amdresearch/aup-learning-cloud/plan-aup-learning-cloud-deployment

Plan an AUP Learning Cloud deployment

Turn a prospective adopter's needs into a concrete recommendation: how many AMD machines (Ryzen AI AIPC, Radeon workstation, or server) and how much networking gear to buy, which chips to pick, what cluster topology to use, and an IP/network plan — ending in a sizing table and a buyer-facing bill of materials (BOM). This is the pre-purchase advisory step that precedes the install/deploy skills.

The single measurable outcome: a defensible BOM + sizing/topology/network plan the user can act on. Full sizing math, the hardware-research method, the network decision table, worked examples, and the BOM template are in [reference.md](reference.md).

Prerequisites

• Web access (to look up the latest AMD silicon and confirm ROCm support).
• No cluster or checkout is required — this skill produces a plan, not a

running system.

• Helpful context: the AUP Learning Cloud

overview and quick start.

Phase 1 — Interview the requirements

Ask, and confirm back, before sizing anything:

1. Courses/toolkits wanted: Computer Vision, Deep Learning, LLM-from-scratch,

Physics Sim, and/or generic CPU/GPU + code-server. This drives both the GPU VRAM tier and which images to enable later.

1. Total headcount and the session pattern: a whole class on at the same

time (scheduled lab) vs self-paced/錯峰 usage.

1. Peak concurrent users, split into GPU sessions vs CPU-only sessions.

If the user only knows the total, estimate peak (see reference) and confirm.

1. Persistence/storage expectations (do notebooks need to survive reboots;

rough per-user disk).

1. Budget band and online vs air-gapped.

Phase 2 — Interview the network environment

1. How many routers, and how many subnets / CIDRs with which IP ranges.
2. Static IP vs DHCP; can a stable/reserved IP be given to one machine.
3. A managed switch and how many free ports (PoE not needed).
4. Internet access from the would-be service machine; any VLANs/firewalls.
5. Whether the machines are bare (can netboot) or will each get an OS.

Phase 3 — Research current AMD hardware

Do not rely on memory — web-search the latest AMD silicon and match it to the requirements:

1. Search current AMD options across form factors: Ryzen AI APUs (mini-PC /

laptop AIPC), Radeon workstation dGPUs, and multi-GPU workstations or servers. Compare by compute (CU/TFLOPs) and VRAM, not marketing tier.

1. Gate every candidate on ROCm support — if a chip is not ROCm-supported it

cannot run the GPU notebooks.

1. Map the chip to an existing chart accelerator key (phx, strix,

strix-halo, 9070xt, r9700, or 9600gre) and the expected amd.com/gpu.product-name node label. rdna4 is an installer detection fallback, not an existing chart accelerator key accepted by gen_configs.py. A new chart key requires configure-aup-learning-cloud-courses work before it can be generated.

1. Prefer multi-GPU chassis (workstation/server) when peak concurrent GPU

users is high enough that many single-GPU AIPCs become impractical to cable, power, and manage. Keep AIPCs for small labs and the demo-like experience.

Phase 4 — Size the cluster

The full formulas and per-notebook config table are in reference.md. The shape of it:

1. Concurrency, not headcount. Convert total users to peak concurrent

(~40-60% of total for self-paced; ~100% for a whole-class scheduled lab).

1. GPU drives machine count (whole-GPU, no sharing). Each GPU notebook in

AUPLC claims a whole, exclusive amd.com/gpu: "1" (request == limit); there is no time-slicing/MIG, and this is the same for every GPU course. So GPUs needed = peak concurrent GPU users. An APU box = 1 GPU; a workstation/server = N cards.

1. RAM/CPU sets the per-machine spec. CPU notebooks are best-effort and pack

densely (RAM-bound): RAM ≈ (concurrent users on the node × max mem/user) + overhead. Pick per-user memory from the course type (reference table).

1. VRAM picks the chip tier. Exclude 4GB iGPUs (780M/890M) for LLM/large

models; steer to Strix Halo (64GB) or R9700 (32GB) for those.

1. Add a control/service node. PXE/NFS/k3s-server overhead; small labs may

co-locate it on a GPU node (state the single-point-of-failure trade-off).

Phase 5 — Plan topology and network

1. Choose the topology (decision table in reference):

• Single-node (./auplc-installer) for one box / demo replica.
• PXE-diskless cluster for bare AIPCs on one flat L2 subnet that can

netboot (relies on the user's existing DHCP/router; the service machine needs a static IP).

• SSH-preinstalled cluster when nodes already have an OS or the network is

routed/multi-subnet.

1. Derive the switch-port count (≈ nodes + uplink) and whether the existing

router(s) suffice or a managed switch is needed.

1. Produce an IP plan: the static service-machine IP, the node subnet/CIDR,

gateway, and DNS — consistent with the topology you chose.

Phase 6 — Deliver the recommendation

Produce, for the user:

• A sizing table (peak concurrency → GPU count → machine count + the chosen

chip/VRAM, with the assumptions spelled out).

• A topology choice and an IP/network plan.
• A bill of materials: machine model + quantity + GPU, plus switch/router and

cabling, framed so the user can purchase (this is what leads to AMD hardware sales). Offer at least an AIPC-based option and a denser workstation/server option when concurrency is non-trivial.

• A handoff: point to install-aup-learning-cloud-single-node (one box) or

deploy-aup-learning-cloud (cluster) to execute, and configure-aup-learning-cloud-courses to enable the chosen courses.

Safety

• Advisory only. This skill plans; it does not install, buy, or change any

system. Never run installer/deploy commands from here.

• State every assumption (concurrency ratio, per-user memory, GPUs per

chassis) so the user can correct them before spending money.

• Always confirm ROCm support for any recommended silicon; never recommend a

chip you could not verify is supported.

• Flag single-point-of-failure trade-offs of all-in-one small labs, and

storage durability (local-path/NFS-on-one-box is disposable without backups).

Reference

Sizing formulas + per-notebook config table, the whole-GPU evidence, the hardware-research method, the network/topology decision table, worked examples, the BOM template, and the interview question bank: reference.md.