Câu đố về chip & phần cứng AI

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

1. What should a team define before comparing AI hardware?
   • The workload, such as training, interactive inference, or on-device processing
   • Only the model parameter count
   • A vendor’s peak operation count
   • The model name
2. Why does peak arithmetic performance alone not predict application speed?
   • Peak figures include monitor refresh rate
   • Memory movement, software, and other bottlenecks can dominate
   • All workloads are compute-bound
   • Arithmetic is unrelated to models
3. What should be checked alongside compute capacity?
   • Only the published weight-file size
   • Only the arithmetic throughput at maximum utilization
   • Memory capacity and bandwidth, including buffers and cached state
   • Only the sum of memory across all devices
4. What does the guide caution about memory spread across multiple devices?
   • It always becomes one seamless pool
   • It eliminates communication costs
   • It guarantees model quality
   • Aggregate memory is not automatically one simple pool
5. What should a team evaluate when using lower-precision numerical formats?
   • Accuracy changes on the intended model and task
   • Only whether model files became smaller
   • Only whether the numerical format is advertised
   • The peak operation rate without running the model
6. When is an advertised hardware capability actually used?
   • Whenever a chip has a high peak score
   • When hardware support, kernels, and the framework use it
   • Only when a benchmark says so
   • When listed in a brochure
7. Which measurements does the guide recommend for reproducible hardware comparisons?
   • Only peak arithmetic
   • Only a vendor demo
   • Latency, throughput, power, and cost per useful task with stated settings
   • Processor count alone
8. One billion parameters stored at 16 bits each takes about how much weight storage in the worked example?
   • 200 MB for the whole system
   • 20 GB including all runtime buffers
   • 2 GB including activations and cache
   • 2 GB for weights alone; other memory needs are excluded