Groq MLE Interview Questions (2026)

Free questions with model answers (4)

• #176 Explain disaggregated prefill/decode serving and when it pays for itself.

  Prefill processes all prompt tokens in parallel and is usually matrix-multiply compute bound.

• #203 Redo the decode roofline argument for an LPU with no HBM — only on-die SRAM — and explain what changes.

  Prefill processes all prompt tokens in parallel and is usually matrix-multiply compute bound.

• #204 Walk through deploying a 70B dense model at 8-bit weights on chips with ~230MB of SRAM each, including unit economics.

  Start with workload, users, data, quality metrics, latency, throughput, freshness, reliability, and cost targets.

• #716 Per token decode latency from memory bandwidth

  Key insight: decode latency is a roofline estimate: seconds per token is approximately bytes moved divided by memory bandwidth.