Benchmark Results Show Apple M1 Beating Every Intel-Powered MacBook Pro
The Apple M1 launched earlier this week, and benchmark results for the new chip are already starting to trickle out. While these initial tests are preliminary, they still paint a very positive picture for Apple’s upcoming CPU.
In Geekbench 5, an Apple M1 is capable of achieving a 1687 single-thread and 7433 multi-thread score. For comparison, the A14 inside an iPhone scores 1598 / 3995. This surge in multi-threaded performance is likely due to its higher core count (+2 high-performance Firestorm CPU cores) and the near-certainty of sustained higher clock speeds, courtesy of its larger chassis. In the past, iPhone CPUs have been heavily throttled due to high power consumption, so what we’re seeing here might represent the A14 finally stretching its legs.
These scores compare extremely well with existing Intel and AMD CPUs on the market. As always, I recommend caution when extrapolating from synthetic tests to real-world performance. Synthetics are useful for low-level feature analysis and they can shed light on interesting microarchitectural differences, but we don’t rely on them as our primary means of evaluating hardware.
The highest-end MacBook Pro 16-inch that Apple has ever sold features a Core i9-9980HK, an 8C/16T CPU with a base clock of 2.4GHz and a 5GHz boost. Measured performance for this CPU is 1149 / 7329 in GB5. Now, this is an older 14nm CPU. Intel doesn’t currently ship an eight-core Ice Lake mobile CPU, but Apple also ships Intel’s quad-core, ICL-derived Core i7-1068NG7. That chip scores 1352 / 4914. Apple is therefore leading in single-threaded performance even against Intel’s Sunny Cove architecture from 2019, and it’s leading in multi-threaded against the highest performing CPU cores that Intel ships in mobile. While Apple is only tying the Core i9-9980HK in multi-threaded, the M1’s Geekbench scores cleanly beat everything below that point, including the Core i9-9880H.
Since folks will be curious about AMD, here’s how things compare. GeekBench 5 clearly likes Zen 3 much more than Zen 2; the Ryzen 7 3700X scores a 1281 / 8220 in GeekBench 5 according to LegitReviews, while the Ryzen 7 5800X scores a 1673 / 11,246. That’s a 1.3x and 1.36x improvement, respectively, which is well above the gains AMD told us to expect from Zen 2 –> Zen 3 on average (1.19x IPC, and up to 1.25x with IPC and clock). Obviously, however, these CPUs are both desktop processors. We don’t know how AMD’s Zen 3 mobile cores will compare to its desktop chips, but the Ryzen 4800U scores ~1130 in single-threaded GB5 and ~6800 in multi-threaded. That’s an average — we’ll be referring to this Lenovo system, specifically, when we compare performance.
To summarize: AMD’s brand-new Zen 3 8-core matches the M1 in single-threaded performance and significantly exceeds it in multi-threaded performance, but it achieves both of these outcomes at desktop-level power consumption. Data from the comparison suggests Zen 3 will compare significantly more favorably than Zen 2, but data also indicates that AMD’s mobile performance in a 15W envelope is much lower than when allowed desktop power levels, which only makes sense. If we assume that AMD can improve single-threaded performance by 1.3x and 1.36x — meaning they can achieve the entirety of this uplift in exactly the same power envelope — our hypothetical Ryzen 5800U scores 1495 / 9248. Even if we make these very optimistic assumptions, the AMD mobile system would still be consuming far more electricity than its Apple counterpart.
The Intel Core i9-9980HK loses almost every single-threaded sub-test to the Apple M1, but it doesn’t lose them evenly. In some cases, the gap is small; in text compression the M1 scores 1292, compared with 1177 for the 9980HK. The Gaussian blur, face detection, and horizon detection tests all dramatically favor the M1, as does the ray tracing test.
