Disaggregated Compute: Faster Inference via Split Prefill and Decode

toptickcrypto · x · 2026-08-20

An experiment demonstrates performance gains by disaggregating compute between a DGX Spark and a Mac M5 Max. Since prefill is compute-bound, the Spark (350 TFLOPS) runs it 5x faster than the Mac. As decode is memory-bound, the Mac (614GB/s bandwidth) runs it 2x faster than the Spark. By running prefill on the Spark and decode on the Mac, overlapping KV cache transfer over 10GbE, the setup outperforms running on a single device.

Original post →

More from coding & agent

coding & agent channel →