BREAKING
HeyGen Agent Triples VAE Decoder Speed
0
x
Inference speedup
0
s
fp32 baseline
0
s
Optimized time
Fewer Trials, Better Result
Manual trials
105
Auto-Optim trials
18
How the Speedup Was Built
fp32 eager
14.47
bf16
8.42
+CUDA graph
7.42
break-free
4.82
final
4.79
Quality Gate Held the Line
Quality gate
61.3 dB
●
Parity vs fp32 reference
●
Cleaner PSNR output
Stop condition
99.8% GPU
●
8 rounds no gain
●
Convolution-bound limit
Toolkit Released on GitHub
AI NEWS BLITZ
HeyGen says an autonomous AI agent tripled its video decoder speed on its own.