VKORE Runs Transformer Training on Inference-Only NPUs at 0.9999 Parity

VIDRAFT unveiled VKORE, a software layer that boosts NPU inference 3.8x for single users and 4.1x in server throughput, while enabling LoRA training of transformers on the same NPUs at 0.9999 numerical parity with PyTorch and about 170x the server's CPU speed. Tests ran on a Korean 8-card NPU server (32 chips) with Google's Gemma 4 26B.

• Gemma 4 26B inference rose from 14.1 to 53.9 tokens/s per user (3.8x)
• Server throughput climbed from 214 to 881 tokens/s (4.1x)
• LoRA training on NPU hit 0.9999 parity with PyTorch, ~170x CPU speed
• Training throughput reached about 54% of a single NVIDIA B200 at comparable cost