COMPUTE.md

# ABOUT COMPUTE.MD

COMPUTE.md is a working idea: a plain markdown file that says what a workload needs from the cloud, part by part. GPUs and how much memory each one has, CPU cores and RAM, fast local disks, a shared NAS for datasets and checkpoints, object storage for everything else, the network between machines, and the databases and caches a SaaS app leans on. An AI agent reads it before it picks machines, holds the list up against what a cloud, a GPU rental market or a hosting provider actually offers, and says plainly where something falls short: too little GPU memory for the model, a disk too slow to feed it, no fast link between the nodes of a training run.

## WHAT IT COULD COVER

• - GPUs: model, count, memory per GPU, and whether the GPUs in a machine are linked
• - CPU and RAM: cores, memory size, and whether the memory is ECC
• - local disks: NVMe scratch size and speed, and whether it is wiped when the machine stops
• - shared storage: an NFS or SMB share (a NAS) mounted on every node, with its size and throughput
• - object storage: buckets for datasets, model weights, checkpoints, logs and user uploads
• - network: bandwidth between nodes, a fast fabric for multi-node training, and traffic out to the internet
• - databases and caches: a relational database, a cache, and a vector index for search
• - serving: a load balancer, containers or machines that scale with traffic, and a CDN in front

## HOW IT COULD WORK

• - a team writes one section per workload: training, fine-tuning, inference, the web app
• - sizes come from the work itself: model weights plus context cache in GPU memory, the dataset on shared storage
• - the agent turns each line into a machine type, a volume or a bucket from the provider's own catalog
• - it shows the match line by line, with gaps marked, before anything is created
• - the file lists needs, never secrets: no keys, passwords or account numbers

## THINGS IT COULD NAME

• - NVIDIA H100 and H200, AMD Instinct MI300X: data center GPUs with 80, 141 and 192 GB of high-bandwidth memory
• - NVLink and InfiniBand: fast links between the GPUs in one machine, and between machines
• - NVMe: the interface behind fast local SSDs
• - NFS and SMB: the file-sharing protocols a NAS speaks; Amazon EFS, Google Cloud Filestore and Azure Files are managed versions
• - the S3 API: the object storage interface many clouds and tools speak
• - Kubernetes: schedules containers onto machines, with device plugins that hand out GPUs
• - PostgreSQL with pgvector, and Redis: the database, vector search and cache many SaaS apps start on

## WHO IT COULD HELP

• - an ML engineer: a fine-tune that needs 8 GPUs with 80 GB each and the dataset on a shared mount
• - a SaaS startup: a web app, a database, a cache, a bucket for uploads and a CDN
• - a research lab: one list, checked against two clouds and a GPU rental market
• - a solo developer: an open model served on one GPU with room for its context

Most cloud mistakes in AI work are mismatches: a model that does not fit in GPU memory, a training run starved by a slow disk, nodes with no fast link between them, a database sized for a demo. Writing the needs down part by part makes them easy to check before anything is rented.

## A SKETCH

```md
---
compute: draft
---
# COMPUTE.md
## Fine-tune, a made-up example
- gpu: 8 per machine, 80 GB each, linked inside the machine
- cpu: 64 cores, 1 TB RAM, ECC
- scratch: 4 TB local NVMe, fine to lose when it stops
- shared: 20 TB NFS share on every node, for the dataset
- objects: one bucket for checkpoints and logs
- network: one machine, so no InfiniBand needed
## Web app
- app: 3 small machines behind a load balancer
- db: PostgreSQL with pgvector, 200 GB, daily backups
- cache: Redis, 8 GB
- objects: a bucket for uploads, behind a CDN
- region: one, close to most users
```

## DISCLAIMER

COMPUTE.md does not rent, price, launch or stop anything, and it has no prices. It lists needs; a provider's catalog, quotas and contracts decide what is really available, and hardware specs change. The sizes in the sketch are made up, not advice. Products and protocols named here are described for reference only, with no affiliation or endorsement. Keep keys, passwords and account numbers out of the file.

COMPUTE.md is an informational page. It is provided as is, without warranty of any kind, and may be incomplete, inaccurate or out of date. Nothing here is financial, investment, legal, tax, insurance, medical, engineering, safety or other professional advice, or an offer or solicitation of any kind. The sketch is an illustration, not a standard or a recommendation. