Semiconductor supply chain and capstone

Capstone title: From Tokens to Throughput. You will trace one request from a tokeniser to a served token, then outward to the hardware supply chain that makes that path possible. Stay honest. Do not invent shipment dates, market shares, or "Reuters today" claims. Name the parts of the chain you can defend.

Hardware supply chain, operator view

• GPUs. A training or serving chip is a lot of SRAM, tensor units, and a memory subsystem. You already estimated KV bytes. That estimate is why HBM capacity shows up in every serious serving plan.
• HBM. High-bandwidth memory sits next to the die. Bandwidth and capacity bound decode. When people say the model is memory-bound, this is the hardware sentence.
• Packaging. The die, HBM stacks, and interposer have to be assembled. Packaging yield is part of why accelerators are scarce even when a design exists. You do not need a fake yield number. You need the mechanism.
• Networking. Scale-up inside a node and scale-out across a rack. Training all-reduce and multi-GPU serve both care. A pretty GPU with a weak fabric is a paperweight.
• Power and data centers. Megawatts, cooling, and lead time for substations. Tokens per second at the product layer sit on joules and permits at the facility layer.

Connect this to week 9 and week 11. Tensor parallel is a packaging and networking decision. Checkpoint size is an HBM and storage decision. A sold-out SKU is a supply-chain decision that your scheduler cannot negotiate with.

Capstone artifact

Write a single document, 800 to 1500 words, with four sections:

1. The model replica you actually trained (size, tokens, what broke).
2. The serve replica (TTFT, inter-token, KV bytes, batch).
3. The systems replica (how you would run it on 1, 8, and 64 GPUs, and how it fails).
4. The hardware replica (GPU, HBM, packaging, network, power) with no invented statistics.

End with earned opinions. An earned opinion cites a measurement you took or a mechanism you can derive. An unearned opinion cites a keynote vibe.

Assignment

• Ship the capstone document in week12_capstone/from_tokens_to_throughput.md.
• Link the seven core project folders. If a project is thin, say so. Do not backfill fiction.
• Re-read week 1's principle. If you still think you needed an 800B pretrain to understand the shape, write why. If you do not, write what the replicas taught you that a tour would not.

Opinion checkpoint

You are done when the document exists and the repos run. Not when the adjectives are impressive.