So I guess this means that we'll have to wait an additional year for the MacBook Pro that finally smashes through this archaic limit of 16 gigabytes of memory :(
Except if Intel wakes up and relieves us from our pains by inserting a Kaby Lake revision of sorts with an updated memory controller capable of driving more low-power RAM. I personally couldn't care less about 10-20% more processing power per watt from Cannonlake and 10nm, it's the memory limit that's really limiting me.
Not to mention they've already done something similar when they switched chips from Power PC to Intel. I find it hard to see how they'd license x86, but who knows, they do have ridiculous amounts of money piled up...
But how many x86 legacy Mac apps would break if they switched to ARM?
Could it be the case that we’re really waiting on Apple’s lawyer team to figure out how to sidestep Intel’s x86 patent, so OSX can emulate x86 with a bunch of ARM chips?
The biggest issue is likely that they don't have performance headroom they could spend on the emulation though, x86-on-ARM would be way slower than Rosetta let alone m68k
>"So I guess this means that we'll have to wait an additional year for the MacBook Pro that finally smashes through this archaic limit of 16 gigabytes of memory."
Can you elaborate? What are the details of the current chips that limits Macbooks to 16 Gigs? Is it power consumption?
16GB is the limit on LPDDR3 (not sure whether it's spec or Intel's implementation), so you can pick either battery life and 16GB with LPDDR3 or 32+ at the cost of battery life (and a mainboard redesign) with desktop-class DDR4.
The DDR3 spec limits each individual DRAM chip to 8Gb (Gigabits, not bytes) - 8 individual chips makes up a full 64-bit bus, so for 16GB of total memory Apple needs 16x8Gb DDR3 chips in the notebook. I'm not sure if Intel CPU's are limited to 16 individual DRAM chips when used with LPDDR3, but even if they COULD be used with 32x8Gb modules it would suck up a lot more power.
DDR4 allows FOUR times the capacity on the same amount of individual chips, instead of a full dual-rank DIMM capping out at 16GB like DDR3 they can be up to 64GB. On the MacBook line with LPDDR4, assuming the same 16 chip limit, you would also be able to configure it with up to 64GB (16x32Gb).
Gb != GB, those are 4GB full-width (x64) modules, making them 512MB per x8 chip, half the size of what Apple already uses in the MBP line's fully maxed out configuration.
Presumably they're already running it in a multi channel configuration, I don't have a modern MBP to check with but you need 8 individual DRAM chips per DIMM (or soldered on equivalent). Since the JEDEC spec for DDR3 caps each individual chip at 8Gb (1GB) they'll have 16 of them, which are probably run in dual-channel instead of dual-rank single-channel.
With the reported performance of the A11 in the new iPhones, my speculation remains that Apple may well be moving to an in-house MacBook (Pro, and whatever) CPU/GPU system.
I don't know what that would mean with respect to system memory. IIRC, I read that the iPhone 8 comes with 2 GB of RAM. But, its predecessor, the 7, doesn't seem to be lacking, despite this.
What I can't help but wonder is: is the 16GB LPDDR3 limit one of spec or implementation? And will an eventual support for LPDDR4 simply bump the hard limit up to 32GB?
The other option Apple might consider by then is keep RAM at 8/16Gb but offer an Optane module in addition? Dunno if that'll be viable but it might be a better option.
It's not even only Apple anymore. As far as I have seen, the Microsoft Surface Pro suffers from the exact same limitation. Practically every manufacturer who wants to build some high-end, but still thin and lightweight laptop runs into this problem.
What's the deal with the dell xps 15 and precision 5520 laptops then? They're 15 inch but they're still pretty thin and light (1.78kg for the precision 5520), but you can get them with 32 GB of RAM and their battery life apparently isn't terrible. I'm planning on getting one and never realised this was a compromise.
What do you need more than 16GB for? Until less then a year ago, I used to have a 4 GB RAM
(+ 4GB SSD swap) laptop, used it for everything, including software development (Android mostly). Even ran virtual machines occasionally. It wasn't that horrible.
Now I have a MacBook Pro with 16GB and this amount seems more than comfortable.
Video and image processing/editing, for starters. Working with large datasets. Running several vms with reasonable memory allocation. Running local DBs. High-end / VR gaming. Etc. I do all of these things daily for both work and pleasure and with 16GB RAM I have run into countless headaches.
Personally I don't use a Mac and could upgrade at any time but I'm waiting out of principle because DDR4 RAM is still seeing exorbitant prices from the shortage.
