The A8 SoC
At the heart of the new iPhones beats a new system-on-a-chip (SoC), the Apple A8. The A8 is at least an incremental upgrade over the A7 SoC from the iPhone 5S, although Apple is fairly coy about the exact details of its silicon designs.
iPhone 6 Plus
|SoC||Apple A6||Apple A7||Apple A8|
|Die size||97 mm²||104 mm²||89 mm²|
|Transistor count||?||>1 billion||2 billion|
|Manufacturing process||32 nm||28 nm||20 nm|
|CPU cores||2 Swift||2 Cyclone||2 "Cyclone++"|
|CPU die area||14.7 mm²||17.1 mm²||12.2 mm²|
|Max core frequency||1.3GHz||1.3GHz||1.4GHz|
|System memory||1GB LPDDR2||1GB LPDDR3||1GB LPDDR3|
|Coprocessors||-||M7 motion coprocessor||M8 motion coprocessor|
Outside of Apple's own statements, the best sources of info on the A8 SoC are this teardown analysis by ChipWorks and Ryan Smith's take on it. The A8 chip has been widely reported to be manufactured at TSMC on a 20-nm fabrication process, and the A8's considerably smaller die size lends credibility to those reports.
We know that the A7 has two copies of Apple's own custom CPU core, dubbed Cyclone. Cyclone is a high-IPC core compatible with the 64-bit ARMv8 instruction set. The A8 SoC in the iPhone 6 and 6 Plus also has dual cores, and those cores are clocked only 100MHz higher than in the A7 in the iPhone 5S. Indications point to these cores being a tweaked version of Cyclone. The size of the CPU cores on the die hasn't dropped commensurately with expectations in light of the die shrink, so the cores themselves must contain more complex logic, larger structures, or both. These changes are likely intended to improve per-clock instruction throughput.
The locations of the the L2 SRAM arrays in the A8's floorplan have changed from the A7, and Chipworks speculates that Apple may have moved to 1MB of dedicated L2 cache per core. However, the more probable reason for the change isn't a size increase but increased modularity. Each core now appears to have its own "slice" of associated L2 cache, so that the cache size can scale up with the core count. Such an arrangement would make sense in light of the fact that the A8X SoC used in the iPad Air 2 has three CPU cores.
Another set of SRAM arrays on the chip is likely a last-level cache, probably 4MB in size and possibly shared not just between the CPU and GPU blocks but by the whole of the SoC. This LLC is also present in the A7.
That's all good news, in my book. While many of its competitors have taken the path of increasing core counts in their latest SoCs, Apple has built one of the highest-throughput mobile CPU cores anywhere. We know even from big desktop PCs that the user experience is often dominated by the performance of one big, hairy thread that's difficult to execute. Apple's decision to pursue higher per-thread performance instead of expanding the core count seems like the smart course.
SoC and CPU performance
Comparing SoC performance across platforms isn't easy, but we do have a handful of reasonably useful benchmarks we can employ. We're spoiled by the extensive instrumentation and easy scripting of the PC platform, I suppose, but many of the tests below (and on the following pages) don't do the work we'd prefer they did. For instance, many mobile benchmarks simply report synthetic scores without reference to anything concrete, like a rate of computation or the time to completion of a task. Where possible, we have reported the results below in concrete units, even if those aren't the most commonly quoted numbers you might see elsewhere.
Also, mobile benchmarking is fraught with shenanigans related to power management policies. We've not yet mapped out this space well enough to effectively counter some of the benchmark detection efforts phone makers have been known to use. That said, all of the numbers we've reported are the median of at least three test runs, and we've discarded any major outliers from the pool of results.
We have several notable devices on hand to compare with the iPhone 6 and 6 Plus. Those include three prior generations of Apple offerings, the iPhone 5S through the iPhone 4. The OnePlus One and LG G3 are competing large-format Android phones, both based on Qualcomm's Snapdragon 801 SoC with quad "Krait" custom CPU cores and Adreno 330 graphics.
