FP32, 16 threads, batch 1
Tokens / sec
BF16, batch 1
Local batch-one measurements excluding tokenization. CPU uses FP32 with 16 threads; RTX 3090 uses BF16.