ship with max RAM configs and you can pay to unlock then down the line or b) that RAM itself is modular.
The option a) is not feasible because of overall scarcity of high-end RAM Apple uses. It would be a bad business move for Apple, as it will significantly impact their ability to produce these machines and cut into their revenues. Not to mention that it will likely cause the reverse effect — people will be less likely to spend money on RAM upgrades since they can do it down the road anyway — and if they don't do it at purchase, they are probably never going to do it at all.
The option b) is not feasible because there are no efficient mounting options for the types of RAM apple uses. I mean, they could potentially use an LGA-type mount with hundreds or thousands of pins for their custom RAM, but that will be expensive, take up a lot of space, fragile, and likely kill the power-efficiency of their mobile solutions.
How would that work? You'd need dozens of DIMMS to have a decent enough bandwidth. The only way I see DIMMs in upcoming Mac Pros is as a slower (large) memory pool after the unified memory (a kind of cache between the SSD and the actual high-bandwidth RAM)
Laptop CPUs are soldered in for a reason. Sockets are inefficient and take tons of space. Want to have good battery life and compact machines? Forget about sockets.
I would think the AS Mac Pro will come with 16 ECC DIMM slots, that need to be minimally populated with 8 or maximally 16 DIMM modules, giving 400 - 800 GB/s bandwidth, maybe more, depending on the DDR5 speed. Each group of 8 DIMMs must match. Similar design are already being done with server motherboards, albeit with lower groups of DIMMs.
Using DIMM modules as lower tier mass storage may work, as it could be configured as RAM disk swap space by macOS but that will present OS overhead and impact performance, while the SSD act as an even larger but lower tier swap space. If Apple is going this route, they will likely be designing a separate slower speed memory controller to manage the ECC DIMMs, maybe 128-bits data + ECC bus, which the DIMMs would have to be installed in pairs.
You know nothing about hardware architecture.
It will certainly happen the same day you can upgrade GPU memory.
Nope, that's pretty much a real thing: https://www.phoronix.com/scan.php?page=news_item&px=Software-Defined-Silicon-Intel
RAM is physically integrated into the SOC to achieve the highest performance possible; it's not going to be user upgradable.
I take your word that this is something that happens in niche retro hardware, but honestly I have a hard time believing that consumers will accept this in most general products like cars and computers. If Apple did this, it would be the day I switch to daily driving another companies machine. My gut feeling is that most people would have the same objection in-principle to software locks on hardware features. If so, I think sales numbers and market factors would be a barrier against this.
For example, I will *never* buy a BMW with CarPlay locked to a monthly subscription because, F-that.
The things Telsa buyers are OK with are not necessarily things that most consumers will be OK with.
The likes of the M1 chip I expect is uniform and consistent in the physical chip. Deviating manufacturing has costs if several SKUs need to be supported. If the chip is fused to the supported config in factory so it can only use 8GB of the 16GB of RAM then cost is controlled. We see the restriction to seven of eight cores of the GPU in the MBA for marketing reasons. These are all about segmentation of the market and are decided long after the parts tape out. We can apply the same logic to the other two classes of Apple Silicon chip.
My reading of this is that licensing might rollout in say the M2 chip and be retroactively applied to the M1 chip if the above is how they manufactured it. We“ll ultimately need to wait and see though.
What is the rational argument agains this practice, provided the prices stay the same? For example, if you could buy the base 14" MBP and then unlock all CPU cores for the same $200 as Apple charges now? Sounds to me like it would give the users more options with our really removing anything.
It's also possible to put a house on tracks and move it down the street. It doesn't mean that this should or can be a standard thing.
Techie Outrage is not a mass market so don’t kid yourself. Customers will be happy to have piece of mind that they can upgrade via additional licensing if they desire. To the morally outraged, you will quickly be ignored.
The problem is not the chip itself but the RAM. Experiments show that the chip will recognize and use whatever amount of RAM is there. But resoldering the RAM is basically out of the question, as is using socketed RAM. So the only option is pre-installing larger amount of RAM on the device and letting the user unlock it for a fee, which is not economically viable due to the scarcity of RAM itself.
You laugh, but when I bought the OG iMac back in the day, that was an option. I shelled out more to go from 2MB to 6MB (yes MB) of VRAM.It will certainly happen the same day you can upgrade GPU memory.