Apple Mini M1 SoC Teardown

Article By : Don Scansen

After years of speculation and months of anticipation, Apple launched their first desktop and laptop computers powered by their own chip design -- M1 SoC...

After years of speculation and months of anticipation, Apple launched their first desktop and laptop computers powered by their own in-house chip design — M1 SoC. Apple and anticipation need no explanation. Buzz is their business.

The speculation for many years was what shape this new design would take. For a long time, pundits theorized that the design using ARM cores would see Apple more or less moving an A-series design from mobile over to (at least low end) Macs. With the announcement of the newest Cupertino Chip and the lineup of supported computers, Apple assured us that the M1 would be a new design.

Like everyone from Shanghai to Chicago to Shlisselburg in this game, a new product, particularly one from Apple, can mean only one thing – TEARDOWN. But I doubt there is anything terribly new to see on the system level. What we need to know about is the M1 itself.

Inside Mac Mini
If there was anything vanilla among the three products Apple announced, it was the Mac Mini. The Macbook Air is fanless, but the Mini is just a box with forced air cooling. Getting the M1 in the least expensive way possible was one goal. There was a bigger one, but that comes later.

After product purchase, the next phase of teardown is the unboxing. Sorry, but I didn’t take videos. A lot of “tech” coverage addresses this, often in nauseating detail. If you like it, great, but that’s just not my thing. I will make one point about the packaging, though. Apple’s origamists did a great job with the paper card used to contain the power cord under the unit. It was distinctively Apple.

Getting the logic board out
Getting the logic board out of a Mac Mini presents little challenge. Cell phones are always more fun – heat, sharp tools, glue, fumes, and occasionally pyrotechnics from soft lithium power packs. The Mac Mini simply offers no risk of injury or property damage. The assembly is nicely designed and comes apart in a very straightforward fashion. That is, if  you have the right screwdriver. As such things often go, I did not. Truth be told, none of the usual retailers here in Ottawa had one either. The “tamper proof” driver required to open the box is designated Torx T6H. The “H” is what tripped me up. This is the type with the recessed center to accommodate the small pillars in the center of the assembly screws. I don’t know if this was done intentionally, but a great many precision driver kits that are commonly available contain a T7H and above. I can hear the designers at Apple. “Tantalizingly close, isn’t it?”

Fortunately, I analyzed the M1 with the support of my good friends at MuAnalysis. Tricky set up aside (thanks Mohamed), a few minutes on the milling machine and voila! A regular T6 driver was quickly converted. I guess the drivers themselves are not tamper proof.

Back to the M1. With screws removed, logic board slid right out of the case. Four screws held the heat sink in place over the M1 and the power management IC (PMIC). Both the processor and PMIC make use of the finned heat sink that is in turn cooled by the internal fan. As you may have seen from Apple marketing material, the M1 processor is co-located on a package substrate (ball grid array – BGA). The two DRAM packages do not benefit from a heat sink.

Further to the point about Apple marketing material, several features of the image of the package were accurate. There are two side-by-side LPDDR4 DRAM chips co-located on the BGA substrate next to the M1 processor (DRAM placed end-to-end beside the processor silicon). In fact, the image exactly replicated the number, position and scale of the surface mount passive components on the package.

The Apple marketers also put a die image where the M1 silicon is. This is obviously inaccurate for a number of reasons, but this was only mild artistic license.

First, the device is flip chip attached to the BGA. That means that the circuitry side is face down. Looking at the final assembled packaged would give you the backside view of the silicon. That would either be a featureless polished slab with a glassy appearance or in some cases, laser etched with part numbering information.

Second, the Apple M1 had an interesting integrated heat spreader covering the die. This thin metal (probably aluminum) covered only the side of the BGA substrate containing the M1 silicon. Usually, these heat spreaders cover the entire top package surface.

Even in this subtlety, Apple found a way to “think different.” This type of system-in-package (SiP) with 2D arrangement of processor and DRAM is uncommon. Apple made it unique.

Universal Memory Architecture
One more quick note about device arrangement for anyone wondering about the 8G versus 16G RAM versions of the computers. The original internet posts of M1 teardowns were from 16G computers which showed a pair of DRAM packages. With half the amount in the 8G computers, some wondered if there would be only one DRAM package. There are still two for 8G systems, but each package contains half the density of the 16G components. It is simply a matter of packaging half the die in each package for the 8G models. My Mac Mini contained Hynix LPDDR4 chips. Other images I have seen on blogs or Twitter were also Hynix devices.

