引導式測驗 · 中等 等級
GPU 與 TPU 的 AI 測驗
在選擇硬體之前,先比較 GPU 和 TPU 架構、軟體支援、系統限制以及團隊所需的特定工作負載證據。
問題 1 的 8
指南列出的支援雲 TPU 工作負載的框架有哪些?
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。
- 指南列出的支援雲 TPU 工作負載的框架有哪些?
- JAX, TensorFlow, and PyTorch/XLA
- CUDA C++ only
- PyTorch CPU and ONNX Runtime only
- TensorFlow Lite only
- TPU 是誰設計的?
- NVIDIA
- 英特爾
- Apple
- 哪一種核心硬體結構使 TPU 在矩陣乘法方面有效率?
- 旋轉的磁碟
- 大型機械繼電器組
- 乘法累加單元的脈動數組
- 單一超快核心
- 哪些系統因素會影響 GPU 與 TPU 的效能比較?
- Only the accelerator’s peak operation count
- Only the number of parameters in the model
- Only the cloud region name
- Model support, software, interconnects, memory, utilization, and deployment cost
- GPU 相對於 TPU,對許多團隊而言實際的優勢是什麼?
- A broad ecosystem of CUDA libraries, tools, and existing GPU-supported workloads
- A guarantee that every model is faster on a GPU
- A requirement that all training use TensorFlow
- A built-in advantage in every cost comparison
- Why does the guide recommend benchmarking the intended workload on the specific platform?
- Peak operation counts always predict end-to-end speed exactly
- Measured results depend on workload shape, kernels, input pipeline, memory, and system scaling
- All workloads use the same hardware path and precision
- A benchmark removes the need to check software support
- What should a team compare before choosing between a GPU and a TPU for training?
- Peak FLOPS alone for an unrelated precision
- The chip’s name and product launch date
- End-to-end time and cost for its model, framework, input pipeline, and target scale
- A benchmark from a different model and software stack
- How does Google describe a TPU pod?
- A single GPU with no network
- A CPU-only cluster
- A model format
- A group of TPU chips connected through a high-speed interconnect