Apple announced its second-generation Apple Silicon yesterday with some new details on the CPU’s architecture and what consumers should expect when it debuts later this year. The Apple M2 is built on an improved iteration of TSMC’s 5nm process and features refined versions of its existing CPU cores. Apple never divulges much information about its CPU designs, but the company gave us enough information to infer a few improvements.
Apple is retaining the same eight-core configuration for the baseline M2 as in M1, but the shared L2 cache is larger — 16MB instead of 12MB. Apple claims that the M2’s efficiency cores are much improved compared to M1, but the high level stats it quotes for L1/L2 cache are identical between both designs.
One of the defining features of Apple’s M-class silicon is its very high memory bandwidth. The M1 was already well-outfitted in 2020 when it debuted with 68GB/s of memory bandwidth. The M2 will double this, to 100GB/s, courtesy of an LPDDR5-6400 memory interface instead of LPDDR4X-4266 and up to 24GB of unified memory. The combined bandwidth boost and total RAM upgrade should allow the M2 to address even large memory needs quite well.
Apple is claiming that the M2 will be 1.18x faster than the M1 in the same power envelope. That’s a very solid improvement for a single-generation on what amounts to the same process node. An 18 percent improvement in approximately 18 months is also reasonable. Apple claims it can deliver these gains with no increase in net power consumption. AMD and Intel like to make such claims when they launch new cores as well, but the accuracy of the x86 manufacturers varies widely depending on which specific product generation is being considered.
Apple offers four performance and power consumption comparison graphs — two against the Core i7-1255U, and two against the Core i7-1260P. The point is to illustrate how the M2 can (according to Apple) offer 1.9x the performance of a 10-core laptop or equal performance in 1/4 the power. Against the 12-core Core i7-1260P, Apple promises 87 percent of the maximum performance of that CPU in just 1/4 the power.
The reason for Apple’s differential positioning here is that Core i7-1255U and the Core i7-1260P are very different types of microprocessors. The Core i7-1255U features just two performance cores and up to eight efficiency cores. In this comparison, Apple emphasizes that the M2 offers much higher raw performance. The Core i7-1260P offers four performance cores and eight efficiency cores, and is faster than the M2 in absolute terms. It also draws far more power at maximum, scaling up as high as 65W. In contrast, we see the 1255U stopping at 30W.
What Apple is arguing, with these slides, is that it can triangulate on x86’s weakness. An x86 CPU can beat M1, but only by using vastly more power.
Apple goes on to make a similar set of claims regarding its GPU — namely, that its significantly more powerful than the GPU inside modern Apple systems:
This may be true, but it’s a smaller consideration than the impact of Apple’s CPU silicon. Nobody associates Intel with top-notch GPU performance. Attacking Intel’s CPU performance, in contrast, hits the company where it hurts. The usual caveats are very much in order here: We’re taking Apple’s word regarding its performance on “industry standard benchmarks” with no disclosure on which benchmarks or what settings. This is not how transparent comparisons are made.
At the same time, the feedback I’ve heard from M1 adopters has been almost entirely positive, though a few individuals I’ve talked to have felt Apple oversells performance and power-efficiency in non-native apps like Zoom. It’s clear that for some users, the M1 series has been transformative, with better battery life and quieter system operation than people are used to.
The threat the M1 and M2 present to the x86 ecosystem has always been real, but indirect. The risk is not that Apple will engulf and swallow the entire existing PC industry, but that other manufacturers may see the advantages Apple is bringing to bear against the existing x86 manufacturers and will decide they deserve a piece of it themselves. Another possibility is that AMD and Intel will opt to move away from x86 themselves to deliver improved performance.
None of these things are going to happen in the short term, but they shouldn’t be seen as crazy ideas, either. AMD and Intel both have ARM licenses and AMD flirted with pivoting to ARM back in the early 2010s before going all-in with Ryzen and x86. Qualcomm is preparing to bring chips to market based on work done by Nuvia, an ARM-focused CPU developer that it bought a few years back. Manufacturers like Ampere are continuing to evolve and improve their own ARM-based products.
The fact that x86 isn’t directly facing competition from companies that play in the Windows ecosystem is a short-term irrelevance. The semiconductor industry moves on long timelines. Intel is still fixing damage to its business in 2022 that occurred because of mistakes and product slips that happened in 2017-2019. AMD began Ryzen development in 2012, launched the chip in 2017, and took until 2019 and the launch of Zen 2 to pull ahead of Intel. The chips AMD and Intel are designing today are the chips that will be in the thick of this fight, 2-4 years hence.
It may turn out that x86 is capable of rising to this challenge. Certainly it was 25 years ago, when a lot of bold predictions about the eventual ascendancy of SPARC, Alpha, and PowerPC all came to naught. But not all challenges are vanquished in a single product cycle or a node transition. Qualcomm currently expects to bring Nuvia products to market in 2023. If those chips prove capable of competing effectively with x86, we’re going to see more companies eyeing the duopoly Intel and AMD have historically enjoyed in server, HPC, desktop, and laptop computing and wondering if they might snag a piece of it for themselves.