Stage contribution to the latency lower bound
These mutually exclusive stage contributions add to the latency lower bound.
Updating calculation.
Optimistic analytical lower bound; not measured.
01
End-to-end latency lower bound across the selected GPU families at one sharding scenario.
| System | Sharding | Latency lower bound | Ideal throughput | Weights / GPU | Peak / GPU | Capacity | Execution path |
|---|
02
Exact stage contributions, plus overlapping whole-model certificates.
These mutually exclusive stage contributions add to the latency lower bound.
Each bar sums one certificate across all stages. The certificates overlap and must not be added together.
03
Operational intensity versus effective per-rank throughput.
Drag to pan; use the wheel or trackpad to zoom. Each roofline is the piecewise minimum of its memory-bandwidth slope and compute plateau. Communication-bound points fall below it because collective topology adds a third floor.
04
Embedding + 93 decoder layers + final norm + LM head.
| Stage | Path | Latency lower bound | Share | Dominant operation | Layer lower-bound certificate(s) |
|---|
Select any operator to reveal the formula, substituted values, units, and result behind every numeric field. Operator rooflines inside a stage can overlap; the stage latency lower bound is the maximum of its critical path and aggregate compute-, HBM-, and communication-demand certificates. Exact ties display every co-limiting certificate.
05
Per-rank accounted model and state against vendor-nameplate HBM.
Physical bytes resident on each rank.
This is not an OOM guarantee. Runtime reserve, allocator fragmentation, workspaces, and unaccounted live activations remain outside the bound.
06
Checkpoint facts used by every operation count in this report.