Pearl reports low proof overhead on two production-scale model setups

Llama 70B · 4×H200

+5.08%

End-to-end overhead, Pearl-reported

DeepSeek V3.2 · 8×H200

+3.9%

End-to-end overhead, Pearl-reported

Pearl Research Labs says its proof-of-useful-work kernel adds 5.08% end-to-end overhead on a Llama 70B setup using four H200 GPUs, and 3.9% on DeepSeek V3.2 using eight. The current path runs on integer hardware through 7-bit weight-and-activation (W7A7) quantization; a floating-point version is in development.

If producing proofs ate a large slice of model-serving capacity, block rewards could fail to cover the cost. Single-digit overhead makes the dual-output thesis more plausible for compatible workloads.

• Reproducible benchmark code and workload configuration.
• Independent replication across more models and GPU generations.
• Latency, throughput, power and dollar cost reported together.
• Production usage and savings after rewards are sold at realistic prices.