引导式测验 · 简单 等级

NVIDIA 人工智能测验

涵盖 NVIDIA 生态系统角色、推理优化、兼容性、输出验证和端到端性能测量。

相关引导路径英伟达人工智能
问题 1 的 8

A request spends 100 ms on inference and 900 ms preparing data. If inference becomes twice as fast, about how long does the request take?

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. A request spends 100 ms on inference and 900 ms preparing data. If inference becomes twice as fast, about how long does the request take?
    • 950 ms
    • 500 ms
    • 900 ms
    • 450 ms
  2. In the NVIDIA deployment workflow described by the guide, what role can an inference optimizer such as TensorRT play?
    • Replace the need to evaluate output quality
    • Optimize a trained model for execution before it is exposed through a serving system
    • Train every model from raw data automatically
    • Store all application data without a runtime
  3. A model is fast on GPU but the application remains slow because input preparation dominates. What should the team investigate?
    • Increasing model parameter count
    • Changing the model vendor without profiling
    • The full data and preprocessing path, not only accelerator throughput
    • Removing output validation
  4. After changing precision or the inference execution path, what should be compared with the original model?
    • 仅其文件扩展名
    • Only the GPU temperature
    • Whether the model has the same marketing name
    • Task outputs and quality on representative evaluation inputs
  5. Which details help make an inference benchmark reproducible?
    • Hardware, software versions, precision, supported operations, workload and input conditions
    • Only the vendor name
    • 仅计算模型的参数
    • The color of the server chassis
  6. Which statement about a larger accelerator is supported by the guide?
    • It guarantees every application runs faster
    • It may not improve end-to-end speed when another stage is the bottleneck
    • It removes all data transfer costs
    • It ensures output quality is unchanged
  7. Which cost dimensions does the guide recommend measuring for a deployed AI workload?
    • Only peak theoretical arithmetic throughput
    • Only model accuracy on training data
    • Latency, throughput, power and cost under the intended application conditions
    • Only the accelerator’s name
  8. Why does the guide separate hardware capability from software-stack performance?
    • Software never affects accelerator performance
    • Hardware specifications directly guarantee service speed
    • All NVIDIA devices execute every model identically
    • Supported operations, kernels, and compatibility determine how hardware is actually used