গাইড লিঙ্ক ক্যুইজ · মাঝারি স্তর
nvidia-smi এবং DCGM কুইজের সাথে GPU ইউটিলাইজেশন মনিটরিং
GPU ব্যস্ত সময়, মেমরি কার্যকলাপ বনাম ক্ষমতা, শক্তি এবং স্বাস্থ্য মেট্রিক্স ব্যাখ্যা করুন এবং প্রকৃত বাধাগুলি খুঁজে পেতে প্রোফাইলিং ব্যবহার করুন।
প্রশ্ন 1 এর 8
What does nvidia-smi GPU utilization approximately report?
প্রতিটি প্রশ্নের পূর্বরূপ দেখুন8 প্রশ্ন
আপনি কুইজ নেওয়ার সাথে সাথে উত্তর এবং ব্যাখ্যাগুলি উপস্থিত হয়।
- What does nvidia-smi GPU utilization approximately report?
- The fraction of peak FLOPs achieved by the GPU
- The percentage of total VRAM allocated to a process
- The share of a sample period with one or more kernels executing
- The amount of device memory currently allocated
- How does memory utilization differ from memory usage?
- GPU busy time versus VRAM capacity in gigabytes
- Memory activity over time versus allocated memory capacity
- Two names for tensor-core activity
- The same counter reported in different units
- A GPU reports high utilization but low throughput. What is a plausible explanation?
- The utilization field directly measures peak FLOPs
- A busy GPU must be compute-bound
- High allocated memory proves device-memory bandwidth is saturated
- The workload may be memory-bound or use inefficient kernels
- What might low GPU activity between training steps indicate?
- The input pipeline or synchronization may be stalling the device
- Low utilization proves all model parameters fit in VRAM
- A longer step means the training run is overfit
- The model must be using tensor cores at peak throughput
- Which DCGM metric can help assess memory-traffic activity?
- SM occupancy ratio
- GPU memory allocated in MiB
- DRAM activity or utilization ratio
- Tensor-pipe activity ratio
- What can temperature, clocks, and power help diagnose?
- Whether the model has class imbalance
- Thermal or power constraints affecting performance
- Whether data-loader workers have duplicate samples
- Whether a model uses grouped-query attention
- When should an engineer use a framework profiler after reading DCGM metrics?
- When a DCGM field shows the exact source line of a slow kernel
- When a kernel fails to compile before any execution telemetry exists
- Before collecting aggregate metrics so the profiler has a baseline
- When aggregate telemetry cannot identify the slow kernel or operator
- Why record sampling period and hardware context with utilization data?
- Sampling interval and device details affect metric interpretation
- The sampling period has no effect because the counter is instantaneous
- Device details determine model output labels
- Recording context guarantees equal throughput across runs