Apple’s A20 Pro: Overhauling a 10‑Year Chip Architecture
On September 9, 2026, Apple performed this surgical redesign on the A20 Pro: it moved memory off the top of the processor and laid it out alongside the core, with a triple-area copper layer making direct contact with the compute die.
Benchmark runs on the phone last just thirty seconds, posting eye‑popping scores. Many users grab a new device and immediately fire up benchmark software, watching metrics climb to peak values and feeling reassured. But close the benchmark app, launch a high‑end AAA game, and the experience starts smoothly. Frames hold steady at max settings, gameplay feels fluid, and abilities render without stutter. Once the ten‑minute mark hits, however, everything falls apart. The back of the phone grows scalding hot, almost too warm to hold. The screen dims automatically without warning, and game frame rates plummet sharply. What was once smooth movement devolves into slow motion, the gameplay stuttering like a slide show.
This flagship chip that dominated benchmarks can barely sustain heavy gaming for ten minutes. Why? A quick online search yields the same simplistic answer: insufficient heat spreaders, cost-cutting by manufacturers, undersized heat pipes, or skimped thermal materials. That explanation is wrong. If you cut open the phone’s mainboard and examine it, you will find the real bottleneck is not the thickness of external heat sinks. It is a decades-old physical engineering problem.
Before unpacking this constraint, let us establish a clear benchmark: what triggers thermal throttling here? Not the chip itself. The processor can operate safely above 100°C. The hard limit comes from human touch. Manufacturers cannot let the chassis become too hot to handle, so they cap the outer casing temperature at roughly 40°C. Once the shell hits this threshold, power delivery to the chip must drop, even if the silicon still has headroom.
A phone’s sustained performance ceiling does not depend on stacking thicker heat spreaders. It depends on how easily heat travels from the silicon die to ambient air. This thermal path forms a series of linked segments; the section with the highest thermal resistance defines the overall limit. Let us examine the first millimeter of that path — where heat exits the core.
Slice open an iPhone mainboard from the past decade and view its cross-section under a microscope. The CPU and GPU, the primary heat-generating compute cores, are soldered onto the bottom substrate of the package. Sitting directly on top of them is memory, another heat-sensitive chip. This packaging scheme is known in semiconductor engineering as Package-on-Package, or PoP. Simply put, delicate memory chips stack vertically on top of the compute die, connected by tiny solder balls. The full stack measures just over one millimeter tall. When the compute core runs at full load, it acts as a blazing hot furnace.
Placing a hot furnace below a layer of memory is like covering a stove with a thick blanket. Apple’s chip designers understood this tradeoff perfectly. Ten years ago, PoP was a shrewd engineering compromise. Back then, the mainboard’s limited physical space was the primary constraint, not thermal dissipation.
Large batteries, expanding camera modules, and multiple antenna arrays all compete for internal space, leaving only a narrow region for the mainboard. If compute and memory chips sat side by side, the mainboard footprint would expand substantially. A larger board forces a smaller battery, hurting battery life and making the phone thicker and heavier.
Stacking memory vertically on top of the die saves precious horizontal board space by leveraging vertical volume. There was a second major advantage back then: signal latency and power consumption for data transfer. The CPU and memory constantly exchange data across thousands of micro interconnects. Stacked PoP uses extremely short vertical wires between the two chips. Shorter traces reduce signal travel time, cut latency, and lower power used to drive those signals.
In that era, Package-on-Package solved space constraints while minimizing data latency and power draw. This stacked architecture made thin, lightweight smartphones possible. It was not lazy design, but the optimal engineering compromise for its time.
Engineering tradeoffs shift as technology evolves. Over the past decade, mobile chip transistor counts have risen relentlessly, pushing up power draw and heat output under full load. The design that once enabled slim smartphones has now become a physical bottleneck.
Follow the heat flow to understand what happens during gaming. When you play, the underlying compute core runs at high load. For context: the chip can hit 12 watts during a 30-second benchmark burst, but sustained gaming power sits between 5 and 8 watts. Even at that power level, thermal limits kick in within ten minutes.
Heat generated by the die has three escape routes. First path: downward conduction. Heat moves through the package substrate, solder balls, and mainboard copper layers. This route spreads heat onto the PCB, but it is thin and convoluted, so thermal transfer is slow. Second path: lateral conduction. Copper traces in the substrate can spread heat sideways, yet the surrounding package area is packed with power management ICs and tiny capacitors, leaving little room for heat to disperse. Third path, the primary route: upward conduction. Heat rising from the die travels through the shield can, heat pipe, graphite film, chassis, and finally into open air. This upward path is the main thermal highway for phone cooling. But heat moving upward immediately hits the memory chip stacked on top. The first millimeter of the main thermal route — the widest potential highway — is blocked by memory with poor thermal conductivity.
