bench de CPU: 512 e 1024, EPYC 16 vCPU, RAM folgada

bench/cpu_speed.json

@@ -0,0 +1,50 @@
+{
+ "512": {
+  "tq3p": {
+   "encode_prompt_s": 17.51,
+   "sampling_s": 259.43,
+   "s_per_it": 64.858,
+   "vae_decode_s": 9.43,
+   "total_s": 286.4
+  },
+  "q4km": {
+   "encode_prompt_s": 12.81,
+   "sampling_s": 171.32,
+   "s_per_it": 42.83,
+   "vae_decode_s": 9.46,
+   "total_s": 193.6
+  }
+ },
+ "1024": {
+  "tq3p": {
+   "encode_prompt_s": 17.19,
+   "sampling_s": 833.03,
+   "s_per_it": 208.257,
+   "vae_decode_s": 35.89,
+   "total_s": 886.1
+  },
+  "q4km": {
+   "encode_prompt_s": 12.17,
+   "sampling_s": 570.16,
+   "s_per_it": 142.54,
+   "vae_decode_s": 36.9,
+   "total_s": 619.2
+  }
+ },
+ "_meta": {
+  "host": "RunPod CPU pod cpu5g, AMD EPYC 4564P 16 vCPU, 124 GB RAM (96 GB disponivel)",
+  "runtime": "stable-diffusion.cpp master-802-e92e86f + sdcpp-cpu.patch, SD_CUDA=OFF, -t 16",
+  "protocolo": "4 steps, euler, cfg 1.0, seed 42, VAE small-decoder; RAM FOLGADA (sem --mmap) => mede o TETO de CPU, sem releitura de disco",
+  "n": "1 execucao por (braco, resolucao) -- s/it e derivado do sampling isolado do log, nao do wall total",
+  "bracos": {
+   "tq3p": "DiT TQ3P (6,1875 bpw) + TE Qwen3-8B TQ3P",
+   "q4km": "DiT Q4_K_M (~4,5 bpw) + TE Qwen3-8B Q8_0"
+  },
+  "veredito": "em CPU o Q4_K_M GANHA: 1,51x a 512 e 1,46x a 1024 -- o inverso da GPU (onde o TQ3P ganha 1,52x). Em CPU o custo e bytes lidos por peso, e o K3 le ~37% mais que o Q4_K_M.",
+  "executado_via": "python3 -m sasori draw (o subcomando novo, exercitado de verdade)",
+  "razao_tq3p_sobre_q4km": {
+   "512": 1.514,
+   "1024": 1.461
+  }
+ }
+}