I am working through AI Systems Performance Engineering by Chris Fregly. The first two chapters cover training compute estimates and the hardware that runs those workloads. I put together these visualizations to follow the calculations and see how the components connect.
Like my visualizations for Inference Engineering, these are companions to specific parts of the book. Each link below includes its chapter so you can open the relevant diagram as you read.
How to use them
Open an interactive page from the links below. Hover over the diagrams or use the Tab key to focus their elements for more detail. The charts also have a table view, and the pages follow your device’s light or dark setting.
The animations below explain the hardware with on-screen labels. They have no audio.
The visualizations
- Training compute: where C ≈ 6ND comes from (Chapter 1, “Toward 100-Trillion-Parameter Models”): breaks down the operations in the forward and backward passes, then compares training compute estimates across models.
- GB200 NVL72: where 1.44 exaFLOPS comes from (Chapter 1, “NVIDIA’s ‘AI Supercomputer in a Rack’”): shows how per-GPU throughput, numeric precision, structured sparsity, and the rack’s 72 GPUs contribute to the quoted peak.
- Grace Blackwell: how the CPU accesses GPU memory (Chapter 2, “AI System Hardware Overview”): traces the paths between CPU and GPU memory and compares the bandwidth available on each path.
The Grace Blackwell superchip, animated
This animation accompanies Chapter 2’s “The CPU and GPU Superchip” section. It walks through the Grace CPU, the two Blackwell GPUs, their memory, and the links between them.
The animation runs for 3 minutes and 33 seconds. You can also watch it on YouTube.
Inside the SM, animated
This animation accompanies Chapter 2’s “Streaming Multiprocessor, Threads, and Warps” section. It shows what a streaming multiprocessor (SM) contains and how the GPU memory hierarchy is laid out. It then shows how an SM runs threads in warps of 32 and hides memory latency by switching between warps.
The animation runs for 3 minutes and 8 seconds. You can also watch it on YouTube.