Batch throughput

	Recommended use
Start with	A model-and-card pair with a current batch receipt
Prioritize	Request mix, output length, admission limits, and aggregate completion rate
Validate	The real HTTP surface at the intended concurrency, not only a kernel benchmark
Read next	Serving, Performance, Testing

Do not assume the fastest single-stream path is the best batch path. Memra selects and qualifies decode modes per request shape; use the board and the server gate for the workload you will run.