Tuesday, November 10, 2020

One More Thing: Apple Silicon Macs

Apple (MacRumors, Hacker News):

Apple today announced M1, the most powerful chip it has ever created and the first chip designed specifically for the Mac. M1 is optimized for Mac systems in which small size and power efficiency are critically important. As a system on a chip (SoC), M1 combines numerous powerful technologies into a single chip, and features a unified memory architecture for dramatically improved performance and efficiency. M1 is the first personal computer chip built using cutting-edge 5-nanometer process technology and is packed with an astounding 16 billion transistors, the most Apple has ever put into a chip. It features the world’s fastest CPU core in low-power silicon, the world’s best CPU performance per watt, the world’s fastest integrated graphics in a personal computer, and breakthrough machine learning performance with the Apple Neural Engine. As a result, M1 delivers up to 3.5x faster CPU performance, up to 6x faster GPU performance, and up to 15x faster machine learning, all while enabling battery life up to 2x longer than previous-generation Macs. With its profound increase in performance and efficiency, M1 delivers the biggest leap ever for the Mac.

Apple (MacRumors: Air, Mini, Pro, Hacker News, Slashdot):

Apple today introduced a new MacBook Air, 13-inch MacBook Pro, and Mac mini powered by the revolutionary M1, the first in a family of chips designed by Apple specifically for the Mac.

David Smith:

fun fact: retaining and releasing an NSObject takes ~30 nanoseconds on current gen Intel, and ~6.5 nanoseconds on an M1…and ~14 nanoseconds on an M1 emulating an Intel 😇

Rich Siegel:

After the whole “iOS 14 is shipping tomorrow” thing, macOS developers get a whole extra day!

Previously:

15 Comments RSS · Twitter

Right at the start of Apple's presentation, they said the M1 is optimized for ultra portables. It's their replacement for Intel's U-series chips. The Macs released today are all consumer-level macs intended for people who don't need more than 8 or 16 gb of RAM, who don't need a lot of speed but like not having to worry about battery life, and who don't care about CPU speeds and prefer to not have to decide between an i5 and an i7.

I consulted the tea leaves, and I think the macs that are intended to please programmers and others who need tons of ram and tons of performance, will be coming next year, with an M1X or M2 or whatever chip inside - and will consume more power, have less impressive battery life, but be even more ridiculously fast and come with options for more RAM, etc.

Also: The ONLY configuration you can choose in today's new macs, other than storage tiers, is between 7 and 8 GPU cores in the Air. Getting the extra core costs $250 more but comes bundled with the $200 storage upgrade. (insert rant about egregious amounts Apple charges for storage here)

It’s unclear how many of these changes are a repositioning, or simply Apple only moving the lower-end SKUs to ARM for now. For example, the ARM Mac mini has no 10GigE option, but you can still buy the Intel version that does. The MacBook Pro being limited to 16 GB RAM and two Thunderbolt ports could be for the same reason.

It sounds like the camera hardware is the same, and the improvements are only in software.

I’m not sure that’s right. The slide says “advanced ISP”. So there’s probably at least a new signal processor in there. The sensor could also be improved (and yet limited to 720p). Resolution is hardly the only relevant indicator for camera hardware quality.

We're back to the days of Cyrix and the MediaGX! Unified memory for the framebuffer and the O.S.: check! "Processor Rating" comparing performance to a "recent laptop" instead of raw GHz: check! Used in lower-end computers: check!

All facetiousness aside, the unified memory means you'll have less than 8 GB / 16 GB actually available for your programs than you had previously since the video frame buffer / textures need it too. The speed of your code may vary depending on your screen resolution and the animations occuring (memory contention).

I'll wait for real benchmarks. (Well written programs don't spend much time retaining or releasing NSObjects.) I am amused that my $75 Raspberry Pi comes with 8Gb of on-chip RAM, and "Apple", I'm sorry ARM silicon... RAM actually matters when processing large images/movies or AI. Similarly, unless "tensorflow support" includes training models, the neural engine is of very little use to machine learning programmers. If Apple assumes you'll have another computer to train your model on, why bother having a Mac? (and no, I don't buy the cloud argument).

So far, I see just another product aimed at consumers to consume, not a tool for technical people other than iOS app-developers. Its hardware is incompatible with many free/open-source software tools: assured pain, for no obvious gain. And given Apple's "piss poor" documentation, one should not expect this to be remedied quickly. Unfortunately, I expect the exodus of professionals to Linux/Windows to continue. I wonder how many Mac apps will now be written in "React" and Javascript. Apple better continue working on their "world's fastest*" Javascript performance.

The hard limit of 2 displays is disappointing, a downgrade in the mini's case.

Also a big downgrade for RAM, as the Intel Mac mini supports 64GB of RAM.

Charging $200 for 8GB of extra RAM is absurd, even by Apple standards. I can buy 64GB of fast DDR4 for under $200, at retail. How much memory will be free after OS boot, GPU, and Rosetta take their share? I feel like 16GB — and 512GB of storage — should be the standard for a 'premium' product by now. That extra 256GB probably costs Apple less than $25, and they're charging $200.

