Qualcomm spent its summit in Maui this week talking less about how fast the Snapdragon 845 is and more about what it can think about, and that shift is the actual news. The performance bump is real and modest, roughly a quarter faster on the CPU and graphics over the 835. But Qualcomm built the whole presentation around five pillars, and the loudest one was AI. For a company that sold raw benchmarks for a decade, deciding that on-device intelligence is the headline feature is a real change in how the fight gets fought, and it is happening because the whole industry just pivoted at once.
The silicon underneath is a maturing of the 835 rather than a reinvention, and that is fine. The Kryo 385 CPU is built on ARM’s new DynamIQ cluster design, four performance cores around 2.8 GHz and four efficiency cores, and for the first time a Snapdragon carries a shared L3 cache, which is the DynamIQ giveaway and the thing that lets the core types share work and memory more cleanly. It stays on Samsung’s 10nm, the second-generation low-power-plus version, so this is efficiency squeezed out of a known node rather than a risky new one, which after the 810 I am happy to see. The Adreno 630 is the visual subsystem, 30% up on graphics and built for room-scale VR with six-degrees-of-freedom tracking and foveated rendering. The Spectra 280 image processor captures far more high-dynamic-range color information than before and shoots 4K in HDR. The X20 modem pushes Gigabit LTE to 1.2 Gbps. Solid, expected, well-executed.
Two things break the pattern. The first is the Hexagon 685 DSP, which Qualcomm is now positioning as an AI engine, claiming roughly three times the on-device machine-learning throughput of the 835. The pitch is that your phone runs the neural-network work, the camera scene detection, the voice processing, the photo tricks, locally instead of shipping it to a data center. The second is a brand-new Secure Processing Unit, a vault-like block physically separate from the main CPU, handling biometrics, mobile payments, and an integrated SIM with hardware-level isolation. Both are Qualcomm reading the same room everyone else just walked into.
Because here is why the AI framing is suddenly everywhere: Apple put a Neural Engine in the A11 this fall, Huawei put a dedicated NPU in the Kirin 970, and Google is doing on-device machine learning in the Pixel. The entire premium tier decided within a few months that AI acceleration is a checkbox flagships need, a long way from the connectivity-and-VR mood that filled the CES floor in January. Qualcomm’s approach is different in an interesting way, it is leaning on the Hexagon DSP it already had rather than bolting on a separate neural block, arguing that a flexible DSP plus CPU plus GPU working together beats a dedicated unit. I am not sure that bet is right. A purpose-built NPU may end up more efficient than a repurposed DSP for this work, and we will not know until real phones run real models. But it is a coherent position and it costs Qualcomm less die area, so I see the logic.
The 845 also doubles as the chip for the things Qualcomm announced it was chasing last year. The first Always Connected PCs running Windows on Snapdragon, the ASUS NovaGo and HP’s Envy x2, got shown at this same summit, and Qualcomm is explicitly pitching the 845 at XR headsets and laptops alongside phones. The current 835 flagships, the OnePlus 5 and 5T among them, are barely settled in, and the first 845 phones already land early next year, with Samsung’s next Galaxy and Xiaomi’s next flagship confirmed. I am parking the Apple royalty lawsuit that is hanging over all of this, because that fight deserves its own post and would swallow this one whole.
What I keep coming back to is that the 845 is the moment Qualcomm’s marketing stopped being about megahertz and started being about what the chip understands (finally, and I am all in into that, usage and needs before benchmarking). The hardware is a clean, confident iteration on a node Qualcomm has figured out. The story Qualcomm wants told is that the phone is becoming a thinking device and Snapdragon is the brain. Whether the DSP-as-AI-engine approach beats Apple’s and Huawei’s dedicated silicon is the question I would bet the next two years on, and I do not know the answer yet.