Files
Robot_Arm/docs/experiment_round34_precision_ab.json
T
cbb 296cd17f98 exp: 轮次35——FP16/BF16与流水线复用都被实测否掉,f32认证需先抬eps到2e-4
- 显卡精度档(B=1,K=8,steps=1):推理 FP32 7.68 / FP16 8.31 / BF16 8.58 ms,整周期不变而FP16 P99 46.6→55.6;
  动作偏差 FP16 0.0050°、BF16 0.0377° ⇒ 该批量下不是算力受限,保持 FP32
- float32 认证:48段×9点×15组对的间隙差 max 1.32e-4 mm(有1例高估)→ 现 eps=1e-4 已被越界,要用需 ≥2e-4;
  判据结果两边一致(各24接受、0 仅f32安全、0 符号翻转),暂不放宽 dtype 守卫
- 流水线复用特征间隙:第一版42.1→31.8ms 但等价性检查不过(我取缓存行索引取错,差3.19mm);
  修正后只剩 −3.1ms 且缓存行与新算差 6.3e-5mm(不同 batch 的结合序)⇒ 放弃
- 顺带修掉轮次34日志里重复的章节标题
2026-10-20 21:01:04 +08:00

58 lines
1.0 KiB
JSON

{
"device": "NVIDIA GeForce RTX 3090",
"torch": "2.10.0+cu130",
"K": 8,
"steps": 1,
"B": 1,
"act_scale": 1.0,
"inference": {
"fp32": {
"p50": 7.681,
"p95": 9.426,
"p99": 10.631,
"max": 15.837
},
"fp16": {
"p50": 8.309,
"p95": 10.791,
"p99": 12.031,
"max": 12.322
},
"bf16": {
"p50": 8.577,
"p95": 11.057,
"p99": 14.174,
"max": 17.1
}
},
"cycle": {
"fp32": {
"p50": 36.507,
"p95": 43.688,
"p99": 46.568,
"max": 50.598
},
"fp16": {
"p50": 36.162,
"p95": 50.43,
"p99": 55.598,
"max": 57.043
}
},
"output_deviation": {
"fp16": {
"act_max_abs": 0.002096,
"act_p99_abs": 0.001059,
"deg_max_abs": 0.005,
"deg_p99_abs": 0.00219,
"would_change_execution_order": false
},
"bf16": {
"act_max_abs": 0.017171,
"act_p99_abs": 0.012661,
"deg_max_abs": 0.03766,
"deg_p99_abs": 0.03008,
"would_change_execution_order": true
}
}
}