Compounding the issue: different chip components have very different temperature tolerances. The underlying compute cores generate massive heat but only perform calculations. They can operate reliably above 90°C. The memory stacked above cannot. Its safe upper limit is roughly 85°C. The difference comes down to how each component works. Memory stores data using tiny capacitors that hold electric charge to represent 0s and 1s.
Capacitors naturally leak charge. Higher temperatures accelerate molecular motion in semiconductors and speed up leakage. This is the fundamental distinction between compute cores and memory: cores process data and discard it after calculation, while memory must retain data continuously. When charge bleeds away too quickly, memory increases refresh rates to recharge capacitors. Once temperatures exceed 85°C, refresh rates double, consuming extra power and reducing the power budget available for the compute core.
Combine these factors. During extended gaming, the compute core runs continuously, and heat accumulates because the primary upward thermal path is blocked by memory. Heat builds up and reaches the phone’s outer shell. When casing temperature hits the 40°C threshold, the system enforces power reduction. Important point: thermal throttling is triggered neither by CPU overheating nor memory exceeding its temperature rating. The entire device’s thermal budget is exhausted.
Memory is not the sensor that triggers throttling. It narrows the first millimeter of the thermal path and lowers the overall system’s sustainable temperature limit. Once core clock speeds drop, compute performance collapses and gameplay stutters. The phone also dims the screen to cut power and heat generation, squeezing extra thermal headroom from this congested system.
This explains why flagship phones suffer sudden performance crashes after ten minutes of gaming. The root cause is not poor software tuning, but a physical bottleneck in that first millimeter of heat transfer. Many cooling marketing claims circulate across the mobile industry: aerospace-grade graphene films, large dual-layer heat pipes, even titanium frames marketed to boost thermal dissipation.
These technologies sound like they will solve phone overheating, yet real-world gaming still leads to hot devices and frame drops after ten minutes. These cooling solutions are not pure marketing hype. The iPhone 17 Pro (2025) introduced heat pipes for the first time. It still used stacked PoP memory, yet its gaming performance improved notably, nearly 70% in some tests.
New silicon contributed to the gain, and heat pipes delivered real benefits. But all these upgrades improve only the latter portion of the thermal path. Where are those large graphene sheets mounted? On metal shields and the back of the battery. Oversized heat pipes cover the outer sections of the mainboard. Titanium or aluminum chassis sit at the very edge of the device.
Notice the pattern: graphene, heat pipes, and metal chassis all manage heat after it has already left the chip. Imagine a pipeline carrying heat from the die to ambient air. Upgrading the downstream pipe diameter helps flow slightly, but a narrow constriction remains upstream. The total throughput of the whole pipeline is governed by that tight choke point.
Only removing the bottleneck unlocks the full thermal path. Many users assume overheating and frame drops stem from insufficient cooling hardware. The reality: optimizations target downstream heat transfer, while the critical first-millimeter bottleneck remained untouched for more than a decade. Once the bottleneck is identified, the solution becomes clear. Downstream thermal tweaks still offer marginal gains, but their returns are diminishing.
As long as memory sits stacked on top of the compute die, the upstream choke point persists. Adding extra copper layers or modifying the chassis will always suffer thermal losses across that stacked memory barrier. To maintain high performance over long durations, one radical hardware change is required: remove the first-millimeter thermal bottleneck.
At Apple’s September 9, 2026 keynote, Apple delivered this redesign with the new A20 Pro chip. What changed? Apple adopted a design previously reserved for Mac processors: memory is fully removed from the top of the CPU and placed alongside the compute die. Apple mentioned this change during the presentation, stating the new package uses custom packaging technology derived from Mac silicon to enable direct die-to-memory interconnects.
Earlier leaked iPhone 18 Pro mainboard schematics already showed memory placed beside the compute core using TSMC wafer-level packaging. Apple did not explicitly phrase it as “relocating memory,” but two lines of evidence support the redesign. This semiconductor technique is called Wafer-level Multi-Chip Module, or WMC. The compute and memory dies are interconnected while still on the raw silicon wafer, without intermediate interposers or packaging substrates.
The old stacked two-die layout becomes a side-by-side planar arrangement. Apple’s M-series Mac chips have used this layout for years, with the core die centered and memory mounted on either side. This marks the first time Apple brought this Mac-style architecture into an iPhone. With memory relocated, heat transfer dynamics transform completely.
The thick “insulating blanket” over the CPU and GPU is removed. The hottest compute hardware is exposed directly at the top of the package. The new heat pipe bonds directly to the compute die, with triple-area copper layers contacting the core without a memory chip in between. Heat generated under sustained load travels unimpeded through that first millimeter straight into the copper heat pipe.
The heat pipe spreads high-temperature heat evenly across the mainboard. The relocated side-mounted memory no longer sits under constant thermal bombardment from the hot compute core. The tightest thermal bottleneck is eliminated, lowering overall thermal resistance of the full path. With the same 40°C chassis temperature limit, the phone can sustain a higher continuous power level.
