// RUN โ€” FITS

Can the NVIDIA B200 run ERNIE-4.5-21B-A3B?

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

ERNIE-4.5-21B-A3B on the NVIDIA B200

QuantFits?Max contextSpeedBeyond context
FP16 yes 508k 139-165.7 t/s ~0.7-0.8 (slow)
Q8 yes 588k 278.1-331.4 t/s ~1.4-1.7 (slow)
Q4 yes 623k 496.6-591.8 t/s ~2.5-3 (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 B200 run ERNIE-4.5-21B-A3B?โ–ถ

Yes. At Q4 it generates ~496.6-591.8 tok/s and fits up to 623k context; beyond that it spills to system RAM offload (depends on your system RAM) and slows to ~2.5-3 tok/s.

How much context fits?โ–ถ

At Q4, up to about 623k tokens stay in fast memory; longer context spills to system RAM offload (depends on your system RAM) and throughput collapses.