Apple’s announcement of the M1 certainly generated some rumblings, but they were insufficient warning of the shockwaves to come. Only after products appeared in the wild and were finally subjected to benchmark testing could the M1 be fully appreciated. The first product from Apple’s chip design team meant for the personal computer line surpassed many competing microprocessors and nearly everything currently in other Apple products, particularly in single core and GPU tests.
Apple was gracious enough to release a die photo (another little detail that put them into the company of AMD and Intel since that is now a traditional new process announcement strategy) and these were quickly annotated by processor uber geeks like Andrei Frumusanu at Anandtech.
We know from the earliest announcements that the ARM-powered M1 would be solidly in the system-on-chip category. For the chiplet designs that are expected to continue to take over traditional CPU designs, analysts may have an easier time identifying functional elements since they will be physically distinct pieces of silicon. Understanding the architecture of an SoC takes a little (maybe a lot) more squinting.
And that effort is well worth it for competitors or anyone interested in the efficient use of silicon real estate. As we know from the most cited advantage of the system-in-package paradigm using chiplets over monolithic designs, SoC products are rapidly becoming complex, difficult and expensive. The trend drives exclusivity in the SoC club as fewer chip companies can participate. Understanding the best in-class products and surveying the state-of-the-art is arguably more critical for the SoC than system-in-package.
A few features of the M1 SoC immediately stand out.
First, there is little in the way of area devoted to system cache memory. Apple designers focused on functionality over memory. The universal memory architecture (UMA) of the M1 system frees up die space by keeping the LPDDR4X DRAM close at hand. The UMA approach allows separate DRAM components to be shared by both the CPU and GPU cores to optimize the SoC floorplan. Keeping the DRAM physically close is a concept borrowed from mobile processors that stack the memory on top of the application processor in a package-on-package configuration. This has been discussed in more detail previously, so briefly, the computer approach is designed for higher performance and heat dissipation while the mobile design is minimalist in both space and power consumption.
A big chunk of the M1 die is devoted to the GPU. But you may be saying to yourself, “This is no surprise.” It isn’t. No one would expect Apple to sacrifice graphics performance by reducing the space devoted to an integrated GPU. Again, at some level, the architecture is borrowing from the mobile application processor design.
Another aspect of the M1 die floorplan revealed by Apple was the preponderance of general logic cells (as opposed to processing cores). This is an advantage of designing purpose-built microprocessors for your own operating system. It is possible to build some functionality into firmware and keep the CPU cores for more demanding tasks. My good friend and excellent technology analyst, Paul Boldt, has made this connection on several occasions. One of his contributions quoted Alan Kay which I repeat here. “People who are really serious about software should make their own hardware.”
Steve Jobs quoting Alan Kay
An interesting opportunity is presented by the flip-chip structure used to assemble most advanced integrated circuits. This was noted previously in my recent column and allowed die imagery to be acquired quickly for floorplan analysis by using the transparency of silicon to infrared. This approach saves both time and cost.
Easy access to the M1 in the Mac mini computer also allowed observation during CPU benchmark testing.
Using a thermal imaging camera, it was possible to locate active die areas by their heat signature. The example image contains a bright yellow region which indicates a single high-performance core active during computer operation. Thanks to early access to benchmark tests like Geekbench 5 created specifically for the M1 silicon, this approach allowed confirmation of the die location of both the high performance Firestorm and efficiency Icestorm CPU cores, the GPU, and the neural engine among others.
The thermal imaging was helpful and opens the door to more sophisticated testing. Other computing platforms, particularly mobile ones, complicate such an analysis, however. But if there is a will, there is a way. More importantly (and something self-evident in the semiconductor business), if there is sufficient budget, it can likely be done.
Although much of the focus on the M1 is related to its architecture (and well it should be), the package design and structure offers more than a few interesting technical bits. The M1 BGA substrate with two side-by-side mounted LPDDR4X fully encapsulated BGA DRAM packages is the same (or very similar) to the iPad versions of the mobile A-series processors, the A12X and A12Z.
Probably the best tool for getting an overall understanding of complex system package structures while still providing detailed imaging for analysis is the x-ray CT scan. System Plus Consulting has released an overview of their CT scan results and identified the integration of both surface and substrate-embedded silicon decoupling capacitors in the M1. This package insight is part of the standard System Plus Consulting Reverse Costing analysis.
System Plus recently upgraded their costing reports with new analysis that includes new pieces not present in their earlier reports. According to System Plus CEO Romain Fraux, “The new Apple M1 report offers two important analyses that are new products for System Plus. We now offer a floorplan — or die architecture — report and a front-end structural analysis with TEM (Transmission Electron Microscopes) in addition to our very popular reverse costing report focusing on die and package-level details. This new comprehensive report greatly increases the value of our benchmarking efforts.”
Roles played by iPad Pro
For a few years now, a lot of discussion and speculation has revolved around the iPad Pro and what it meant for Apple. Was the iPad Pro going to be a laptop replacement? This thinking also fueled discussion of a possible transition of the mobile processor and iOS tandem over to the personal computer products.
No doubt, the iPad Pro was part of this journey for Apple. The repeat of the package design across iPad, Macbook Air and Pro, as well as Mac mini certainly indicates a great deal of commonality.
Product benchmarking focuses on cost. Competitive intelligence teams operating inside chip companies may not have to look far to compare performance specs against their own designs. But a big competitive advantage could come from the cost structure compared to the competition. Taking a product apart down to bare silicon will reveal those details to a reasonable degree of granularity given the right experience, industry insight, and modeling.
What makes the work so fascinating is the opportunity to reveal the innovations that keep the semiconductor business marching forward.