Tensor Cores Quiz

Answers and explanations appear as you take the quiz.

1. What core mathematical operation are Tensor Cores specifically designed to accelerate?
   • Matrix multiply-and-accumulate (D = A x B + C)
   • Sorting large lists of numbers
   • Random number generation
   • Encrypting and decrypting data
2. Which NVIDIA GPU architecture first introduced Tensor Cores?
   • Pascal
   • Volta
   • Turing
   • Maxwell
3. Why do Tensor Cores often use reduced precision like FP16 or FP8?
   • Lower precision is required by law for AI
   • Higher precision is impossible on a GPU
   • It lets more numbers be processed per clock, boosting throughput while keeping accuracy acceptable
   • It makes the chip physically smaller
4. To preserve accuracy, what precision do Tensor Cores typically use for the running sum (accumulation)?
   • The same low precision as the inputs
   • No accumulation is performed
   • Integer-only with no fractions
   • A higher precision such as FP32
5. How do everyday AI frameworks like PyTorch make use of Tensor Cores?
   • Libraries like cuBLAS and cuDNN automatically tile large matrices into small blocks the cores can process
   • Developers must hand-write assembly for every model
   • Tensor Cores only work with custom hardware no software can reach
   • They cannot be used by high-level frameworks at all
6. Which newer numeric format did the Blackwell architecture add to push inference throughput even higher?
   • FP64 (double precision)
   • FP4 (4-bit floating point)
   • ASCII
   • FP128
7. Why should a team validate model quality when changing Tensor Core input precision?
   • Reduced precision guarantees identical outputs
   • The model’s vocabulary changes automatically
   • Lower-precision arithmetic can change numerical results, so accuracy should be checked for the task
   • Tensor Cores remove all rounding effects
8. Which computer-vision feature is a practical example of Tensor Core acceleration?
   • A database index with no matrix operations
   • A file compressor that never runs a neural network
   • A CPU-only spreadsheet calculation
   • Neural upscaling such as NVIDIA DLSS in games