Right now on my MPB I'm sitting at 14.68GB used with another 3.73GB in swap. Between Chrome, electron-based desktop apps, and Docker containers it's easy to hit that limit. This is just full stack web dev too, nothing rendering related.
(name process finpitch finheight gatepitch interconnect pitch)
ITRS 14nm - 42 - 42 - 70 - 56
Intel 14nm - 42 - 42 - 70 - 52
Samsung 10nm - 47 - 49 - 68 - 51
TSMC 10nm - ?? - ?? - 66 - 44
Samsung 10nm ~= Intel 14nm by dimensions (actually a bit worse). No idea how you'd go about benchmarking this stuff.
So it appears as though Intel is still ~years ahead of others, as they've been shipping 14nm for 3 years, and Samsung just started shipping this year (but they called it 10nm so they don't look as bad).
Yes. The A11 Bionic CPU Apple is shipping in the newly announced iPhones is on a 10nm min-feature-size process built by TSMC. They announced revenue from the process this summer.
Samsung is ahead of TSMC by a few more months; they've been shipping wafers since March of this year.
7nm or 10nm are purely marketing terms. According to Wikipedia, Samsungs and TSMCs "7nm" process is comparably to Intels 10nm. And their 10nm process is only slightly better than Intels 14nm process. Don't let marketing fool you. Samsung and TSMC aren't really ahead of Intel.
Can you specify the reference? If you're talking about feature sizes like fin height/pitch, tor pitch, IC... sort-of -- but only if you believe Intel's marketing material on what they are going to ship. Because they haven't shipped anything, they haven't had to commit to what that process _actually means_.
According to this, Samsung and TSMC are ahead of Intel, since these two have shipped their “10nm” chips which are ahead of Intel’s top-of-the-line shipped chips in transistor density. When/if Intel ships their 10nm chips, however, the leader position in transistor density would be regained by Intel (by around a factor of two!). Unless TSMC/Samsung manage to squeeze out something better than their current tech before then.
Are Samsung and TSMC's 10nm equivalent to Intel's 10nm? I vaguely recall from studying semiconductors that different companies will call things the same when in fact one is substantially smaller and harder. No idea if that is the case here.
The 10nm moniker comes from ITRS (International Technology Roadmap for Semiconductors). You may be thinking of e.g. memory technology versus CPUs versus FPGAs or something -- but in this case they're roughly comparable. It's true that the specifics are arguably less ambitious (fin height and pitch, gate pitch... are all tighter) -- but that's only if you compare Intel's marketing material to TSMC and Samsung's actually extant wafers :)
Right but samsung's current 10nm process only meets ITRS's 14nm guidelines... looking at tables it appears as though Samsung's 10nm is about the same (actually worse) as Intel's 14nm.
Sort-of. The successor (Ice Lake) is planned to be on a 10nm+ process (so an advanced, 2nd gen version of the 10nm process). However, if the issues with Cannon Lake are mostly about moving to a smaller feature size (i.e. it's mostly a die shrink), and Ice Lake is more of an arch update than a feature size reach, it makes sense that Cannon Lake gets pushed back more than Ice Lake would be. (Yes, it'd necessarily be later -- but if you tackle most of the hard problems now, maybe not _that much_ later, and it makes sense to skip a gen.)
> Can a Coffee Lake CPU work on a Kaby Lake / Skylake motherboard?
No. Although the socket has the same number of pins it is not compatible with Skylake/Kaby Lake. The socket will probably be called LGA 1151-v2 to distinguish it from the earlier socket.
Since Intel nm is not tsmc nm, maybe Intel delays 10nm for different reasons. Like realizing that with a bit more time (that their smaller (sic) nm allows) they could produce much better dies.
So what's the future? Plasmonics? TFETs? Exciton FETs? Nanoelectromechanical switches? Actually looking at the data and realizing the interconnect is the bottleneck?
For Apple, Apple should just go ahead and ship Macs with their ARM chips. Their 10nm A11 chip is already as fast as their current generation MacBook Pros, and that's with a 2 Watt cell phone power budget. Give it a 5 Watt tablet power budget or a 15-35 Watt Laptop power budget and it would wreck anything Intel could produce.
I'm not even sure what the issue is holding them back on switching their Macs to ARM cores? They're already forcing devs to ship byte code to the App store.
If you want to run any software besides what's on the iOS App Store, you need a 64-bit x86 CPU. Apple hasn't released a macOS toolchain for ARM yet, and they need to do that a couple years before releasing such a computer. Otherwise their computer will launch with zero third-party applications (besides the open source ones available on brew).
