NVIDIA is the default, not the only choice. This module draws the other machines the way the previous topic drew NVIDIA’s: AMD’s chiplet GPUs and its first rack, Google’s TPUs and their optical pods, AWS Trainium and the other hyperscaler chips, and the SRAM-first designs from Cerebras and Groq. The first lesson maps the whole field on two axes and compares the software stacks, which decide more purchases than the silicon does.
| # | Lesson | The question it answers |
|---|---|---|
| 9.a | The Accelerator Landscape | What families of AI chips exist, what does each bet on, and how do I compare them? |
| 9.b | AMD Instinct | How are MI355X and MI455X built, and where do they beat NVIDIA? |
| 9.c | Google TPU | What is a systolic array, and how do 9,216 TPUs act as one machine? |
| 9.d | AWS Trainium and the Hyperscaler Chips | Why do AWS, Microsoft and Meta design their own chips, and what are they? |
| 9.e | SRAM Machines: Groq and Cerebras | How do chips without HBM decode so fast, and what do they give up? |
Facts were checked on 16 October 2026. Many of these products are announced well before anyone outside the vendor can measure them; each lesson marks which numbers are claims.