NCA GPU Architecture & Fundamentals 5

Question 1

What is the function of NVIDIA's RT Cores in Turing and later architectures?

Accepted Answer

Accelerate ray-box and ray-triangle intersection tests for real-time ray tracing

Answer

RT Cores are dedicated fixed-function hardware units that accelerate bounding volume hierarchy (BVH) traversal and ray-triangle intersection, the core bottleneck in real-time ray tracing.

Question 2

What is the difference between NVIDIA's 'compute capability' version and a GPU's CUDA toolkit version?

Accepted Answer

Compute capability describes hardware feature support (SM architecture); CUDA toolkit version describes the software SDK release

Answer

Compute capability (e.g., 8.0 for A100, 9.0 for H100) encodes the hardware SM feature set, while the CUDA toolkit is the software release that may support one or more compute capability levels.

Question 3

How does NVIDIA's Ampere architecture's sparsity feature improve Tensor Core throughput?

Accepted Answer

It doubles effective Tensor Core throughput by skipping zero-valued weights in 2:4 structured sparse matrices

Answer

Ampere's Tensor Cores support 2:4 structured sparsity, where exactly 2 of every 4 weights are zero, allowing the hardware to skip zero multiplications and double MMA throughput for sparse models.

Question 4

What is the purpose of NVIDIA's GPUDirect RDMA technology?

Accepted Answer

Allow GPU memory to be accessed directly by network adapters without CPU involvement, reducing latency

Answer

GPUDirect RDMA allows network adapters (InfiniBand or RoCE) to read and write GPU memory directly via PCIe, bypassing CPU and system memory and reducing communication latency for distributed training.

Question 5

In the context of NVIDIA GPU execution, what is a 'CUDA stream' and how does it enable performance optimization?

Accepted Answer

A sequence of CUDA operations that execute in order, allowing overlap of independent streams for concurrency

Answer

A CUDA stream is an ordered queue of GPU operations; operations in different streams can overlap (kernel execution, memory copies), enabling better GPU utilization through concurrency.

(NCA) NVIDIA-Certified Associate Practice Test

(NCA) NVIDIA-Certified Associate Practice Test

NCA GPU Architecture & Fundamentals 5