AIMS AI Research Foundations
Make AI models fit, run fast and cost less
A hands-on guide to GPUs and Model Acceleration: how to measure compute and memory, find the bottleneck, and use the tricks real teams use to train and serve models.
Hi, I'm Ben. I've heard "you need GPUs for AI" a hundred times, but I couldn't tell you why. I couldn't say how much compute a task needs, how much memory a model needs, or why my training run crashed.
I didn't know how to find a bottleneck, make a GPU run faster, or serve a model to real users. Let's figure it all out together, one idea at a time.
Your journey
Why GPUs?
CPUs vs GPUs, and why AI is a parallel problem.
Module 2Measuring Compute and Memory
Estimate FLOPs, training time, cost and GPU memory.
Module 3GPU Bottlenecks
Find what limits a GPU, then see how FlashAttention moves less data.
Module 4Single-GPU Memory Optimization Tricks
Quantization, mixed precision and gradient accumulation.
Module 5Multi-GPU Training
Data, tensor and pipeline parallelism, plus ZeRO.
Module 6Inference Engineering
Prefill, decode, KV cache, batching and speculative decoding.
Practise and go further
Quiz
12 questions with instant feedback and explanations.
Choose Your GPU
The group worksheet, with answer checks and a plan builder.
Lecture slides
Browse all 78 slides or download the deck.
Open source
Built in the open
See how this course is made
Everything lives in a public GitHub repository. Reuse the materials and contribute for improvement.
github.com/rexsimiloluwah/aims-gpus-lecture-website View on GitHub