Compute & Systems for AI
The physics of intelligence
Why AI runs on GPUs, what FLOPs and memory bandwidth actually constrain, and how inference gets fast and cheap.
Advanced3 modules · 10 lessons~96 min930 XPAligned with Stanford systems primers · NVIDIA docsStart with: Deep Learning
Progress0/10
Module
Why GPUs Won
Silicon lotteries: the hardware that made deep learning inevitable
Parallelism vs Sequential9 min · 85 XPParallelism vs Sequential
FLOPs, Bandwidth & the Roofline11 min · 100 XPFLOPs, Bandwidth & the Roofline
TPUs & the Accelerator Landscape9 min · 90 XPTPUs & the Accelerator Landscape
Module
Making Models Fit
Precision, memory math, and splitting work across a fleet
Precision & Quantization10 min · 95 XPPrecision & Quantization
KV Cache Memory Math10 min · 100 XPKV Cache Memory Math
Distributed Training11 min · 105 XPDistributed Training
Module
Inference Engineering
Throughput, latency, and the economics of serving tokens
Batching: Throughput vs Latency9 min · 90 XPBatching: Throughput vs Latency
Speculative Decoding & Speedups10 min · 95 XPSpeculative Decoding & Speedups
Serving Economics9 min · 90 XPServing Economics
Edge Deployment8 min · 80 XPEdge Deployment
Capstone project
Find the bottleneck
Profile a training or inference workload and identify the real bottleneck — data loading, memory, compute? Fix one and measure the improvement.
Deliverable: A short performance report: profiling evidence, the fix, and before/after numbers.
Milestones
Finish every lesson to earn the AI Systems Certificate.