The failures are part of the result.

Can batching reuse one parameter stream?

batch        1     8     32    128
bytes read   ─────────────────────
useful work  ████████████████████▶

Measured: Ps equals batch size; streamed and resident outputs agree bit-for-bit.

Can scalar CPU compute keep up with cached storage?

The tested engine stayed compute-bound. This gives purpose-built arithmetic a concrete job; it does not prove FPGA speed.

Does serial MoE preserve a real model result?

All 24 streamed MoE layers preserved top-1, retained 9/10 top logits, and stayed below the 0.002 absolute-error gate.

Does asynchronous buffering automatically help?

Seven-run cases ranged from −5.5% to +11.7%; cache-bypass proxies gained 2–5%. Correct plumbing, no universal speed claim.