Hedra speeds up Krea 3 diffusion transformer 6.9x on AMD MI355X

Krea says Hedra Labs made the diffusion transformer behind Krea 3 6.9 times faster on AMD Instinct MI355X, cutting generation time from 22.9 seconds to 3.3 seconds.

Krea's post says the output keeps 98.3% of BF16 quality by SSIM, a similarity measure.

Krea also reports that MXFP6 GEMM runs 2.8 times faster than BF16 and that GPU transfer waiting fell from 18% to 2% of time. Krea attributes the work to Hedra's autonomous research system.