Three chips in and Google Tensor is on life support
Google’s custom Tensor chips in Pixel phones are being criticized for weak performance, poor power efficiency, and inferior modems compared to Qualcomm-based rivals, leading many AI features to run in the cloud rather than on-device as originally touted. Commenters debate whether this is primarily a technical lag, a consequence of Google’s business incentives to keep processing and data in its data centers, or simply a normal misalignment between fast-moving AI models and slow silicon design cycles. Some users report smooth UI and useful ML features, but many cite overheating, bad battery life, and unreliable connectivity as reasons to avoid current Pixels or wish Google would return to Snapdragon.
On‑device vs. cloud AI and privacy
- Many see Google’s continued cloud offload for new AI features as driven by data collection, not just hardware limits.
- Others note Google could process locally and upload summaries later, but that full raw-data uploads after local processing would look worse optically.
- Counterpoint: running everything in the cloud is expensive; compute costs and GPU/TPU scarcity may push vendors toward more on‑device work over time.
Tensor performance and design trade‑offs
- Broad consensus: Tensor has underwhelming general performance and efficiency versus competitors, especially Qualcomm.
- In ML, some argue decisions were made years before the current LLM boom; large, high‑quality models naturally favor the cloud.
- Some criticize the article’s “on life support” framing as clickbait; others say missing Google’s own “on‑device AI” promises makes the phrase fair.
Battery life, thermals, and modem issues
- Multiple reports of poor battery life and overheating on Pixel 6–8 series, often linked to Samsung modems and weak reception.
- Some users have acceptable or good battery life, suggesting workload, coverage, background apps, and ROMs matter.
- Several returned Pixels mainly due to modem power draw and unreliable data connectivity; potential buyers for GrapheneOS are reconsidering.
User experience of ML features
- Many praise Pixel UI smoothness, speech‑to‑text, call handling, and camera features as “class‑leading” in daily use.
- Others see most ML features as gimmicks compared with fundamentals like reception and battery.
- Apple’s on‑device ML (OCR, photo search, face/pet recognition) is cited as a strong example of practical, private features that users heavily rely on.
Business and ecosystem considerations
- Maintaining both on‑device and cloud code paths is seen as expensive; product teams may prioritize one path that scales to all Android devices.
- Google wants AI features on non‑Pixel phones to drive subscriptions and ecosystem usage, not just hardware sales.
- Qualcomm’s licensing power and Apple’s much greater leverage are mentioned as reasons Google opted for its own SoC + Samsung modem despite trade‑offs.
AI hardware and model constraints
- Several comments explain that ML/“AI” chips are highly parallel vector/matrix units optimized for low‑precision math and bandwidth, not general LLM workloads.
- LLMs are described as heavily memory‑bandwidth‑bound and too large for current phones; smaller local models vs. big cloud models is an explicit trade‑off.