Bit-Level IEEE-754 Standard Inspector
Sign (1b)
Exponent
Mantissa
Decoded Value
1.0000
Exponent Bias
7
Max Representable
448
Min Pos Subnormal
0.015625
Dynamic Scale & Quantization Loss Simulator
Tensor Outlier Scale (Std Dev Multiplier)
2.5x
MSE Loss
0.00214
Clipping Ratio
0.42%
SQNR (Signal/Noise)
26.8 dB
Layer-by-Layer Outlier Saturation Heatmap (32 Layers)
Hardware VRAM Footprint & Bandwidth
Latency & Token Throughput Projections
Memory Bound Limit
Yes (Token Gen)
Gen Speedup (FP8/BF16)
1.88x
Quantization Analysis Proof Summary
Model Configuration
Ling-3.0-flash
8.0 Billion Parameters
Active GPU Hardware
NVIDIA H100 SXM
3.35 TB/s Bandwidth
BF16 Baseline Weight VRAM
16.00 GB
Baseline FP16/BF16 Footprint
FP8 Quantized Weight VRAM
8.00 GB
50.0% Weight VRAM Savings
Estimated Token Throughput
418 tok/s
1.88x Speedup over BF16