I remember the Rosetta era when they switched from PPC to Intel and had their VM layer inbetween. You could go to your process list and see what was using x86 vs PPC via Rosetta and later even x86 64-bit. Eventually Rosetta was removed from the default install and I think it's gone entirely now, so you can't run really old PPC apps without an older version of MacOS and an older Mac.
A Rosetta equivalent that runs x86 on ARM is way harder from a pure technical perspective than running PPC on x86. Going from a weak memory model to a strong one is waaaaay easier than the reverse.
Except isn't X86_64 ISA technically property of AMD, not Intel?
I know some of the later stuff (SIMD / AVX / SSE / etc) is Intel, but technically I think they could leave that off and still be successful assuming that would be all that's needed to not be infringing (which I don't know if that's the case or not).
Software emulation is fine (otherwise QEMU would be pretty dead), it's just hardware accelerated emulation and actual HW implementations of the ISA that Intel does not like.
The A11 may need to thermal throttle right after the benchmark completes (which is actually reasonable for many workloads, but games).
IIRC the previous gen A10x in the iPad Pro beat some low-end macs, and usually the next gen non-x is a bit faster than the previous gen x... so, I'd expect the A11 to be faster than some low-end macs.
Could Apple be arranging 3rd party conversions in secret? Or, could the conversion be done seamlessly, at the compiler level?
Anyway, I think ARM macs have been on the cards for a few years now... in a way, the iPad Pro can be seen as a way of experimenting with Mac-level power, as a preparatory step.
Their hesitation might partly be due to a kind of identity crisis: if we can develop iOS apps on a Mac, and that Mac is ARM, why can't we develop on an iOS device itself? But Apple doesn't want open iOS devices.
This is just speculation since I don't own an iPad Pro but I read a comment on reddit that during the beta version of the latest iOS the iPad Pro had thermal management issues where it would get very hot quickly. Could it be that apple is intentionally removing thermal throttling at launch and then enables it after a month to cheat benchmarks and reviews?
I think it's more likely that one of the purposes of the beta was to determine when thermal throttling was ideal, given real-world workloads (which are hard to fine-tune without knowing exactly how people will use it).
It's a pity thermal throttling is necessary; but at least in a Mac form-factor, there would be more cooling options.
Actually, looking at geekbench, the intel processors still have a slight edge per core (4500 vs 4200 in the single core case). You're right about power consumption though.
But regardless, there's so much legacy software that would need to run on a virtual machine and thus be really slow. Probably not worth it - yet. But if intel keeps stalling and Apple's A line keeps excelling like they have been so far, it's only a matter of time.
JIT translation has come a long way and MS plans to do exactly this with x86 software runnning on ARM in the near future. It’s not impossible that A12s with custom silicon working in concert with software JITing could provide a decent experience.
I would be a lot of work though. You don’t just have to translate the compiled user apps, you need to make things like AirPlay work, and that's an example of a feature that uses specific hardware acceleration. And some things like Thunderbolt might not work at all without Intel’s help.
They don’t sell hundreds of millions of macs. It might not be worth it. Why not just wait a year? It’s not like Windows laptops have access to better chips than they do.
> And some things like Thunderbolt might not work at all without Intel’s help.
If my memory is correct, Thunderbolt is mostly an Intel invention, hence they likely own the patents, and as such I wouldn't expect it to work without their support.
Doesn't AirPlay work already on current iPhones? And Intel announced they'll be dropping thunderbolt's licensing fees making it free to implement on anything starting 2018.
I haven't really followed the news. This is a surprising claim, but I admit I haven't checked up on it. Are their chips actually that competitive with a maxed out Intel quad core macbooks? It took a long time to get fully swapped from x86 to ARM. In the scheme of things Intel chips seem like such a marginal cost to Apple, why bother changing?
What other cross-platform benchmark do you suggest? People are comparing it in Geekbench because it is the only big benchmark that is available on both platforms.
snythetic benchmarks are notorious for being unrelistic and unreliable.
just because there isnt an alternative dosent mean that geekbench is accurate, id be pretty skeptical of claims that arm chips with a fraction of the transistors and power budget are somehow magically just as fast as big full fat x86 chips.
Power delivery, sure, but you’re off on transistor count. Wikipedia list a ‘Dual-core + GPU Iris Core i7 Broadwell-U’ as 1.9B transistors [1], and the A11 is 4.3B.
