Apple announced its Apple Silicon M2 processors today, saying that the new second-gen chips offer up to 18% more performance in unspecified multi-threaded CPU-focused tasks, while the revamped 10-core GPU offers up to 35% more performance in unspecified graphics workloads. Apple also bumped maximum memory capacity up to 24GB of LPDDR5, and the next-gen 16-core neural engine is 43% faster than its predecessor and can process up to 15.8 trillion operations per second.
The new M2 chips come fabbed on a second-generation 5nm process, presumably TSMCs N5P, and have 20 billion transistors. The first M2 processors will debut in the MacBook Air and the MacBook Pro, which will be available next month.
The Arm-powered Apple Silicon chips have reinvigorated the company's PC products with the now well-known M1, M1 Pro, M1 Max, and M1 Ultra chips enabling the company to cut ties with Intel processors and move to more advanced chip fabrication technology and the Arm microarchitecture. That trend continues as Apple moves forward to the second-gen 5nm process and a more performant chip architecture that sees advancements in all of the key architectural units.
| Row 0 - Cell 0 | Apple Silicon M2 CPU | Apple Silicon M1 CPU |
| Performance Cores | 4 | 4 |
| Efficiency Cores | 4 | 4 |
| Transistors | 20 Billion | 16 Billion |
| Core Architecture (P- / E-Core) | Avalanche / Blizzard (not confirmed) | Firestorm / Icestorm |
| Memory | 24 GB of LPDDR5 (100 GB/s) | 16GB of LPDDR4X-4266 (68 GB/s) |
| Share L2 Cache (P- / E-Core) | 16 MB / 4 MB | 12MB / 4 MB |
| Instruction Cache (P- / E-Core) | 192 / 192 KB | 192 / 192 KB |
| Data Cache (P- / E-Core) | 128 / 64 KB | 128 / 64 KB |
| Neural Engine | 15.8 TOPS | 11 TOPS |
The M2 processors come with up to eight CPU cores, the same as their predecessors, with four high-performance and four efficiency cores. The M2 chip appears to be based on the A15 Avalanche+Blizzard architecture with ARMv8.5-A, and not Armv9. The high-performance cores have a beefed-up cache, with 16MB of shared L2 cache compared to M1's 12MB L2. The four efficiency cores have an unchanged cache capacity hierarchy compared to the M1 processors.
As with all performance claims from a vendor, you should take these with a grain of salt. As you can see in the above album, Apple claims that the combination of these CPU cores offers up to 18% more performance than the M1 in an unspecified multi-threaded CPU workload — this means that the 18% performance improvement is not indicative of increased instruction per cycle (IPC) throughput.
This unknown threaded benchmark also doesn't tell us which group of cores, be they efficiency or performance, contribute the most to the increased performance. The performance cores (P-cores) handle latency-sensitive work for high-performance applications, while four efficiency cores (E-cores) step in for background and threaded workloads. It is common knowledge that the A15 architecture's E-cores deliver a larger performance increase than the P-cores, so we might see muted gains with the M2 in lightly-threaded work.
The company also claims the M2 chips offer 1.9X the performance of a 10-core Intel Core i7-1255U processor paired with 16GB of memory, but with both chips constrained to the same power limit — not at peak performance. Apple also claims the M2 offers the same peak performance as the 10-core chip but at 1/4th the power. Moving up to a 12-core Intel Core i7-1260P, Apple claims it delivers 87% of the peak performance while using 1/4th of the power.
| Row 0 - Cell 0 | Apple Silicon M2 GPU | Apple Silicon M1 GPU |
| Cores | 10 Cores | 8 Cores |
| Teraflops | 3.6 | 2.6 |
| Gigatexels/Second | 111 | 82 |
| Gigapixels/Second | 55 | 41 |
The GPU has also seen an overhaul, increasing from the eight cores present on the M1 chips to ten cores, which Apple says contributes to a 35% gain in GPU performance, again with an unspecified workload. The M2's GPU is spec'd at 3.6 teraflops, a sizeable 38% increase over the 2.6 teraflops from the M1's GPU.
The media engine supports up to 8K H.264, HEVC, has ProRes encode/decode, and features 'increased bandwidth' that allows it to play back multiple 4K and 8K streams. As before, the chip only supports two displays, with one external up to a 6K resolution. As Apple has done in the past, we expect the company to ship different M2 models with varying numbers of GPU cores.
Apple claims the M2's GPU delivers up to 25% more performance at the same power as the M1 and up to 35% more performance at peak power. However, in a rather meaningless comparison, Apple compared its GPU to a Core i7's integrated GPU that isn't meant for any serious work. Apple claims a 2.3X advantage over Intel's iGPU at the same power, and the same peak performance at 1/5th the power.
Apple feeds both the CPU and the GPU with up to 100 GB/s of memory bandwidth via up to 24GB of on-package LPDDR5 memory, a 50% increase in bandwidth over the prior-gen M1 chips. That comes courtesy of the step up to LPDDR5 over the M1's LPDDR4X. The M2 also has a 50% increase in memory capacity (the M1 topped out at 16GB). The LPDDR5 memory communicates across a 128-bit wide bus.
Specialized silicon for hardware-accelerated workloads is becoming more of a centerpiece in all chips, and Apple has also made inroads here as well. Apple claims that its next-gen 16-core neural engine is 43% faster than its predecessor, processing up to 15.8 trillion operations per second compared to the M1's 11 trillion operations. Surprisingly, Apple accomplishes more work with the same number of neural cores as found on the M1, attributing the increased performance to architectural enhancements. However, we don't know if Apple ended up dedicating more die area to these units to improve performance.
Apple's move to a more advanced process node than Intel and AMD continues with the move from the first-gen TSMC 5nm process (5N) with the M1 to the second-gen 5nm process, presumably TSMC's N5P. Apple spreads the M2 design across 20 billion transistors, a 25% increase over the M1 processor.
As you can see above, the M2 processor is also larger than its predecessor. This seems to be a necessity given that N5P doesn't come with density improvements and Apple added two more GPU cores, but other alterations to the design (perhaps including smaller functional units) appear to have yielded a chip that is roughly 18% larger (assuming that Apple's graphic is to scale).
Compared to the TSMC N5 process found in the M1, the N5P process is said to be 7% faster at the same power, or reduce power consumption at the same clocks by 15% (you can't have both).
Apple will wrap these chips in the MacBook Air and MacBook Pro chassis (click the link for more info), with the former coming as a fanless design while the latter will feature an active cooling solution (fan) to enable higher performance in more demanding workloads. Both the Air and the Pro will be available in July, but Apple hasn't given a specific launch date.
On the whole, the increased performance from the M2 seems in line with the increased transistor budget and die area, implying that the M2 might not have as much of a stellar performance-per-watt ratio as its predecessor. Additionally, the CPU performance numbers touted by Apple don't seem quite as impressive as some have expected, which isn't surprising because the company likely picked most of the low-hanging architectural fruit with the first-gen chips, and also benefited from the step to a much newer and denser process node. This time around, the relatively small step from the N5 process node to N5P brings smaller performance and power benefits while not offering a density increase, and the microarchitectural gains seem much slimmer from afar. As always, the final verdict will come in third-party benchmarks.