After resolving that first-millimeter constraint, Apple’s September 9, 2026 keynote shifted its focus away from short burst benchmark scores. The presentation highlighted a 40% improvement in sustained performance and doubled performance compared to products released two years prior. Apple’s confidence to market sustained performance as the key selling point comes from moving memory off the compute die. Peak benchmark performance is like a 30-second sprint; sustained performance means stable long-run operation after removing thermal barriers.
If side-by-side memory delivers such major advantages, why did Apple not adopt this design years ago? Apple engineers certainly considered it, but there was a prohibitive cost: die area. Stacked PoP occupies one unit of board area. Side-by-side placement requires space for both dies.
iPhone internal space is already maxed out. Batteries and camera modules cannot shrink, so widening the mainboard was not feasible. Leaked PCB drawings challenge that intuitive assumption. Measurements show the A20 Pro package footprint is nearly identical to the previous A19 Pro. Memory moved to the side, yet total package size stayed unchanged. Where is the recovered space? It comes from removing intermediate packaging layers.
Traditional packages require interposers and stacked substrate layers. Wafer-level packaging eliminates those components by interconnecting dies while still on the wafer. The extra horizontal space needed for side-mounted memory is offset by the removal of stacked intermediate layers. This redesign became viable not because the compute die shrank, but because wafer-level packaging finally made the area tradeoff practical.
What role does the 2nm process play? Many viewers focus on raw performance gains. For a mobile device capped by chassis temperature limits, however, the most critical metric is performance per watt. Power budget is fixed. 2nm manufacturing delivers more compute from each watt of electricity. Sustained performance fundamentally boils down to performance per watt.
What about the latency tradeoff? Moving memory to the side lengthens interconnect traces, erasing some latency advantages of the old PoP stack. Apple did not detail this point, and independent test data is not available. The area equation balances out, yet planar packaging costs more and is harder to manufacture. What demand justifies this high investment?
The motivation extends beyond smoother gaming. The core driver is on-device large language models. In 2026, deploying billion-parameter large models locally on phones became the industry’s primary trend. This workload differs drastically from gaming or video playback. When you run deep inference on-device, every generated token requires loading billions of weight parameters from memory and feeding them into compute units.
Generating multiple tokens per second creates massive data movement. Many assume slow on-device AI comes from insufficient compute power. That is incorrect. For large model inference, the bottleneck is usually memory bandwidth, not raw compute throughput. Even abundant CPU/GPU cores sit idle waiting for data if memory interconnects cannot supply parameters fast enough. This is the well-known memory wall. Memory bandwidth defines the upper limit of on-device AI performance.
Apple implemented two key changes in the A20 Pro to address this: first, a 50% increase in memory bandwidth; second, a dual 32-core Neural Engine. The 50% bandwidth upgrade is not arbitrary. Under the old PoP stack, memory connected to the die through a limited array of solder balls. Solder ball count caps the number of interconnect traces and restricts memory bus width. Widening the bus was impossible with stacked memory.
Relocating memory to the side removes that physical constraint, enabling the wider memory bus. Apple stated in the keynote this is the widest memory interface ever used in an iPhone. Leaked board specifications list a 96-bit bus. Moving memory and expanding memory bandwidth are two outcomes of the same architectural redesign.
The second major benefit from relocating memory emerges under long-duration heavy workloads. On-device large models are not short benchmark sprints. Complex reasoning and document summarization can run continuously for minutes, requiring sustained high load. The 32-core Neural Engine runs continuously while memory streams data across the upgraded bus, generating significantly more heat than traditional workloads.
Without this redesign, keeping memory stacked on the hot compute die would trap heat in the first millimeter. Chassis temperature would quickly hit the thermal limit, forcing power cuts and stalling the large model. This shift was not a voluntary upgrade; it was forced by technical requirements. Sustained high heat from local AI demanded memory relocation and an open thermal path.
These connected engineering choices make stable local large model inference on iPhones possible. Returning to the original ten-minute bottleneck: mobile phones of the past decade were built for 30-second benchmark sprints. Extended use traps heat in the upstream thermal path, triggering chassis temperature limits and sharp throttling. Removing that first-millimeter barrier lets heat flow freely out of the package, turning the phone into a device capable of sustained high performance.
One important disclaimer: all figures cited here — 2nm process, 40% sustained performance uplift, doubled performance versus two-year-old hardware, 50% higher memory bandwidth, 32-core Neural Engine, triple-area copper direct die contact — are Apple official claims from the September 9, 2026 keynote. Independent benchmarking and full device teardowns have not yet been released.
If you enjoy what I do, consider supporting me! Every little bit means the world! ❤️
☕ Buy me a coffee