Nvidia reports 42.5% inference throughput gain on DGX Spark

Nvidia reports that three software optimizations increased Ling-3.0-flash decode throughput by 42.5% on DGX Spark across all 1,319 GSM8K questions. Mean end-to-end latency fell by 24.5%, with no statistically detectable accuracy change. The configuration combines specialized Humming kernels, FP8 compression of the output layer's weights and speculative decoding that drafts and verifies multiple tokens together. The first two reduce work per decoding step; the third reduces the number of steps. Combining them requires the draft model to share the target's entire output-layer module. Without that integration, throughput falls below the unoptimized baseline. Nvidia cautions that a separate synthetic test's 185.1% throughput gain does not establish performance on other workloads.