Because being able to dual boot Windows or even Linux is a serious consideration for important segments of the market. Companies that buy computers for their employees are going to be more hesitant to blanket approve Macs for their employees if they can't easily be repurposed to support their Windows (or Linux) need.
Getting all software developers on board to do another architecture transition is also a massive undertaking. Apple has lots of experience in this area. Steve Jobs was asked how long the PowerPC to Intel transition would take and he said 7 years. That's because current Macs still need to be supported and some key vendors like Microsoft and Adobe were extremely slow at transitioning their products.
LLVM bitcode is not architecture agnostic. You still need different builds for ARM vs Intel.
Well, not all of them! x86/64 compatibility is going to be necessary for awhile, and emulation isn't quite there yet. Especially with one of ARM's main advantages being power efficiency.
But seriously, I agree that Apple made a HUGE mistake in completely missing the ARMtop 'Chromebook' market. Google really slammed that niche open - they've gone so far as to trademark 'Apptop' - and even Microsoft is beating them to the punch.
I mean, the 'Apple Apptop'...well, coulda shoulda woulda. 'Google Apptop' sounds good, too; both would be equally incomprehensible names to someone two generations ago.
"x86/64 compatibility is going to be necessary for awhile"
Not necessarily, if Apple can get the developer community onboard with releasing fat Intel/ARM binaries with updates for any already-owned software. Xcode obviously supports both instruction sets, and fat binaries date back to the good old NeXT days. :-)
Not really...Netbooks promised a fully-fledged laptop/desktop replacement, which is not the goal of these highly mobile and very light ARM machines. They aim for battery life, ease-of-use, and flexibility more than all-purpose work.
But I still do use it for EDA software, designing models to 3D print, etc etc. But you often need to compile binaries specifically for ARM to do so, which requires steps that most users won't want to bother with.
Chromebooks are products that are cheap, sturdy, functional, and target both entry-level users who don't need a computer to do much as well as powerusers who want extensibility and relatively strong ownership of the device.
>Their 10nm A11 chip is already as fast as their current generation MacBook Pros, and that's with a 2 Watt cell phone power budget. Give it a 5 Watt tablet power budget or a 15-35 Watt Laptop power budget and it would wreck anything Intel could produce
Haha, what? I'm gonna need some serious citations on your findings there
Because most of the software people run is compiled for x86-64.
I don't know whether the thing about the App Store is true, but even if it is, the App Store is not the only way to download software on macOS, and in fact I suspect only a small minority of people primarily get macOS software there.
TSMC and Samsung, two competing foundries, have already successfully shipped 10nm wafers. They'll also be in the upcoming iPhone, which is a pretty high-volume part.
It's worth noting that node size refers to very different things between different manufacturers, we have very little idea about whats different/similar between TSMC/Samsung/Glofo/Intel's '10nm' nodes.
It's very possible that TSMC and Samsung are using a less technically challenging (and potentially less technically capable) definition of 10nm than Intel.
Or, it could be that TSMC and Samsung are actually technically ahead of Intel.
Yep, that's a fair point. In particular, because Intel isn't actually shipping anything at 10nm, all we have to go off of is their marketing material for what 10nm means.
From Intel themselves [1] TSMC and Samsung 10nm are denser than Intel 14nm. But they're closer to Intel 14nm than Intel 10nm. We'll have to see when it ships, but the expectation is that Intel 10nm will be ahead of TSMC/Samsung in density.
However, both TSMC and Samsung already ship 10nm, so it's also an important point to keep in mind. And then their 7nm processes are expected to leapfrog Intel 10nm. More details in [2], keeping in mind it's mostly announcements and estimations, so keep the salt at hand.
It's true that lithographic technologies ("10nm") don't map to a specific feature size, but TSMC 10nm is denser than Intel 14nm. You may be thinking of Intel 14nm vs TSMC 16 nm vs Samsung 14 nm, which all have comparable fin pitches, and it's certainly true that for the same technology process, Intel has been the most ambitious about what feature size that translates to in the last few years.
Anyone else find it funny that Intel's first CPU with the SHA-1 instruction set is going to be released about a year after SHA-1 was broken? (Yes, there are still certain uses for SHA-1. But the timing is still hilarious)
Except if Intel wakes up and relieves us from our pains by inserting a Kaby Lake revision of sorts with an updated memory controller capable of driving more low-power RAM. I personally couldn't care less about 10-20% more processing power per watt from Cannonlake and 10nm, it's the memory limit that's really limiting me.