SM Clock (Streaming Multiprocessor Clock)
GPU core compute clock frequency in MHz, scaling between base and boost.
What it is
The SM clock is the operating frequency of the GPU's streaming multiprocessors in MHz, reported via DCGM_FI_DEV_SM_CLOCK, and is the primary determinant of raw compute throughput for CUDA kernels. SM clocks scale dynamically between idle (~210 MHz), base (1095 MHz on A100, 1620 MHz on H100 SXM), and boost frequencies (1410 MHz on A100, 1980 MHz on H100 SXM) based on workload demand, temperature, and available power headroom.
Why it matters
When thermal throttling or power capping engages, SM clock is the first frequency reduced -- each 100 MHz drop translates to roughly a 5-7% reduction in compute-bound throughput. An H100 stuck at base 1620 MHz while its 7 siblings boost to 1920-1980 MHz delivers roughly 18% less compute per cycle, making it the straggler that gates every AllReduce in the job. DCGM_FI_DEV_GPU_UTIL may still read 95-100% on a throttled GPU, masking the throughput loss entirely.
How to monitor
Track DCGM_FI_DEV_SM_CLOCK per GPU and compare against expected boost clock and against peers within the same node and job. Alert on intra-node clock divergence of more than 100-200 MHz during active workloads. Correlate with DCGM_FI_DEV_CLOCK_THROTTLE_REASONS to identify the throttle cause. Factryze flags SM clock anomalies within seconds to prevent silent throughput degradation from accumulating into hours of wasted training time.
DCGM_FI_DEV_SM_CLOCKRelated terms
Automatic GPU clock reduction when die temperature exceeds 83-90C safe limits.
GPU HBM/GDDR memory frequency in MHz that determines memory bandwidth.
Percentage of time GPU streaming multiprocessors are actively executing kernels.
Monitor this automatically
Factryze correlates GPU signals in real time: errors, clocks, and fabric health.
Get Started Free