Beyond phones, we have Nvidia's Shield Tablet, based on a Tegra K1 GPU with quad Cortex-A15 CPU cores and Kepler-class graphics. This device is an 8" tablet with a larger power envelope than a smartphone, but the architectural comparison should be interesting, with that caveat kept in mind. Also on hand is the Asus Memo Pad ME176C, a low-cost tablet based on Intel's Atom "Bay Trail" Z3745 SoC; this SoC features quad "Silvermont" cores and Intel HD Graphics.
Many of you are probably hoping for comparisons against one or two other significant competitors in the high-end smartphone space. We don't have those results below, but stay tuned.
We don't know the exact DRAM configuration of the new iPhones. Since they use LPDDR3, the channels are narrower than in desktop memories, either 16 or 32 bits wide. The Stream results suggest the possibility of dual 32-bit memory channels running at 1333 MT/s, if the SoC is squeezing out every last drop of bandwidth. There's some warrant in the die images for the presence of dual SDRAM interfaces in the A8's I/O ring.
We do know that the A8's CPU cores are the only ones here that appear to reach the peak potential of the SoC's memory subsystem in Stream's copy test using a single software thread. Also, even with multiple threads, none of the other devices can match the peak transfer rates of the iPhone 6 and 6 Plus.
Geekbench runs natively on both iOS and Android, and it offers us a look at performance with just a single thread and with multiple threads. You can click on the button below to toggle between our single- and multi-threaded results.
Apple has carved out a substantial lead for itself in smartphone CPU performance, at least among the contenders we've tested. The A8 is far and away the fastest SoC here in single-threaded performance. What's more, even though the A8 has only two cores on tap, its multithreaded performance is more than competitive. Only the Shield Tablet, with four Cortex-A15s in a larger power envelope, outperforms the A8 in Geekbench's multithreaded tests.
|CPU clock frequency||1.3 GHz||1.4 GHz||7.7%|
|Geekbench floating point||1341||1578||18%|
The A8's single-threaded CPU performance has risen by 16% overall in Geekbench, with the gains coming in both integer and floating-point math. As the table above illustrates, the performance improvements outstrip the increase in CPU clock frequency. The A8 may have a new dynamic frequency boost mode we don't know about, but in all likelihood, its CPU cores have been tweaked for increased per-clock throughput.
Geekbench has a ton of component tests, but I'd like to call out one especially interesting result. The AES encryption test illustrates the impact of tailored acceleration instructions built into the ARMv8 instruction set. Apple's A7 and A8 SoCs are the only beneficiaries among the devices we've tested, but one can expect to see a similar boost in this test for other ARMv8-compatible SoCs.
Obviously, the iPhones' huge lead in this test is a bit unusual. I was concerned that the results from this component test would throw off Geekbench's overall integer and composite scores. After a little noodling around in a spreadsheet, though, I'm satisfied. Turns out Geekbench uses a geometric mean to compute its overall indexes, so outlier scores shouldn't have an outsized impact on the results.
The new iPhones continue to perform well in these cross-platform, browser-based benchmarks. The closest competitors here, the L3 G3 and OnePlus One based on Qualcomm's Snapdragon 801, are clearly outclassed.
Speaking of which, I've quietly slipped in some results from a couple of desktop processors, just to illustrate how close these mobile SoCs come to matching x86 CPUs with power envelopes nearly an order of magnitude higher. Remarkable, really.
BaseMark OS II
Intel has evangelized WebXprt pretty enthusiastically. We don't always like it when a company backs a particular benchmark, but given Intel's history there, I'm sure there's nothing to worry about. Right?
Regardless, the new iPhones take the top spot in WepXprt's overall index thanks to strong performance in each workload. However you slice it, really, the A8 SoC has some of the highest CPU performance of any mobile device we've tested.