Naturally, Apple needed to maintain the symmetry and signal paths for the high-speed RAM chips by using two components regardless of system memory size. No thought was given to the idea of creating two package designs for one computer platform. Although my only goal may appear to be higher word count with this description of the M1 package design, it is an important aspect.

It may well be the most important detail of the M1 design. Apple highlighted the new “Universal Memory Architecture” or UMA in the launch (and now on the website).  UMA is the model of efficiency. It allows all the processing cores – both CPU and GPU – to share one large pool of memory. Keeping the DRAM physically close and shared between CPU and GPU cores allows the SoC design to shed some on-chip SRAM cache and reduce the chip footprint.

This is not to say that the proximity is a milestone. Mobile processors keep the DRAM even closer. Whether it’s a Qualcomm Snapdragon processor or one of Apple’s own A-series chips, an LPDDR4 memory component is stacked over it in a 3D packaging concept known as package-on-package (PoP).

It may seem like the M1’s memory placement takes a step back, but the next-door location (as opposed to the apartment on the next floor in PoP) is required for laptop and desktop computer systems. Phone and tablet processors are designed to run cooler and allow for the extra thermal resistance of the PoP reducing heat transfer out of the chip. The non-mobile computing platforms need to have higher performance where the processor is going to generate more heat. The M1 die needs to have a large surface available to contact a heatsink. (For more on thermal design power, consider performance differences between the M1 Macbook Air and the more powerful M1 Macbook Pro which adds a fan for heat extraction.)

The M1 package inspection has so far been superficial. Although the configuration is intriguing, the BGA substrate itself appears to be generic. But Apple’s mobile processor packaging and board level integration have been pushing boundaries. I expect there are some very enticing features, perhaps downright innovations, hidden under the hood. Patience will be rewarded.

Inside M1

M1 functionality, many but not all represent physical blocks on the SoC (source: Apple)

It took a while, but I managed to meander my way back to the new Apple silicon. Much of the anticipation of the M1 focused on the SoC itself and what might come out of the design team. Apple provided more than a few clues. A functional block illustration included operational details many of which can be attributed to physical IP within the SoC:

  • High-performance CPU cores
  • High-efficiency CPU cores
  • Always-on processor
  • GPU
  • High-bandwidth SRAM caches
  • Neural engine
  • Machine learning accelerators
  • Secure enclave
  • Cryptography acceleration
  • Image signal processor
  • Audio processor
  • Thunderbolt / USB 4 controller
  • Gen 4 PCI Express

Apple did not identify the die locations of any of these blocks even if it was suggested by the graphics used to describe the CPU, GPU, and neural engine. Those illustrations were stylized which was a good indication of their inaccuracy. However, Apple did show something that looked very much like a genuine optical image of the SoC die layout on their graphic of the physical product. As discussed above, this was an inexact representation in the sense that it was not true to the assembly onto the package substrate. But it turned out to be a precise version of the SoC physical design.

M1 SoC
Click the image to enlarge. (Source: Apple, MuAnalysis)

This was verified with the assistance of MuAnalysis in Ottawa, Canada. Once the logic board was out and the integrated heat spreader was popped off, it was a quick stop for some backside photos. Martine Simard-Normandin and her team used laser-scanning microscopy (LSM) at infrared wavelengths in order to image through the chip to reveal the circuit patterns on the active die surface. Silicon is transparent to IR which makes flip chip technology a treat for people in my business. A quick flip of that mirror image showed it was a dead ringer (although not color obviously) for the version released by Apple marketing.

M1 SoC

Let’s recap:

  • Teardown – 20 minutes (not including tool manufacturing)
  • IR backside LSM imaging – 15 minutes

The TTRI (time to real information) is a lot faster than you might expect. This approach netted a detailed die image in a little over a half hour.

M1 SoC exposed

Now, the real analysis begins. The eight CPU cores can be quickly identified. But it was already possible to nail those down from the Apple die photo. The same holds for the neural engine although pinpointing the boundary needs more resolution than Apple provided. That leaves a long list of remaining functional blocks to identify. Fortunately, the pandemic leaves a lot of free time to focus on such a task.

Comparing other designs, looking for cloned areas, matching, measuring, and sometimes guessing are the conventional approaches to this type of reverse engineering. By leaving the flip chip SoC in place and using IR imaging through the die’s backside to acquire a die image, the product remains functional. Now, it is time to see what can be learned from monitoring the M1 while it’s put through its paces on the lab bench. I think it will be “illuminating.” Stay tuned.

Subscribe to Newsletter

Leave a comment