Honestly, I'm more excited about the $100 Raspberry Pi 400 that also released this month. That device has a lot more potential to make the world a better place.

M1 apparently uses HBM memory. The maximum capacity allowed by the HBM-2 standard is 24 Gb, and 64 Gb for the new HBM-3 standard. (HBM was invented by AMD for GPUs). A teardown might tell us which is being used.

https://www.tomshardware.com/reviews/glossary-hbm-hbm2-high-bandwidth-memory-definition,5889.html

And here I was thinking "Maybe Apple's new ARM Macs will be so groundbreaking that I'll regret switching to PC a few months ago." LOL, nope. The MBP only has 1 functional I/O port and a headphone jack? What a joke! That's like an iPad with a permanent keyboard and a different OS (maybe that's the point?) My wife's new $850 Dell laptop has 1TB SSD, 8GB RAM, 4 USB ports (3A+1C), headphone, full HDMI, dedicated power jack, and mini SD. It's like Apple wants to go back to the 90s again with these prices and limitations. Maybe the Mac's golden years of ~2005-2012 were just an anomaly.

It looks like the MacBook Air form factor will benefit most from this transition, by being able to go fanless. The fans on the current Air can get quite loud even when browsing websites. So as long as your workload can fit inside of 8-16GB minus what the GPU/Rosetta takes, Apple's popularity in this segment will come down to the quality/stability of macOS these next few years.

Once the higher-end Macs come out, Apple's ARM chips will probably be the fastest retail silicon money can buy. Unfortunately, even if that does happen, I'd still be hesitant to switch back because of Apple's ever-tightening control over what software can be run, and how it can be run (see Little Snitch).

"Is the RAM really on the chip?"

Assuming the diagrams, descriptions, and motherboard renders/photos are accurate, then yes.

Yeah, miss me with this new “Apple Silicon” for at least 2 years. It took Apple literally until *last week(* to make Catalina as “stable” as Mojave. I don’t trust the quality, and my gut says they still have too much technical debt. I didn’t have any trouble with Apple since the 90s until 2 years ago starting with Mojave and I have trust issues. Since last week I’ve been able to achieve more than 3 days of uptime with Catalina and I’ll stay on das Cat for the next 2 years TYVM.

And then with Big Sur there’s the whole network API thing where Apple exempts itself from firewall/network filters via the “new network API” and dropping kexts. And deprecating hostfiles? They work in Catalina, do they work in Big Sur?

These don’t seem to be sysadm securable workstations anymore, I’m wondering if there’s going to be a new “Hardening OS X” doc that will document use of Silicon in secure environments-- or if we need to park secure network switches/Pi-Holes or whatever in front of these machines when we didn’t have to do this before and don’t with Windows or any other vendor’s kit?

Sorry, these come off to me as glorified iPads and I have enough of those.

1. The RAM is not HBM, not sure where you get that idea. It is LPDDRX, just as you get them on iPad, Dual Channel.

2. Integrated Graphics has always had "Shared" Memory. So having Less Memory because of GPU using it is nothing new.

3. Is the RAM really on the chip? Yes. As it has always been on iPad. The M1 is basically A14X.

4 .What are the performance differences between the three new Macs? Basically same chip but thermals. Forget about all the talk on binning. They are literally the same. On MacBook Air, without a Fan you should only be using 10W Peak, 8W sustained. With a Fan you could do up to 30W or more. Consider a Single Core reaches 5W, If you have 4 High Performance Core running at Max Speed that is already 20W. And if you add GPU and Memory Controller power usage and thermals quickly adds up.

5. The GPU difference is one core, something Apple use to increase yield.

6. Thunderbolt Controller, ( or more accurately it should be USB 4 Controller ) is now integrated into the SoC. Compare to Intel's solution where you need a separate TB controller chip. Hence why you get 4 on Intel MacBook. On iPad there is only one port, if the SoC had 4 Port Apple will be wasting some die space. I would not be surprised a larger die, more powerful SoC for MBP 16 would come with 4 USB 4 Port.

7. The maximum RAM for them all is 16 GB > Apple is sticking to LPDDR4X, as A14 is more of a port of A13 to 5nm. The new redesign A15 will likely come with LPDDR5 controller, where it would have the option of 32GB Memory. The ARM of MacBook Pro 16 is not expected until next year. And given the much larger and demanding bandwidth I would expect that specific variant for higher TDP SoC would come with Quad Memory Channel. Hence giving MBP 16 up to 64GB of Memory. ( At least I hope so )

I got it from Apple's Finnish Page: https://archive.ph/nFks7

HBM-muisti : HBM memory

They've updated the page since yesterday. https://translate.google.com/translate?sl=auto&tl=en&u=https://www.apple.com/fi/macbook-air/

[…] let’s go back to reality and be a bit more critical. First, as impressive as these new machines may seem, there is a long list of no shows in them: no […]

So... M1 does allow model training: https://machinelearning.apple.com/updates/ml-compute-training-on-mac
and it's not too shabby. Interesting.

Leave a Comment