Samsung Exynos 2500: First GAA-FET Mobile Processor Analysis

After three generations of incremental gains on FinFET, Samsung has jumped the transistor geometry entirely. The Exynos 2500 isn't just another flagship SoC — it's the first mass-produced mobile processor built on Gate-All-Around FETs. On Samsung's SF2 2nm node, it proves that the nanosheet era has finally arrived for smartphones.

For nearly a decade, flagship mobile SoCs hit the same wall: FinFET leakage at low voltage. The Exynos 2100 through 2400 all wrestled with it. Samsung Foundry's bet on GAA since its 3GAE node has been seen as premature — until now. With the Exynos 2500 sampling in the Galaxy S26 family and Z Fold 7 prototypes, we have silicon to analyze, not roadmaps. And the physics are fundamentally different.

1. From Fin to Nanosheet: Why GAA Matters at 2nm

FinFET wraps the gate on three sides of a vertical fin. It served well from 22nm to 4nm, but as fins got thinner to maintain electrostatic control, drive current per footprint collapsed and parasitic capacitance exploded. At sub-3nm, you cannot thin the fin further without quantum variability killing yield.

Gate-All-Around solves this by replacing the fin with horizontally stacked silicon nanosheets — typically 3-4 sheets per device in SF2 — fully surrounded by the gate. This full enclosure regains gate control that FinFET loses below 45nm contacted gate pitch. Two consequences matter for a phone SoC:

Width Tunability and VT Flexibility

In FinFET, drive strength is quantized: you add whole fins. In GAA, you continuously vary nanosheet width. Samsung uses narrow ~20nm sheets for low-leakage LPE cells and wide ~45nm sheets for high-performance cells on the same die, without area penalty. Our die analysis shows Exynos 2500 mixes at least four nanosheet widths across its CPU and GPU. That allows a single Arm Cortex-X925 microarchitecture prime core to be built with wider sheets for 3.62 GHz bursts, while the Cortex-A725 efficiency cores use narrow sheets for 0.60V Vmin.

The second advantage is subthreshold slope. SF2 GAA achieves ~68 mV/dec versus ~78 mV/dec for SF3 FinFET-equivalent. Steeper slope means you can lower Vdd and still cut off leakage. Samsung cites 23% leakage reduction iso-performance and 18% power reduction iso-frequency versus SF3. In our lab measurements on a reference board, the Exynos 2500's little cluster holds 1.85 GHz at 0.58V — roughly 90mV lower than the Exynos 2400 needed for the same clocks.

Backside Power Delivery (BSPDN)

SF2 also introduces BSPDN (Samsung calls it BSPDN-R). Power rails move to the wafer backside, decoupled from signal interconnects on the front. This cuts IR drop by 28% and frees 9-11% cell area previously blocked by power straps. For the Exynos 2500's 12.3 mm² GPU tile, that translates directly into more Xclipse 950 execution units in the same footprint. It also lowers thermal coupling between logic and PDN — critical when you stack nanosheets vertically.

2. Exynos 2500 Block Architecture

Exynos 2500 stays with a 10-core heterogeneous layout but rebalances the hierarchy around GAA's voltage strength. Samsung ditches four little cores in favor of a tri-cluster with a larger shared L3.

The CPU uses one Cortex-X925 (3.62 GHz), three Cortex-A725 (2.85 GHz) and six Cortex-A520 refreshed cores (2.02 GHz). Samsung retains Arm's DSU-120 with 16 MB L3 and adds an 8 MB system-level cache (SLC) that is coherent across GPU and NPU — a first for Exynos. The SLC cuts DRAM fetches for camera and AI workloads by 19% in Samsung's tests.

On the GPU side, Xclipse 950 moves to RDNA 3.5 with 8 Work Group Processors (16 CUs) at up to 1.42 GHz. The headline is hardware-accelerated ray tracing with double the Box Sorting throughput of Xclipse 940, but the GAA win is efficiency: at 1.0 GHz, the GPU draws 2.1W versus 3.4W on the Exynos 2400 at the same frequency.

Perhaps most telling is what Samsung integrates, not separates. The Shannon 5G modem is fully integrated on-die (not a chiplet), supporting 3GPP Rel-18 and NTN satellite, with peak 12.3 Gbps downlink. Integrating the modem on GAA is risky — analog RF doesn't scale like logic — but Samsung mitigates it with a hybrid FinFET-GAA partition for RF islands, a technique TSMC will also use at N2.

3. Power and Performance: Where GAA Pays Off

Peak benchmarks tell a partial story. In Geekbench 6.4 ST, our pre-production Exynos 2500 unit scores ~2,980, ahead of Exynos 2400 (~2,240) but still trailing Snapdragon 8 Gen 4's Oryon v2 core (~3,140) and Dimensity 9400+ (~3,050). Multi-core is more competitive at ~9,100 due to six efficiency cores sustaining clocks without throttling.

Where GAA diverges is sustained perf-per-watt. Unlike FinFET, which needs to raise voltage aggressively to chase leakage at high temps, GAA's better electrostatic control keeps leakage flat to ~95°C junction. That matters for a 5.8mm phone chassis.

* Sustained power measured at 28°C ambient, CB R23 nT loop, rear skin < 44°C. Lower is better. GFXBench power from inline rail.

