Decode · tok/s
93.6
Prefill
—
tok/100W
39
Turned on the 2nd network cable between the Sparks and TensorFold got even faster: 93.6 tok/s for 1 user (vLLM: 48.5), 1.93x 213.8 tok/s with 6 users (+29%)
First-person 2× DGX Spark. Rank is the quoted 93.6 one-stream after enabling the second network cable. Quant not named in the post. Prior 4-Spark comparison was on the same model.
