Mac ARM: more details on Apple M1

Apple has made huge promises about the performance and efficiency of the new M1 system-on-chip, but hasn’t released any common benchmarks. The first results have now appeared in Geekbench 5, showing the M1 in multi-core performance far beyond the currently fastest x86 mobile processors with TDPs of 15 to 30 watts, namely Intel Core i7-1185G7 (4 cores) and AMD Ryzen 7 Pro 4750U (8 cores). But the results on the iPhone 12’s Apple A14, which is closely related to the M1, reveal more about the M1 technology.

Eight cores, GPU and AI

According to Apple, the M1 contains eight processing cores, a GPU with eight cores or clusters of shaders, and a neural engine with 16 cores for artificial intelligence and machine learning algorithms. The GPU can also be used as an accelerator for some computing tasks.

The M1 System-on-Chip (SoC) also contains controllers for I / O interfaces (Thunderbolt / USB4, PCIe 4.0) and functions previously handled by the separate T2 security chip with integrated SSD controller, including Secure Enclave (SE). The latter is important for biometric authentication (Face ID, Touch ID) and payment functions.

The M1 also has a RAM controller that connects up to 16GB of memory, presumably with two 64-bit channels for LPDDR4 or LPDDR4X.

Fire and ice

An internal “structure” connects the individual function blocks and an additional cache that intercepts memory accesses. The CPU cores are divided into two groups, each with its own L2 cache. Each individual CPU core, in turn, has an L1 cache, as is customary with x86, separately for instructions, abbreviated L1I and data (L1D).

The M1 has four particularly powerful and particularly efficient CPU cores. Apple hasn’t officially confirmed this, but experts say the code names for the strong cores are “Firestorm” and the most efficient ones are “Icestorm”.

According to Apple, each Firestorm core has a 192Kbyte L1I cache and a 128Kbyte L1D cache. These data buffers are much larger than, for example, the current Intel “Tiger Lake” (48 KByte L1I / 32 KByte L1D) and AMD Zen 2 (32 KByte each) CPUs. However, x86 processors also have so-called micro-op caches (µOP), which work particularly efficiently.

Apple’s different ARM core groups can still access a shared L2 cache: the four Firestorms out of a total of 12 Mbytes, the four Icestorms over 4 Mbytes. Furthermore, there is the fabric cache, which Apple has not disclosed the capacity of and which, from the point of view of the CPU cores, acts as an L3 cache.

The “Apple Silicon” A14 Bionic is very similar to the M1 and has as many efficient Icestorm cores and a Neural Engine with 16 cores. The A14 has only two powerful Firestorm cores with a smaller L2 cache (8 MB) and a GPU with 4 instead of 8 cores. Also, the A14’s arithmetic units are generally lower clocked because it can dissipate less heat in iPhones and iPad Air 4s and consume less battery than Macs.

Clock frequencies

Apple does not reveal the nominal clock rates of the A14 and M1, and neither does the Thermal Design Power (TDP). The latter is usually around 5 watts for smartphone SoCs like the A14, but even then a smartphone would get very hot in the long run and an 11 watt hour (3.7 volt / 3000 mAh) battery would drain. in just over 2 hours. On average, the power consumption of the chip is much lower, especially since the display usually needs more power and the modem also consumes electricity.

When it introduced the M1, Apple made comparisons to other 10-watt chips, so it’s probably between 10 watts in the fanless MacBook Air and 15 to 20 watts in fan-cooled mode. The “U” class processors from AMD and Intel have a TDP of 15 to 28 watts, but here too the quality of the cooling system of the respective notebook manufacturer is paramount. These processors can be set to a wide range via configurable TDP (cTDP) and thus provide more or less computing power because they have to slow down quickly under sustained load. This also applies to the M1, as Apple explained at launch: in the 13-inch MacBook Pro with fan, its sustained performance is higher.

Geekbench-Daten

However, the results of the online database of the Geekbench 5 benchmark provide information on the clock rates of the A14 and M1. There, the 2.99 GHz value can be found for an A14 and a value of 3.2 in the M1 result from a fanless MacBook Air GHz. In a 13-inch MacBook Pro or Mac mini with a fan, the M1 might have a slightly faster clock.

However, these clock rates are significantly lower than the turbo rates of the Core i7-1185G7 (4.8 GHz) and the AMD Ryzen 7 Pro 4750U (4.1 GHz). For higher clock rates, both ARM and x86 cores require higher core voltages, associated with significantly greater power loss. The “wider” data paths are arranged, ie the more arithmetic units and transistors run at higher clock voltage and clock frequency, the more losses increase during the clock.

The results the A14 and M1 achieve in Geekbench 5 are even more impressive. In single-core running, which certainly works on a strong Firestorm core, it’s 1596 (A14) and 1732 points (M1). The 8.5 percent advantage of the M1 is largely due to the higher clock rate; the larger L2 cache does not have a major impact.

Compared to a Core i7-1185G7 in an MSI notebook, the M1 is ahead by 23% in multi-core classification and 7.5% faster in single-core. But there are also Geekbench 5 values ​​of a Core i7-1165G7 under Linux which show it is nearly on par in single-core performance, but here it is much weaker in multi-core.

Intel’s Tiger Lakes, like the aforementioned Core i7-1185G7 and 1165G7, have only four cores each and Simultaneous Multi-Threading (SMT aka Hyper-Threading), while all eight cores of the M1 run simultaneously – fast and efficient together. . It is therefore interesting to compare it to the 15-watt notebook processor with the most powerful multithreading capabilities to date, the AMD Ryzen 7 Pro 4750U with eight cores and SMT, which is a total of 16 threads. But even that is around 15% lower than Apple’s M1 in Geekbench 5 multi-core, because its Zen 2 cores offer 33% lower single-core performance than a 3.2 GHz Apple Firestorm. With the Ryzen 5000U (Cezanne) with Zen 3, this could change in 2021.