device: NVIDIA GeForce RTX 5060 Ti, SM 12.0
device: theoretical 448.0 GB/s DRAM, 23704 Gflop/s FP32
device: clock ramp for 30 s at 288 blocks x 256 threads
device: 1567 telemetry samples
compute: self-checks
compute: self-checks passed
compute: peak throughput, 288 blocks x 256 threads x 200000 iters
compute: dependent chain latency
compute: warp MMA
memory: bandwidth sweep, 288 blocks x 256 threads
memory: pointer chase
memory: chase ws=32KiB+ld_default
memory: chase ws=2MiB+ld_default
memory: chase ws=32MiB+ld_default
memory: chase ws=256MiB+ld_default
memory: chase ws=32MiB+ldg
memory: chase ws=32MiB+ldca
memory: chase ws=32MiB+ldcs
memory: chase ws=32MiB+ldlu
memory: shared memory bank conflicts
memory: tile copies into shared (ld/st, cp.async, TMA)
memory: atomics
exec: occupancy sweep over threads per block
exec: occupancy sweep over shared memory
exec: barriers
exec: cooperative grid barrier, 72 blocks
exec: warp divergence
exec: warp primitives
launch: CPU-side submit cost
launch: clock sync offset -1790076050813199360 ns, window 4864 ns
launch: 20000 round trips for the tail distribution
launch: cudaEvent costs
launch: allocation
launch: stream create/destroy
launch: CUDA graphs
pcie: bandwidth sweep
pcie: small synchronous transfers
pcie: concurrent streams
pcie: compute and transfer overlap
pcie: host registration
pcie: unified memory migration
pcie: zero-copy chase over host memory