Question 1

What's the difference between Hopper and Grace Hopper?

Accepted Answer

Hopper is a GPU architecture powering datacenter GPUs like H100 and H200. Grace Hopper is a Superchip that combines a Grace ARM CPU with a Hopper GPU via NVLink-C2C interconnect. The key difference is that Grace Hopper includes both a CPU and GPU in one package with unified memory access, while Hopper alone refers to just the GPU architecture.

Question 2

Why did NVIDIA split consumer and datacenter architectures?

Accepted Answer

Starting in 2022, NVIDIA released separate architectures for consumer (Ada Lovelace) and datacenter (Hopper) GPUs. This allows each architecture to be optimized for its workload: Ada focuses on ray tracing and gaming, while Hopper focuses on Tensor Cores and AI training. Datacenter GPUs need features like enhanced FP64 precision, error correction, and NVLink that gamers don't need.

Question 3

What are Tensor Cores?

Accepted Answer

Tensor Cores are specialized hardware units designed to accelerate matrix multiplication operations used in deep learning. Each generation adds new capabilities: Gen 1 (Volta) introduced FP16, Gen 2 (Turing) added INT8, Gen 3 (Ampere) brought TF32 and structural sparsity, Gen 4 (Hopper/Ada) added FP8 and Transformer Engine, and Gen 5 (Blackwell) introduced FP4 precision. They provide 10-20x speedup over CUDA cores for AI workloads.

Question 4

What does "2-die design" mean in Blackwell?

Accepted Answer

Modern chip manufacturing has physical size limits (reticle limit). Blackwell overcomes this by using two GPU dies connected via an ultra-fast interconnect at 10TB/s, making them function as a single GPU with 208 billion transistors. This multi-die approach allows NVIDIA to create larger, more powerful GPUs than would be possible with a single die while maintaining high manufacturing yields.

Question 5

Can I mix different architectures in a cluster?

Accepted Answer

Technically yes, but it's not recommended for performance-critical workloads. Different architectures have different capabilities (Tensor Core generations, precision support, NVLink bandwidth) that can create bottlenecks. For distributed training, it's best to use identical GPUs. However, for inference serving or development/testing, mixing architectures is fine.

Frequently Asked Questions

What's the difference between Hopper and Grace Hopper?

Why did NVIDIA split consumer and datacenter architectures?

What are Tensor Cores?

What does "2-die design" mean in Blackwell?

Can I mix different architectures in a cluster?

Still Have Questions?

Stay Updated on GPU Specs