int8, one scale: step 7.336e-04   predicted RMS error step/sqrt(12) 2.118e-04   measured 2.117e-04

weights: Gaussian std 0.02, 512x1024.  outliers: 524 entries (0.1%) multiplied by 30, largest |w| 0.093 -> 1.768
format                bits/weight   weight error: clean   with outliers   output error with outliers
int8, one scale       8.00          0.0106                 0.1476          0.1457
int8, scale per row   8.03          0.0079                 0.0451          0.0446
int8, scale per 32    8.50          0.0053                 0.0094          0.0094
GGUF Q8_0             8.50          0.0053                 0.0094          0.0094
int4, one scale       4.00          0.1920                 0.7393          0.7413
int4, scale per row   4.03          0.1427                 0.4703          0.4733
int4, scale per 32    4.50          0.0970                 0.1157          0.1135
GGUF Q4_0             4.50          0.0859                 0.1074          0.1054
FP8 E4M3, one scale   8.00          0.0265                 0.0265          0.0265
FP8 E5M2, one scale   8.00          0.0526                 0.0520          0.0519

clean data, scale per 32: int4 error / int8 error = 18.2   the step ratio 127/7 = 18.1
Q4_0: largest-magnitude weight of each block, worst relative error after the round trip 4.79e-04

FP8 E4M3 error by weight size    std:   0.1000   0.0200   0.0040   0.0008
  scale 1 (plain cast)                0.0265   0.0330   0.1410   0.6905
  scale = amax/448                    0.0265   0.0265   0.0265   0.0265

int8 matmul with int64 accumulation, then rescaled, against dequantize-then-multiply: largest difference 8.9e-07 on outputs of size 0.516

layout                                       bits/weight   GiB per billion weights
fp16 or bf16                                 16.00         1.86
ComfyUI float8_e4m3fn, scalar scale          8.00          0.93
ComfyUI mxfp8, 8-bit scale per 32            8.25          0.96
ComfyUI nvfp4, 8-bit scale per 16            4.50          0.52
GGUF Q8_0, fp16 scale per 32                 8.50          0.99
GGUF Q4_0, fp16 scale per 32                 4.50          0.52
GGUF Q4_K, 144 bytes per 256                 4.50          0.52
