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.
What Today's 27" Intel iMac Refresh Tells Us About the Apple Silicon Mac Release Order
- Thread starter Yebubbleman
- Start date
- Sort by reaction score
That sounds like a pretty big gamble....it reminds me of the days of the Wii U and the Playstation 3. 2 game consoles that failed to attract a significant portion of the gaming developer community and what resulted was a lack of games on both consoles and in the case of the Wii U, a shortened lifecycle, which Nintendo quickly replaced with much easier to develop for console, called the Switch.
There are downsides to making a bespoke and completely proprietary platform that developers have to spend time learning - Wii U and PS 3 are just 2 examples that spring to my mind.
True, but "performance" is not one dimensional. See my post above. I agree that is seems nearly impossible not to get better or equal performance (as in compute) of AS compared to Intel as things stands. However remember that GPU scaling is still an issue to consider.
You're not wrong there at all. However, most of the big players are on-board already and much faster than they were in 2005-06's PowerPC to Intel transition. So, the only real casualty I see here are Mac game developers that develop games that don't get regularly updated as well as those that would've otherwise ported from a Windows/XBox/Playstation port. I do believe we will, most sadly, see a lot less of that than we even saw in the PowerPC era, let alone the Intel era.
Yes I know but are these benchmarks running Mac OS on AS? I am not fully convinced that there is 1:1 relation between iPadOS and MAcOS number in for instance Geekbench. If it is, Apple will have hard time not to outperform Intel. However, I do not think that raw performance as in will be the primary goal for the first release.
What is required for Apple to be able to credibly ditch discrete GPUs is higher bandwidth memory subsystems.
The upcoming game consoles (PS5/XBox sx) show that you can achieve excellent graphics performance from a SoC - if you can feed it. The die area won't be a problem on TSMC 5nm. Memory bandwidth still needs to be adressed.
On the low end, I guess iPad Pro level is fine, which should mean an 128-bit wider bus to (upgraded) LPDDR5. That should allow pretty much twice the graphics performance of the A12z.
The next step up, if you want to match the dGPU of the 16" MacBook Pro, is either 256-bit wide to LPDDR5 or HBM, just as on the dGPU. HBM will be a bit restrictive as far as total RAM size, but offers great bandwidth obviously.
As far as iMacs go, Apple pretty much needs to go with GDDR6 (as the new generation consoles) or HBM to match the currently best build to order dGPUs Apple offers. And those dGPUs are what Apple offers right now, not what will be on the market at the time the new AS iMacs will be introduced.
Matching dGPU performance is a much taller order than CPU performance, and absolutely requires higher bandwidth memory subsystems. It will still be difficult due to power draw concerns.
The upcoming game consoles (PS5/XBox sx) show that you can achieve excellent graphics performance from a SoC - if you can feed it. The die area won't be a problem on TSMC 5nm. Memory bandwidth still needs to be adressed.
On the low end, I guess iPad Pro level is fine, which should mean an 128-bit wider bus to (upgraded) LPDDR5. That should allow pretty much twice the graphics performance of the A12z.
The next step up, if you want to match the dGPU of the 16" MacBook Pro, is either 256-bit wide to LPDDR5 or HBM, just as on the dGPU. HBM will be a bit restrictive as far as total RAM size, but offers great bandwidth obviously.
As far as iMacs go, Apple pretty much needs to go with GDDR6 (as the new generation consoles) or HBM to match the currently best build to order dGPUs Apple offers. And those dGPUs are what Apple offers right now, not what will be on the market at the time the new AS iMacs will be introduced.
Matching dGPU performance is a much taller order than CPU performance, and absolutely requires higher bandwidth memory subsystems. It will still be difficult due to power draw concerns.
If there is a differential, it's not substantial. It's mostly the same OS anyway.
Also, I'm not sure why you think that matching or exceeding performance of their Intel models won't be a main objective. It's THE objective, second only to performance per watt improvements.
Not sure if this counts - it's MacOS running the ipad version of geekbench:
New benchmarks and details about iPhone and iPad apps emerge from Apple Silicon Macs - 9to5Mac
At WWDC last month, Apple officially detailed its plans to transition the Mac lineup to custom Apple Silicon processors. As...
9to5mac.com
More or less 1:1 mapping of geekbench CPU results. It does look like the DTK is running Geekbench compute faster than an ipad though (about 25%).
In any case, Apple has already stated that the ASi chips coming out far exceed the DTK A12Z chip - that's one of the reason why they don't allow benchmarking on the DTK - it is in no way representative of what is coming out.
Apple don't want people to judge the DTK by benchmarks running on it and concluding this is what Apple will come out at the end of the year.
Well, he did say "most powerful super-computing cluster is ARM-based" (https://en.wikipedia.org/wiki/TOP500), which is factually correct based on the metrics quoted.
