引導式測驗 · 中等 等級
為機器學習選擇 GPU 測驗
將 GPU 記憶體、頻寬、運算、互連和軟體支援與定義的訓練或推理工作負載相符。用於目標部署設定。
問題 1 的 8
What determines whether a model's working state fits on a GPU?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What determines whether a model's working state fits on a GPU?
- Weights, gradients, and activations alone regardless of inference cache
- Peak FLOPs relative to the model parameter count
- Available memory relative to weights, activations, optimizer state, and cache
- Memory bandwidth relative to the model input rate
- How does memory bandwidth differ from memory capacity?
- Both terms describe the number of bytes installed
- Bandwidth measures data transfer rate; capacity measures how much can fit
- Capacity measures transfer rate; bandwidth measures storage size
- Bandwidth and capacity are both theoretical FLOPs
- When can interconnect become important in multi-GPU training?
- When one GPU processes a workload without exchanging data with peer GPUs
- Only when data is first read from host storage before training
- When each model layer fits on one device and no peer transfer occurs
- When workers exchange gradients, activations, or model partitions
- Why verify framework and driver support before selecting hardware?
- A device may lack compatible or optimized kernels in the current software stack
- Peak FLOPs guarantee every model operator will use the fastest path
- Framework support is independent of driver and runtime versions
- A GPU makes every precision format equally fast
- Which benchmark comparison is most informative?
- Use the same model but different precision and batch sizes on each GPU
- Compare product-sheet peak values for different arithmetic formats
- Benchmark the same model, precision, workload, and latency target on each candidate
- Compare the same model at different sequence lengths and request rates
- What can constrain batch-one autoregressive LLM decoding after weights are loaded?
- Training optimizer state and full-dataset size after weights are loaded
- Memory bandwidth and KV-cache capacity can constrain batch-one decoding
- The number of concurrent requests has no impact once model weights fit
- The inference engine removes all memory traffic after model loading
- Why can a high-peak GPU scale poorly across several devices?
- Adding GPUs automatically raises linear scaling efficiency
- Peak FLOPs specifications include cross-GPU communication costs
- Training never synchronizes model or gradient state
- Communication or synchronization can dominate useful computation
- What should guide final hardware selection?
- Use measured end-to-end workload performance and total cost
- Choose by peak throughput for a precision the workload cannot use
- Choose by nominal capacity without runtime memory overhead
- Choose by throughput measured at a different latency target