# NVIDIA systems: a workload-to-procurement lab

Original LockedIn FDE practice. Start with the [visual NVIDIA learning path](/training/nvidia).
Use Python 3.11 or later. No GPU, account, package installation or cloud purchase is required.
The case, prices and memory allocations are synthetic. This is a planning and reasoning lab,
not a GPU benchmark, vendor certification or authorization to buy hardware.

## 1. Map the system before choosing a product

The fictional Field Desk drafts maintenance responses using approved procedures. It serves
one authenticated team, keeps data in an approved location, and asks a person to accept a draft.
It must refuse cross-tenant data access and return a controlled response when capacity is exhausted.

Draw application → model service → GPU execution. Put the approved source store beside the
application and a human decision after the draft. Place these labels at their responsibilities:
GPU, CUDA, Nemotron, NeMo and NIM. NeMo development tools do not have to sit in every request path.

Keep a decision packet with: task, component, exact version, alternative, evidence needed and owner.
Explain the diagram to someone before using an acronym. They should be able to point to the model
and the hardware separately.

## 2. Run the capacity boundary

Download [nvidia-capacity-lab.py](/templates/nvidia-capacity-lab.py) into an empty working folder.

```shell
python3 --version
python3 nvidia-capacity-lab.py test
python3 nvidia-capacity-lab.py memory --concurrency 4
python3 nvidia-capacity-lab.py memory --concurrency 10
```

The default weight estimate is 16 decimal GB (about 14.901 GiB). At concurrency four, the
total including fixed overhead and safety reserve is 37.6 GB against 48 GB. At concurrency ten,
it is 49.6 GB and the estimate is refused. The per-request allocation is a supplied assumption;
it is not derived from a real Nemotron architecture. A refused estimate is a successful boundary
demonstration, so the CLI still exits normally and reports `fits_estimate: false`.

Change `--bits 8`, then `--request-gb 4`. Explain what changed and what did not. Lower stored
precision changes arithmetic but does not establish acceptable quality. Actual memory depends
on the selected model, precision, runtime, context and concurrency. The program models one device;
adding several devices requires a placement and interconnect design, not simply adding capacities.

Produce two outputs and a paragraph naming which inputs need real measurements.

## 3. Inspect an environment without pretending to have a GPU

On a host you are authorized to inspect, `nvidia-smi` can show devices and driver information.
Separately record the exact container digest, framework version and device execution check.
If you do not have such a host, mark these observations as unverified. Do not install drivers,
allocate resources or invent a transcript for this exercise.

The current [CUDA guide](https://docs.nvidia.com/cuda/cuda-programming-guide/index.html)
and the chosen runtime's support matrix define compatibility. A host driver command does not
prove that a framework inside a container is executing on the GPU.

## 4. Select a model and an improvement method

Use the [Nemotron directory](https://developer.nvidia.com/topics/ai/nemotron) to choose a
specific model card. Record the exact identifier/revision, modalities, context, license and
runtime requirements. Leave performance unmeasured until it is actually measured.

Write five cases: an ordinary request, missing evidence, conflicting procedures, a foreign-tenant
request and an adversarial instruction embedded in a document. Keep expected behavior and data
boundaries fixed when comparing candidates.

Now diagnose three failures: stale knowledge, inconsistent formatting and unauthorized tool use.
Propose the simplest relevant repair. Read the [NeMo suite map](https://docs.nvidia.com/nemo/)
and explain which specific component, if any, would help an improvement experiment. Keep held-out
cases separate from the material used to improve the system. Guardrails do not replace server authorization.

## 5. Specify the private model service

Use the current [NIM documentation](https://docs.nvidia.com/nim/index.html) to record a
model/runtime candidate. Identify the authenticated caller, source boundary, memory admission
rule, queue limit, timeout, retry budget and cancellation behavior. Do not put secrets into the packet.

Name readiness checks separately from task-quality checks. Include a permitted ordinary request
as well as a denial. Write what happens under overload and who owns rollback. TensorRT LLM,
Dynamo and NIM have distinct responsibilities; verify the exact release's supported combination.

## 6. Compare renting and owning

```shell
python3 nvidia-capacity-lab.py compare --hours 160
python3 nvidia-capacity-lab.py compare --hours 600
```

These hypothetical equivalent-capacity bundles use a rate of 8 currency units per hour,
an 80,000-unit purchase over 36 months, and 800 units of monthly operation. Rental is 1,280
at 160 hours and 4,800 at 600 hours. The ownership estimate is 3,022.22 per month; the
arithmetic crossover is approximately 377.78 hours. This is not a current NVIDIA price.

Replace assumptions only with comparable, dated evidence. Include storage, networking, software,
support, electricity, cooling and staff in the operating-cost assumption. The simple formula
excludes taxes, financing, residual value, migration and unmodeled downtime. Unknown costs stay
unknown in the decision packet; they are not evidence of zero cost.

Add a managed endpoint and a smaller pilot as alternatives. State one change in usage or operating
requirements that would reverse your recommendation.

## 7. Prepare a complete quote and acceptance request

Use the [official where-to-buy directory](https://marketplace.nvidia.com/en-us/enterprise/data-center/where-to-buy/)
and [certified-system information](https://www.nvidia.com/en-us/data-center/products/certified-systems/)
as research starting points. Do not place an order as part of this lab.

Ask for exact SKU, GPU count and memory, CPU/RAM, interconnect, NICs, storage, rack/power/cooling,
software entitlement, warranty, support, installation, delivery date and acceptance conditions.
Distinguish a chip/card quote from a server or rack quote. Confirm provider region and capacity terms
for rental. A product announcement is not a local availability guarantee.

Current-generation references checked 19 October 2026:
[Blackwell and Blackwell Ultra](https://www.nvidia.com/en-us/data-center/technologies/blackwell-architecture/)
and [Vera Rubin](https://www.nvidia.com/en-us/data-center/technologies/rubin/).
Compare supported precision, memory, bandwidth, topology and the complete system. Recheck
documentation and supplier availability before making a purchase decision.

End with a one-page recommendation: workload, chosen option, alternatives, estimated costs,
observed evidence, unresolved assumptions, acceptance workload, receiving owner and next decision.
The evidence should be understandable without the presenter supplying missing facts.