You obviously wouldn't be using Fujitsu A64FX in a desktop / workstation; that's understood. The point was that ARM can used in powerful computers. Apple's SoC will be designed to work well with Apple hardware.
How well it really performs will be revealed later this year.
Last edited:
Not sure if this counts - it's MacOS running the ipad version of geekbench:
New benchmarks and details about iPhone and iPad apps emerge from Apple Silicon Macs - 9to5Mac
At WWDC last month, Apple officially detailed its plans to transition the Mac lineup to custom Apple Silicon processors. As...
More or less 1:1 mapping of geekbench CPU results. It does look like the DTK is running Geekbench compute faster than an ipad though (about 25%).
In any case, Apple has already stated that the ASi chips coming out far exceed the DTK A12Z chip - that's one of the reason why they don't allow benchmarking on the DTK - it is in no way representative of what is coming out.
Apple don't want people to judge the DTK by benchmarks running on it and concluding this is what Apple will come out at the end of the year.
Even when running the Intel version of Geekbench 5 for Mac under Rosetta 2, the Apple Silicon DTK ran comparable to the native performance of the current 2020 MacBook Air. Given that we're most certainly getting faster stuff on the first Apple Silicon Macs that means that Rosetta 2 will AT LEAST MATCH the performance of the CURRENT lowest end Mac. That's honestly not bad considering the PowerPC to Intel Rosetta, at best, matched the performance of a PowerPC G3 (which, at that time was MUCH slower than the lowest-end Mac sold at the onset of that transition).
If you have a look at some of the recent Max Tech videos on YouTube, many developers usering the Apple DTK are reporting very good performance of x86 apps converted with Rosetta 2, and relatively quick development cycles rebuilding x86 apps for ARM. So there may not be a long delay before software vendors have ARM versions of their apps, and Rosetta 2 provides reasonably good performance until they do.
Not sure why you think Metal is "a joke". I've seen several reports that it runs slightly faster for video rendering, even on Premiere Pro which competes against Apple's Final Cut Pro:
Metal VS OpenCL | Neat Blog
Read another article about Neat Video and Neat Image.
I don't see a technological reason why Apple won't be able to match Mac Pro performance (at least a 16-core Intel Xeon W) within 2-3 years. There are already ARM-based chips that can match these Xeon CPUs, and Apple can optimize their SoC to get the best out of their hardware & software ecosystem.
Do you really think Apple didn't analyse whether they can do these before making the decision to move to Apple Silicon? There are very unlikely to release a machine that is slower than the current Mac Pro, and there is no indication that they will exclude the Mac Pro from the transition.
BTW, the semiconductor element used in microprocessors is "silicon". Silicone is used in breast implants.
I agree with this. New ASi Macs need to demonstrate an improvement over the equivalent Intel models, but not eclipse the flagship models. This is why the entry-level / mid-level machines will be the first to transition.
The ASi MacBook Pro 13 (maybe 14?) will need to beat the current one by a reasonable margin - maybe 20% single-core, 50% multi-core, 50-100% better GPU, and better battery. But it still needs to be somewhat less powerful than the MBP 16.
Same story with an ASi iMac 24" and the current Intel 8-10 core iMac 27"
Given the uproar leading up to the 2019 Mac Pro; I can't fathom they'd exclude it. It WILL be fundamentally different in many ways on the other end of the Apple Silicon transition, but it will surely not be left out.
?
The only important thing is that an ASi Mac's performance is greater than the Intel model it replaces. Given that, just as the Core Duo iMacs were faster than their G5 predecessors but not as fast as the Power Mac G5 Quad, you'll have ASi 13" Mac Notebooks that eclipse the 2020 Intel 4-port 13" MacBook Pro but still don't touch the Intel 16" MacBook Pro. But again, the ASi replacements to the higher end Intel Mac models need more time in the oven anyway (hence the point of this thread to begin with.)
If you want to get best performance out of Apple GPUs (and simplify your code), yes, you need to use Metal and Apple-specific rendering techniques. At the same time, most developers use one of the popular game engines (Unity, UE etc.) that take care all the platform-specific stuff for you. There are also open source wrappers that allow you to use standard APIs such as Vulcan on Apple's platforms. Finally, let's not forget WebGPU — an upcoming standard for high-performance GPU programming for web, which is partly based on Apple Metal.
At the same time, you don't need to code specifically for Apple GPUs in order to take advantage of their increased power and memory efficiency. They approach rendering differently to mainstream GPUs and that works for any application. It is just that in some cases, things can be done much more efficiently (and simpler) on the Apple GPU than on the popular GPUs due to their architectural difference.
I also would like to point out that Metal is the AP used on iOS and so far, that platform seems to be striving. Metal is very easy to learn, very flexible, and it comes with a set of good tools. If you are a graphics programmer, picking up Metal takes literally 30 minutes. It's a very straightforward API.
Last edited:
Exactly, second to performance/watt ratio is the king. Performance will follow the improved performance/watt ratio.
