Apple's A20 Pro posts record Geekbench 7 single-core score on TSMC's 2nm process
A six-core phone chip out-running 16-core desktop parts in single-thread tests shows why core design, clock speed and memory bandwidth matter more than core count for many tasks.
Apple's A20 Pro, the processor in its newest iPhones, has set a new single-core record in Geekbench 7, according to Tom's Hardware on 12 September. An early submission puts the chip at 4,006 points single-thread and 11,460 multi-thread, gains of 23.3 percent and 27.1 percent over the A19 Pro. Tom's Hardware cautions that the result comes from a single early listing and may not reflect shipping devices.
The comparisons that drew attention are against desktop parts. The A20 Pro leads AMD's 16-core Ryzen 9 9950X3D by 26 percent and Intel's Core i9-14900KS by 32 percent in the single-thread test, though both desktop processors remain far ahead once many threads are involved. Against Intel's Core Ultra X9 388H, the A20 Pro is 48.7 percent faster single-threaded while the 16-core laptop chip is 63 percent faster multi-threaded. Within Apple's own line-up it is 7 percent ahead of the M5 and 20 percent ahead of the M4 on single-thread work, but trails those 10-core chips by 39 and 27 percent in multi-thread scores, and sits only 5 percent behind the eight-core M3.
Among phone processors the gap is wider: 31.5 to 33.7 percent single-thread over Qualcomm's Snapdragon 8 Elite Gen5 and Xiaomi's XRing O3, both of which the report says perform roughly at the level of the two-year-old A18 Pro, with larger margins over MediaTek's Dimensity 9400, Google's Tensor G5 and Huawei's Kirin 9050 Pro.
The A20 Pro keeps a six-core layout: two "super" performance cores running up to 4.93 GHz and four efficiency cores, paired with a memory interface delivering 50 percent more bandwidth than its predecessor, reportedly through a 96-bit memory bus. It is Apple's first chip on TSMC's N2 2nm-class process, following the A19 Pro on N3P. Tom's Hardware notes that the A20 Pro's 7.1 percent single-thread lead over the M5 matches its 7.1 percent higher clock, suggesting the two share the same core design.
Why it matters
The result shows how much a process step plus a wide core design can deliver even before GPU and NPU details are known. Phone silicon is now competitive with laptop and desktop parts for the latency-sensitive, single-threaded work that dominates everyday use. Apple reusing its largest cores across phones, tablets and laptops points to a converging design across its product line. Rival phone chipmakers face a widening gap on per-core performance rather than a narrowing one.
The result shows how much a process step plus a wide core design can deliver even before GPU and NPU details are known.
- vs Apple A19 Pro+23.3%
- vs AMD Ryzen 9 9950X3D+26%
- vs Intel Core i9-14900KS+32%
- vs Intel Core Ultra X9 388H+48.7%
Figures: Percentages as reported by Tom's Hardware from an early Geekbench 7 submission
What you can learn from this
- Single-thread and multi-thread scores measure different things. A single-thread test stresses one core's design, clock speed and memory path, while a multi-thread test rewards having many cores that can run in parallel. That is why a six-core phone chip can win one column and lose the other to a 16-core desktop part. Most interactive tasks, from scrolling to compiling a single file, are limited by the fastest core, which is why single-thread results matter so much for perceived speed.
- Clock speed and instructions per clock multiply together. Performance per core is roughly frequency times the work done each cycle, and the second term depends on front-end bandwidth, cache hierarchy and branch prediction. A near-5 GHz phone core is unusual because high clocks cost power, so it only works when the process node makes each switch cheaper. The observation that a 7.1 percent clock gain gave a 7.1 percent score gain is a neat illustration of the same core scaling linearly with frequency.
- Memory bandwidth feeds the cores or starves them. A faster core does nothing useful while waiting for data, so bandwidth and latency to memory matter as much as the arithmetic units. Widening the memory bus, as reported here, raises bytes delivered per second without raising clock speed, though latency matters too, because a stalled pipeline cannot be rescued by bandwidth alone. This is why chip upgrades so often pair a core redesign with a memory subsystem change rather than shipping one without the other.
- Process nodes shrink transistors, and the gain shows up as clocks or power. Moving to a denser node like TSMC's N2 lets designers fit more logic in the same area and switch it faster at the same power. Designers can spend that budget on bigger cores, higher clocks or longer battery life, and the mix chosen determines what the benchmark reveals. The contrast the report draws with the earlier N3B transition, which delivered under 10 percent, shows the gain depends on what the architects do with the node, not the node alone.
- A single early benchmark is a data point, not a verdict. Public benchmark databases accept uploads from pre-release devices, engineering samples and misconfigured runs. Reputable coverage flags this, as Tom's Hardware did, because thermal limits, firmware and sustained-load behaviour can move final results. Waiting for multiple consistent submissions from retail units is how you separate a trend from a fluke.
We teach this
How to use this in practice
- Benchmark the machine on your desk. Run a free CPU benchmark such as Geekbench or Cinebench on your own computer and record single-thread and multi-thread scores along with the core count and clock speed reported by the system. Run it twice, once on a power-saving profile or on battery, to see how much clock speed changes the result, and compare the ratio between the two scores with the numbers in this story. Done looks like a note with your scores and one sentence on whether your machine is core-limited or clock-limited for the work you do.
- Profile a task to see whether it is single-threaded. Open Task Manager or a similar monitor, run a task you do daily, such as a build, an export or a large spreadsheet recalculation, and watch whether one core pegs at 100 percent while the rest idle. Done looks like a screenshot and a note about which of your regular jobs would benefit from a faster core versus more cores.
- Sketch the path from memory to core. Draw a diagram with main memory, the last-level cache, per-core caches and the execution units, and label the bandwidth or latency order of magnitude at each step from any reference you trust. Add a note about what a wider memory bus changes. Done looks like a one-page diagram you could use to explain why bandwidth upgrades accompany core upgrades.
- Read one chip's spec sheet against a benchmark listing. Pick any processor you own, find its official specification page, and check that the core counts, clock speeds and cache sizes match what a public benchmark entry reports for it. Note any mismatch, which is often a sign of a throttled or misreported run, and repeat with a second listing for the same part to see how much run-to-run variation is normal. Done looks like a short comparison table with a verdict on whether the listing is trustworthy.
Sources
- Apple's A20 Pro shatters Geekbench 7 single-core record — 2nm chip beats desktop Intel Core i9 and AMD Ryzen 9 by up to 32% — Tom's Hardware
Our reporting is an original summary; full coverage is at the links above.
Don't just read about it — build it.
Square 1 teaches the skills behind the headlines, with every line of your work graded by AI. Find your starting point in 3 minutes.
Get your free skill report