Skip to main content

Cleanba impala threads ASAP max GPU

Created on April 10|Last edited on April 16


10M20M30M40M50Mglobal_step100002000030000
Select runs that logged charts/learning_rate
to visualize data in this line chart.
10M20M30M40M50Mglobal_step-400-2000200400
Select runs that logged charts/avg_episodic_return
to visualize data in this line chart.
510152025Time (minutes)0100200300400
510152025Time (minutes)10203040506070
actor threads (A100)
1
cleanba impala 8 A100 baseline (a0_l1_d4)
1
actor threads asap (A100) prime candidate
1
a0_l1+2+3_t4
1
a0_l1+2+3_t8 w/ no .item() logging
1
a0_l1+2+3_t8_n60
1
a0+1_l2+3+4_t4
1
a0+1+2_l3+4+5_t4
1
a0_l1+2+3_t4 w/o actual env step
1
cleanba impala a0+1_l2+3+4_t4 w/ thread_affinity
1
cleanba PPO baseline
1
TPU
1
a0_l1+2+3_t4 (TPU equivalent)
1
a0_l0_t4 (TPU equivalent)
1
a0_l0_t4
1
a0_l0_t4 old actor threads
1
a0_l0_t4_n30
1
a0_l0_t4_n30_nmb1
1
a0_l0_t4_n120_nmb2
1
a0_l0_t4_n30_nmb2
1
a0_l0_t4_n16_nmb2
1