The aim would be to sell laptop that has a full day battery life under heavy load or iMacs with other form factors or passive cooling. To sum up, to build significantly different computers enabled by the improved performance to watt ratio.
For the high end lines (>=27 inch iMacs, MBP16, Mac Pro), performance will be the driver at he expense of passive cooling, slimmer form factors or increased battery life.
I think there has only been one rumor that there will be a 14” MBP and it came fairly recently (a month ago?). Almost everything else has been wishful thinking by forum posters based on the 15” being replaced with 16” going back to last year.
I was making another itemised response but this is getting needlessly wordy.
My bottom line is that Apple have always sold refurbished Macs and sometimes these are just leftover out of date models. Very few of the machines you claim were kept available were actually kept in production or weren't kept available for very long. Those that were were mostly for other reasons, not to prolong the life of an obsolete technology.
Apple has a history of ruthlessly dragging us along in a forward direction. Its in their DNA. They will want everything moving to their own silicon as soon as possible.
They showed all of the apps I use to make a living in the frickin' keynote. Better than my current Intel MBP I might add.
I agree. As this thread points out, Apple seems to have carefully positioned its current offerings, and it's clear where an APU + LPDDR5 will cut it. It's the exceptions that are the most interesting.
Do you think either of these solutions would be viable?
1. Using HBM2E as all-system-memory. I think the MBP16 is going to be deeply uncomfortable with four stacks of 16GB HBM2E. Would it work if they used 2.4Gbps or 2Gbps to limit power consumption?
2. Using HBM2E as cache. Apple could just put a stack of HBM on the GPU / APU package, call it cache and stick with LPDDR5 as main memory? This seems like the efficient option.
I'd also like to add to this that Unity and Unreal have already expressed that they'll support Apple Silicon.
From what I know, HBM2 uses less energy than DDR4, so that shouldn't be much of a factor (if a 16" can handle 64GB of DDR4 it should be able to handle 64GB of HBM). I assume that multiple HBM stacks can be connected via a single bus? And even at 2.0 Gbps you are looking at whopping 260GB/sec of bandwidth — 5x faster than DDR4.
I don't think that option 2. is viable. Using HBM for cache makes the entire system much more complex and I am not sure that the benefit would be that great compared to a large SoC-level cache. Once you are using HBM, you might as well go full HBM.
We don’t know what they’ve had in their labs, presumably for years. We’re comparing the A12z/A13 chips to Xeon’s, but there’s a chance they’ve had higher-end prototypes similar to the i7 - i9 range working for years in their labs.
Limiting bandwidth to save energy won’t be necessary. Remember that the current MacBook Pro already has a GPU option that has two stacks of HBM, with an intel CPU and memory subsystem in addition!
Dual stacks of HBM2e offers a maximum of 2x24GB or 48 GB in total. That’s passable for a top end MacBook Pro or iMac, but it may be that capacities can double by the time a big iMac is ready to introduced.
A price we have to pay for all these high bandwidth options is that RAM won’t be user installable, and there will be hard upper limits to capacity. I’d say the compromise is easily worth it though.
I would tend to agree with leman that this is unnecessarily complex for an iMac. For a Mac Pro, I’m not quite as sure. The Mac Pro is by far more difficult to asses unless you have Apples data on how much memory is actually installed in the systems that are currently in use.
You know it escaped me that the MBP16 used DDR4 and not LPDDR4. I might have known it at one point and forgotten it again later...
That is good, though. I still have my reservations about the proximity to the APU, and you have heard me spin that yarn before. You can read my heat/area sermon again below
Regarding the first option, there are several things I would like to point out.
The first is that we have seen no evidence that 24GB stacks of HBM2E are being manufactured by anyone. I would not presume they exist or are viable at this stage.
The second is that the Macbook Pro 16 is already configurable to up to 64GB of RAM. For a lot of reasons, the case for using 48GB of HBM2E over 64GB of DDR5 is weak, and I think we should design with capacity parity in mind.
Now, my concern is not about the energy itself so much as heat and the proximity to the APU. 64GB means four stacks of HBM2E, all of which is more heat-dense than DDR. At 2.8 Gbps, one of these stacks would be consuming about 5W of power each, and they are going to be flanking the APU.
Something like this:
Again, that's 5W for each of those four squares assuming they are running at 2.8Gbps. Running them at the 3.2Gbps or 3.6Gbps advertised by Samsung and SK Hynix is I think a total nonstarter in the 16" Macbook Pro. I'm still of the mind that speeds should be brought down further.
Just slap a big old heat spreader on that puppy and and you are good to go
As you say, the RAM will probably be clocked a bit lower in a laptop, more like 2-3 watts per die. Pair it with a 2048 bit memory interface and you still have some incredible memory bandwidth in a compact laptop. Imagine what having 300GB/s memory bandwidth will do for CPU performance.