01
Xu et al. · 2024 · arXiv
Why?
Map of the field before the numbers: compression, hardware acceleration, and edge–cloud tradeoffs. Read first so the measurement papers have a place to hang.
Loading…
One moment while we fetch this page.
Measuring LLM throughput, energy, and thermals on Android phones — what the first mobile harnesses established, and what peak tok/s hides under sustained load.
For
Engineers shipping or profiling on-device LLMs on Android who need a short map of what to measure and which papers already measured it.
Covers
On-device LLM survey map, MELT measurement vocabulary, COTS SoC comparisons with llama.cpp / MLC-LLM, and sustained thermal load including Galaxy S24.
Leaves out
Training or fine-tuning on phone, GUI agents (see Evaluating Mobile-Use Agents on Android), NPU systems papers beyond the measurement arc, or cloud serving.
The full collection in the community’s reading order.
01
Xu et al. · 2024 · arXiv
Why?
Map of the field before the numbers: compression, hardware acceleration, and edge–cloud tradeoffs. Read first so the measurement papers have a place to hang.
02
Laskaridis et al. · 2024 · MobiCom
Why?
The first systematic on-device LLM study with a real harness (Android, iOS, Jetson): throughput, memory, energy, accuracy, and thermal behaviour. Sets the measurement vocabulary everyone else inherits.
03
Xiao et al. · 2024 · TMC
Why?
Deep COTS SoC comparison (Qualcomm, MediaTek, HiSilicon, Apple) with llama.cpp and MLC-LLM: 4-bit as the practical sweet spot, Armv9 instruction wins, Adreno vs Mali, and DVFS under repeated runs.
04
Tummalapalli et al. · 2026 · arXiv
Why?
Twenty warm iterations of the same prompt on Galaxy S24 Ultra, iPhone 16 Pro, Hailo-10H, and a laptop GPU. Peak tok/s lies; thermal floors and energy proportionality are the real deployment numbers.