ywkang
fca24feac5
- intercube_allreduce: add single-cube fast path that skips intra-SIP mesh reduce and goes directly to inter-SIP exchange. Fixes IPCQ deadlock when TP launches kernel on one cube per SIP. - distributed.py: derive effective cube dims from tensor shard placement instead of hardcoding topology mesh size. - pyproject.toml: add matplotlib>=3.7 to dependencies. - pe_dma.py (prior commit): add MMU translation in pipeline DMA path. 577 passed, 0 failed (was 529 passed, 10 failed). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
38 lines
2.5 KiB
CSV
38 lines
2.5 KiB
CSV
algorithm,sip_topology,n_sips,n_elem,bytes_per_pe,bytes_per_sip,latency_ns
|
|
intercube_allreduce,mesh_2d_no_wrap,6,8,16,256,3508.4249999999993
|
|
intercube_allreduce,mesh_2d_no_wrap,6,32,64,1024,3515.55
|
|
intercube_allreduce,mesh_2d_no_wrap,6,64,128,2048,3525.0499999999975
|
|
intercube_allreduce,mesh_2d_no_wrap,6,128,256,4096,3544.049999999992
|
|
intercube_allreduce,mesh_2d_no_wrap,6,512,1024,16384,3667.049999999992
|
|
intercube_allreduce,mesh_2d_no_wrap,6,1024,2048,32768,3837.049999999992
|
|
intercube_allreduce,mesh_2d_no_wrap,6,2048,4096,65536,4177.049999999992
|
|
intercube_allreduce,mesh_2d_no_wrap,6,4096,8192,131072,4857.049999999959
|
|
intercube_allreduce,mesh_2d_no_wrap,6,8192,16384,262144,6217.049999999945
|
|
intercube_allreduce,mesh_2d_no_wrap,6,16384,32768,524288,8937.049999999937
|
|
intercube_allreduce,mesh_2d_no_wrap,6,32768,65536,1048576,14377.049999999872
|
|
intercube_allreduce,mesh_2d_no_wrap,6,49152,98304,1572864,19817.049999999872
|
|
intercube_allreduce,ring_1d,6,8,16,256,3073.1299999999937
|
|
intercube_allreduce,ring_1d,6,32,64,1024,3079.8799999999947
|
|
intercube_allreduce,ring_1d,6,64,128,2048,3088.879999999992
|
|
intercube_allreduce,ring_1d,6,128,256,4096,3106.8799999999865
|
|
intercube_allreduce,ring_1d,6,512,1024,16384,3225.8799999999865
|
|
intercube_allreduce,ring_1d,6,1024,2048,32768,3391.8799999999865
|
|
intercube_allreduce,ring_1d,6,2048,4096,65536,3723.8799999999865
|
|
intercube_allreduce,ring_1d,6,4096,8192,131072,4387.879999999965
|
|
intercube_allreduce,ring_1d,6,8192,16384,262144,5715.879999999957
|
|
intercube_allreduce,ring_1d,6,16384,32768,524288,8371.879999999932
|
|
intercube_allreduce,ring_1d,6,32768,65536,1048576,13683.879999999903
|
|
intercube_allreduce,ring_1d,6,49152,98304,1572864,18995.879999999917
|
|
intercube_allreduce,torus_2d,6,8,16,256,2190.4799999999923
|
|
intercube_allreduce,torus_2d,6,32,64,1024,2196.479999999993
|
|
intercube_allreduce,torus_2d,6,64,128,2048,2204.4799999999905
|
|
intercube_allreduce,torus_2d,6,128,256,4096,2220.479999999985
|
|
intercube_allreduce,torus_2d,6,512,1024,16384,2325.479999999985
|
|
intercube_allreduce,torus_2d,6,1024,2048,32768,2471.479999999985
|
|
intercube_allreduce,torus_2d,6,2048,4096,65536,2763.479999999985
|
|
intercube_allreduce,torus_2d,6,4096,8192,131072,3347.4799999999777
|
|
intercube_allreduce,torus_2d,6,8192,16384,262144,4515.4799999999705
|
|
intercube_allreduce,torus_2d,6,16384,32768,524288,6851.479999999952
|
|
intercube_allreduce,torus_2d,6,32768,65536,1048576,11523.479999999923
|
|
intercube_allreduce,torus_2d,6,49152,98304,1572864,16195.479999999952
|