00Four-line snapshot
Calculations and hypothesis
Not measured. Every calculation, projection and hypothesis on this page lives here; everything below this bar is fact or measured result.
00Dataflow pipeline: one draft step, replayed from the silicon trace
Pipeline stages run left to right like a CPU pipeline diagram: host, PCIe, HBM, on-chip memory, GEMV, seams and attention, QSFP. The MLPV card sits in the lower band. Press Play, or drag the scrubber, to replay the median tree of the F3t census (tree issue→done 2,219.31 µs).
- A box lights up while the FSM state that uses it is running.
- The bar at the bottom of each box is its occupancy over the last 140 µs (about one ATT item, C14).
- Arrows carry tokens while they move data.
- Click any box, arrow, lane or bar to see its JSON entry and sources.
burst: DMA or link frames (packet train)
per-cycle stream (one word every clock)
feature / activation regionKV regionweight region
dotted arrow = boot-time or per-request path (not active inside a tree)
Inspector
Click a box, arrow, lane label or timeline bar.
Per-block timeline of the same tree measured
Rows 1-2: the ATT0 and MLPV0 FSM states read by the host every ~3.4 µs. The other rows map those states onto the blocks they use (the mapping is from the RTL and is listed per lane in the JSON). Hatched = parked (PIPE_WAIT). Click a bar to move the playhead there.