FP8 MoE router projection
FP8-quantized MoE router gate projection that computes expert routing logits. Projects hidden states to expert scores using blockwise FP8 quantization with BlockWise1x128 scaling for activations and BlockWise128x128 scaling for weights. The weight matrix is padded to 128 experts for BlockWise128x128 compatibility.
Reported evidence · last observed 2026-08-22. Reported by source; not independently reproduced. No public source. License unknown.
Current records
Workloadsuite of 16 cases · mean latencyhidden_size = 3584 · num_experts = 64 · hidden_blocks = 28 · padded_experts = 12816 cases
Not measured on H100 for this workload. Challenges →
18.69µs81% SOL1.00×Reported · license unknown · no source2026-08-22
Measured exactly what you asked. Reported by source; not independently reproduced.
29.16µs80% SOL1.05×Reported · license unknown · no source2026-08-24
Measured exactly what you asked. Reported by source; not independently reproduced.
310.1µs78% SOL1.16×Reported · license unknown · no source2026-08-24
Measured exactly what you asked. Reported by source; not independently reproduced.
410.2µs78% SOL1.17×Reported · license unknown · no source2026-08-23
Measured exactly what you asked. Reported by source; not independently reproduced.
510.5µs77% SOL1.21×Reported · license unknown · no source2026-08-15
Measured exactly what you asked. Reported by source; not independently reproduced.
610.7µs77% SOL1.23×Reported · license unknown · no source2026-08-23
Measured exactly what you asked. Reported by source; not independently reproduced.
710.8µs77% SOL1.24×Reported · license unknown · no source2026-07-31
Measured exactly what you asked. Reported by source; not independently reproduced.
811.3µs76% SOL1.30×Reported · license unknown · no source2026-07-31
Measured exactly what you asked. Reported by source; not independently reproduced.
911.4µs75% SOL1.31×Reported · license unknown · no source2026-08-14
Measured exactly what you asked. Reported by source; not independently reproduced.
1011.4µs75% SOL1.31×Reported · license unknown · no source2026-08-14
Measured exactly what you asked. Reported by source; not independently reproduced.
1111.4µs75% SOL1.31×Reported · license unknown · no source2026-08-14
Measured exactly what you asked. Reported by source; not independently reproduced.
1212.0µs74% SOL1.38×Reported · license unknown · no source2026-08-06
Measured exactly what you asked. Reported by source; not independently reproduced.
1312.0µs74% SOL1.38×Reported · license unknown · no source2026-08-14
Measured exactly what you asked. Reported by source; not independently reproduced.
1412.9µs73% SOL1.49×Reported · license unknown · no source2026-08-03
Measured exactly what you asked. Reported by source; not independently reproduced.
1513.3µs72% SOL1.53×Reported · license unknown · no source2026-07-18
Measured exactly what you asked. Reported by source; not independently reproduced.
1613.4µs72% SOL1.55×Reported · license unknown · no source2026-07-14
Measured exactly what you asked. Reported by source; not independently reproduced.
1715.8µs68% SOL1.82×Reported · license unknown · no source2026-07-27
Measured exactly what you asked. Reported by source; not independently reproduced.
1822.1µs60% SOL2.54×Reported · license unknown · no source2026-07-19
Measured exactly what you asked. Reported by source; not independently reproduced.
1924.3µs57% SOL2.79×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
2026.9µs55% SOL3.10×Reported · license unknown · no source2026-07-13
Measured exactly what you asked. Reported by source; not independently reproduced.
2145.3µs41% SOL5.21×Reported · license unknown · no source2026-07-14
Measured exactly what you asked. Reported by source; not independently reproduced.