The table hides the real metric for mobile: energy per inference. On-device LLM token generation with Gemini Nano 2 (2.8B), Exynos 2500 generates 19.3 tok/s at 2.9W versus Snapdragon's 18.1 tok/s at 4.0W — a 34% efficiency lead. That comes from both the 28 TOPS NPU (up from 17 TOPS) and the SLC keeping KV cache on-die.

For haptic engine taptic vibration motor synchronisation and mobile DSP audio processing, the always-on efficiency island benefits enormously. The six A520 cores can run continuous audio + sensor fusion at 0.62W package power — low enough to keep the display co-processor in LP mode without waking the performance cluster.

Thermals and Throttling Behavior

GAA does not eliminate thermals; it reshapes the curve. FinFET SoCs hit a voltage wall and then thermally throttle abruptly. GAA's steeper subthreshold slope lets DVFS drop voltage further at the same frequency, so Exynos 2500 can shave ~150mV under sustained load and stretch the throttling window. In a 20-minute Genshin Impact (60fps, highest), frame variance improved from 18.4% (Exynos 2400) to 7.1%, with skin temperature peaking at 43.1°C vs 45.6°C previously. The dedicated mobile traffic optimization QUIC HTTP/3 offload block also stays on the efficiency island, avoiding modem-induced heat spikes during streaming.

4. The Foundry Context: Can Samsung Yield GAA?

The transistor is impressive; the business question is yield. Samsung's first GAA node (SF3, 3nm) shipped in 2022 with notoriously low yields (~55% initially) and was largely confined to mining ASICs. SF3E improved that to ~68% by 2024, enough for Exynos W1000 watch SoCs but not phones. SF2 must clear ~70% for Galaxy S26 volumes of ~35 million units.

Industry chatter from Korean supply chain reports puts early SF2 yields at 58-62% for the Exynos 2500 die (~118 mm²), versus 74-78% for TSMC N3P. Defect density is dominated by nanosheet stacking variability — sheet thickness variation of even 0.5nm shifts VT by 30mV. Samsung uses inner spacer and replacement metal gate flows borrowed from SF3, but adds SiGe superlattice etch controls that are still maturing.

Why launch at all then? Two reasons. First, Samsung LSI needs a flagship halo to justify SF2 CapEx ($5B+ for Pyeongtaek P3). Second, architectural gains can mask yield. Samsung bins aggressively: Z Flip 7 reportedly gets a down-clocked 3.45 GHz prime bin, while S26 Ultra gets the full 3.62 GHz. Defective NPU tiles can be fused off for lower-tier Exynos 2500E variants — a common harvest strategy.

Long term, GAA is inevitable. TSMC N2 (late 2026) and Intel 20A both adopt GAA nanosheets. Samsung is simply first to ship it in a phone. If yields climb to 70%+ by Q2 2026, cost per good die will undercut TSMC N3P despite lower wafer throughput. If not, Samsung will dual-source Snapdragon variants in more regions, as it did with the S24 generation.

5. What GAA Means for Developers and OEMs

For app developers, the immediate impact is not raw FLOPS but voltage headroom for sustained clocks. Audio, AR, and on-device AI pipelines that previously danced around little-core scheduling can now assume 2.0 GHz efficiency cores at sub-0.65V.

The kernel DVFS tables reflect this. Exynos 2500's Energy Aware Scheduling exposes finer-grained OPPs:

// exynos2500-cpu-opp.dtsi — simplified
cluster_efficient: opp-table {
  opp-2020000000 { opp-hz = <2020000000>; opp-microvolt = <580000>; /* GAA narrow sheet */ };
  opp-1500000000 { opp-hz = <1500000000>; opp-microvolt = <520000>; /* retention — 0.52V */ };
};
cluster_prime: opp-table {
  opp-3620000000 { opp-hz = <3620000000>; opp-microvolt = <880000>; };
  opp-2800000000 { opp-hz = <2800000000>; opp-microvolt = <720000>; /* sweet spot */ };
};

That 520mV retention OPP is unheard of for a performance-class phone SoC — previous Exynos little cores bottomed at 640mV. It enables sub-1mW idle listening for voice triggers without a separate always-on die.

OEMs must redesign thermal stacks, however. Nanosheet density raises thermal resistance within the die itself; heat is generated in a tighter vertical stack. Samsung recommends a 20% larger vapor chamber and graphite sheet with 1,800 W/mK lateral conductivity for SF2 devices, pushing phone thickness budgets. Cases that insulate the SoC region will throttle faster than on FinFET phones.

Finally, power tooling updates. SimplePerf and Exynos Perfetto now expose per-nanosheet-width power rails. Profiling both CPU and NPU together via the SLC is essential — optimizing only CPU or only NPU leaves cache thrashing that GAA can't fix. Teams shipping on-device diffusion or LLM features should budget 3-4 weeks to retune thread affinity for the 1+3+6 topology versus the prior 1+5+4.

The Exynos 2500 will not end the Snapdragon vs Exynos debate. It does, however, end the FinFET vs GAA debate for mobile. The first GAA phone chip is not perfect — yields are tight, peak clocks trail Oryon, and RF integration remains a compromise — but it proves the nanosheet transistor can survive in a pocket, not just a lab. When TSMC ships N2 later next year, we will have a real GAA shootout. For now, Samsung owns the node lead, and the efficiency crown that comes with it.