kern
/ live
run
trace
anatomy
bottlenecks
shapes
feed
UTC
--:--:--
Beijing
--:--:--
◐
API service hours:
daily 13:00–22:00 Beijing time (UTC+8), equivalent to 05:00–14:00 UTC.
notices
UTC
Loading…
—
—
serving
per minute
live GPU trace
CUPTI kernel timeline & 100 µs hardware performance counters
focus
prefill → round
1 round
1 layer
full slice
⟲ reset
color flame by
role
DRAM BW
tensor %
window —
drag brush or scroll below to zoom from seconds down to single kernels
execution chrono-flame
step → layer / stage → CUPTI kernel
hardware counters
device-wide CUPTI PM sampling @ 100 µs
DRAM read
DRAM write
tensor pipe %
SM active %
Hover either strip for synchronized software call stack + 100 µs GPU counter readout · click any step or layer block to zoom in
the model, as kern runs it
measured GPU time & manifest contract across finalized CUPTI traces
▶ sweep
every layer, every role
GPU time
SOL %
what limits GPU speed
every call
step time by shape
measured execution time, per minute (recent active window)
feed
UTC, one line a minute