// RUN — FITS

Can the NVIDIA H200 run AWD-LSTM-DRILL + dynamic evaluation† (WT2)?

Yes — here's how fast and up to what context, with real measured numbers.

AWD-LSTM-DRILL + dynamic evaluation† (WT2) on the NVIDIA H200

QuantFits?Max contextSpeedBeyond context
FP16 yes 1955k 51529.4-61411.8 t/s ~429.4-511.8 (slow)
Q8 yes 1956k 103058.8-122823.5 t/s ~858.8-1023.5 (slow)
Q4 yes 1956k 184033.6-219327.7 t/s ~1533.6-1827.7 (slow)

Estimate (memory-bound). Beyond "max context" the KV cache spills to system RAM offload (depends on your system RAM) and speed drops to the "beyond" figure.

FAQ

Can the NVIDIA H200 run AWD-LSTM-DRILL + dynamic evaluation† (WT2)?

Yes. At Q4 it generates ~184033.6-219327.7 tok/s and fits up to 1956k context; beyond that it spills to system RAM offload (depends on your system RAM) and slows to ~1533.6-1827.7 tok/s.

How much context fits?

At Q4, up to about 1956k tokens stay in fast memory; longer context spills to system RAM offload (depends on your system RAM) and throughput collapses.