Subtest performance against AMD’s 4800U is distinct from Intel in terms of which tests Apple wins by which margins, but they aren’t much different in overall outcome. AMD’s Ryzen 7 4800U can hit 1318 faces/second compared to Intel’s 750, but the M1 reports 2209. Intel’s 9980HK scores 1218 in the N-body physics test compared with Ryzen’s 936, but the Apple M1 scores 1769. Interestingly, Intel retains leadership in the machine learning subtest, with a score of 1332 versus 1169 on M1 and 965 on 4800U.
Multi-Threaded vs. Single-Threaded
Multi-threaded tests show some differences compared with the ST figures. In ST, the M1 swept both Intel and AMD, nearly across the board. In multi-threading, it’s much more of a fight. The M1 wins AES-XTS compression by a country mile, but Intel is ahead in text compression, 8284 to 7162 (AMD) to 5528 (Apple). AMD decisively wins the image compression test, 10,392 versus Intel at 9,524 and Apple at 7,213. AMD narrowly wins PDF rendering while the M1 wins text rendering over both x86 competitors. Intel’s camera benchmark performance is narrowly higher than AMD or Apple, and it decisively wins the rigid body physics test. Intel’s scaling in the machine learning test is terrible, however, and Apple wins this test in multi-threading.
What does this tell us about M1 scaling versus Intel and AMD chips? That’s unclear. It’s possible the chip’s performance is still constrained by thermals, or by the split-core configuration Apple is using, or both. Scaling factors for the AMD Ryzen 7 4800U, Intel Core i9-9980HK, and Apple M1 are 6.0x, 6.38x, and 4.4x, respectively. Given that Apple is working with a thermally constrained environment and a mix of high-performance and high-efficiency cores, that’s not too surprising. This is also why the M1 isn’t just going to annihilate x86’s market position in the next year or two — Intel and AMD are both more competitive in MT than ST, and more workloads are multi-threaded these days. These results are a real threat to x86, but they aren’t its death rattle, even if they translate well to real-world applications.
It would be inappropriate to draw conclusions from the results of a single synthetic tests, but nothing in these results implies good things for Intel or AMD. Even if GeekBench 5 runs particularly well on Apple hardware or ARM chips more generally, we’re still seeing a CPU with a 5W nominal TDP take on chips with 15W nominal TDPs and walk out an early winner.
It’s entirely possible that sustained testing under load will reveal that the M1 has to throttle quite a bit over time because it lacks a fan, and that Apple’s higher-end MacBook Pros will still outperform it in real-world tests. This is actually the outcome I expect, because while Apple takes a lot of heat for its pricing, professional customers are not going to purchase $2,000 – $3,000 Intel Core i9 laptops if they can get better performance in audio and video rendering suites with a much less expensive machine.
GeekBench 5 is not a real-world benchmark and we’ll wait for real-world tests before we draw conclusions. But nothing about these results points to weak places in the architecture, and none of its sub-scores highlight problems AMD or Intel can immediately exploit. Considering that we’re looking at a 4+4 CPU going up against full octa-core solutions, there’s plenty of margin for GeekBench 5 to be overstating Apple’s performance without changing the result: The M1 appears, at the least, to be highly competitive with x86 core-for-core and clock-for-clock. Further testing will determine the degree to which this is true, or conversely, establish that GB5 is too Apple-friendly to make a useful cross-platform benchmark.
With tests like Cinebench R23 now baking in support for M1, we’ll have an answer to this question sooner rather than later.
Also — just to revisit the “98 percent of PC laptops” bit — the multi-threaded subtests actually show why this isn’t true. “Faster” has a broad contextual meaning, and the Core i9-9980HK is well above the M1 in text compression, image compression, SQLite, PDF rendering, Clang, N-body physics, rigid body physics, and the HDR subtest. Even within a benchmark that shows Apple’s M1 to exceptionally good light, there are distinct and particular workloads where the Core i9-9980HK isn’t just competitive, but victorious by a large enough margin that lower-end Intel CPUs would stand a good chance vis-à-vis the M1 as well.
By the same token, however, Apple’s marketing overreach on that stupid claim shouldn’t lull anyone into ignoring this chip. It represents the most potent threat to x86 dominance that I’ve seen in my entire career.