Skip to content

Llama Cpp Benchmarks

valhala llama.cpp benchmark — 08.jul.2026

Section titled “valhala llama.cpp benchmark — 08.jul.2026”

Hardware: HP ML350p Gen8, 2× Xeon E5-2620 (Sandy Bridge, AVX1-only), 12 threads, 36 GB installed / 12.5 GB usable (legacy BIOS MMIO hole), FreeBSD 15.1, neo RAIDZ1 on P420i SAS

Build: llama.cpp b9940, source-built with -DGGML_NATIVE=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DGGML_F16C=OFF -DGGML_BMI2=OFF

Metric14B R1 Distill Qwen Q5_K_M8B R1-0528-Qwen3 Q5_K_MRatio
File size10.5 GB5.7 GB1.8×
RSS at rest~10.2 GB~6.0 GB1.7×
Prompt processing~430 ms/tok (2.3 tok/s)~200 ms/tok (5.0 tok/s)2.2×
Generation~968 ms/tok (1.0 tok/s)~555 ms/tok (1.8 tok/s)1.7×
100-token response~100 sec~56 sec1.8×
Long-context prompt (97 tok)194 ms/tok, 47s total

Both models can load simultaneously in 12.5 GB (RSS ~16.2 GB — kernel compresses inactive pages). Sequential loading required on spinning SAS (parallel = disk thrash). Under memory pressure: kill 14B, optionally spin up 8B to free ~4 GB.

ModelLatency tierUse case
14BAsync/batch (~1 tok/s)Non-blocking reasoning, summarization, overnight classification, plan validation
8BAsync/batch (~1.8 tok/s)Same workload, faster turnaround — still not interactive on this CPU

Neither hits interactive latency on Sandy Bridge AVX1-only. Haswell+ (AVX2+FMA) would roughly double throughput. Route blocking work to cloud; use valhala for zero-cost async agent processing.