backend,weights,pp512_tok_s,tg128_tok_s,ppl512,argmax_agree,threads,note llama.cpp(plain AVX512),Q4_K_M,180.50,13.39,NA,NA,16,"llama-bench -r 5; GGML_NATIVE=OFF, AMX=OFF, CPU_REPACK=OFF" llama.cpp(default AMX+repack),Q4_K_M,112.77,12.19,NA,NA,16,"llama-bench -r 5; native build" oneDNN,bf16,75.52,4.79,NA,NA,1,"llama-bench -r 5; -t 1 + OMP_NUM_THREADS=16; oneDNN AMX bf16 w" oneDNN,int8,90.09,7.67,NA,NA,1,"llama-bench -r 5; -t 1 + OMP_NUM_THREADS=16; oneDNN AMX int8 (s8/col)" oneDNN,int4,NA,NA,NA,NA,1,"no AMX int4 kernel; oneDNN 3.7 uses ref_int8 (~961 ms per 4096x4096 matmul at M=1)" OpenVINO,f32,125.60,2.22,28.8218,1.0000,16,"IR from HF fp32; INFERENCE_PRECISION_HINT=f32; tg=median of 128 steps" OpenVINO,bf16-hint,209.43,3.93,28.8452,0.9746,16,"IR from HF fp32; hint=bf16 (weights stay f32)" OpenVINO,int8-asym,219.17,4.70,29.1091,0.9512,16,"NNCF weight-only INT8_ASYM on IR; hint=f32" OpenVINO,int8-sym,226.68,NA,29.3413,0.9414,16,"NNCF weight-only INT8_SYM on IR; hint=f32" OpenVINO,int4-asym,223.65,5.20,32.6830,0.7031,16,"NNCF weight-only INT4_ASYM group=128; hint=f32" OpenVINO,int4-sym,222.48,NA,42.9488,0.7012,16,"NNCF weight-only INT4_SYM group=128; hint=f32" OpenVINO-GenAI,Q4_K_M,NA,NA,NA,NA,16,"LLMPipeline 2026.4.1 load fails: gguf_tensor_to_f16 (54 Q5_0 tensors unsupported); arch openpangu-r unsupported" OpenVINO-GenAI,f32-prefill-IR,NA,NA,NA,NA,16,"manual prefill IR loads; generate fails: Port attention_mask not found (needs stateful KV model)"