2026-05-26 dual-spark bench: GSM8K 94-98%, AIME 4/5, concurrency wall 12→125 tok/s
#7
by pastapaul - opened
Comprehensive 2x DGX Spark TP=2 validation: 8 OOB findings (F001-F008), driver alignment to 580.142 closes 23% cross-pair variance, GSM8K 8-shot 94-98% holds published baseline, AIME-2024 think-max 4/5 correct with 0 truncations at 65K (longest reasoning 14,725 tokens), concurrency scaling 10x agg (seq=1 → seq=16) at 16-128K ctx, OOM ceiling at 1M x seq=1.