LLM Benchmarking on ODROID-XU4: Why 4 Cores Beat 8 Cores

1. The Asymmetric (big.LITTLE) Architecture

The Samsung Exynos 5422 SoC inside the ODROID-XU4 features 8 total physical cores split into two distinct clusters:

Cluster	Cores	CPU Architecture	Relative Speed
LITTLE Cores	0 – 3	Cortex-A7	Low Power (BogoMIPS ~48)
big Cores	4 – 7	Cortex-A15	High Performance (BogoMIPS ~108)

Thread Synchronization Barrier: During matrix multiplication in llama.cpp, all active threads must sync up after computing each layer. If you run 8 threads (-t 8), the 4 fast Cortex-A15 cores finish early and sit idle waiting for the 4 slow Cortex-A7 cores to finish their calculations.
Memory Bandwidth Contention: Running all 8 cores creates a bottleneck on the system's memory bus, starving the fast cores of the data throughput required for fast inference.
Thermal Throttling: Maximizing all 8 cores rapidly generates excessive heat, causing the SoC to automatically reduce clock speeds to prevent overheating.

To maximize tokens-per-second, use taskset to bind the execution exclusively to the 4 high-performance cores (Cores 4–7):

taskset -c 4-7 ./build/bin/llama-bench -m qwen0.5b.gguf -t 4 -r 1 --no-warmup