7. Key Quality Metrics

Performance-related metrics must remain key acceptance criteria throughout Stages 3-6. Stage 3 focuses on operator performance, Stage 4a on communication throughput and linearity, and Stage 4b on training/inference accuracy and end-to-end performance. Before advanced cooperation, Day 0, distribution, or showroom plans, the efficiency target relative to the NVIDIA baseline under equivalent compute must be confirmed. The exact percentage will be defined later according to the KT2 requirements.

Stage	Acceptance criteria	Pass standard	Required hardware
Stage 0	MOU signing	Effective after both parties sign or seal	N/A
Stage 1	Environment readiness	Chip is recognized by the OS and PyTorch can recognize the backend	1 development board
Stage 2	Triton kernel compilation pass rate	>=95%, CI machine availability >=99%	1 CI machine
Stage 3	FlagGems operator test pass rate	>=80%, core operators 100%	1 CI machine and 1 debug machine
Stage 4a	Homogeneous collective communication throughput ratio	>=99% of the native library	2-machine multi-card cluster
Stage 4b	Model accuracy deviation	<2% relative to the NVIDIA baseline	Reuse Stage 4a devices
Stage 5	CTS core tests passed	100% passed or waiver approved	1 CI machine and 2 SVT machines
Stage 6	Number of models released through FlagRelease	>=10 models	1 dedicated FR machine
Advanced cooperation	Equivalent-compute performance comparison	Efficiency reaches X% of the NVIDIA equivalent-compute baseline	Multi-card cluster, FR machine, or CI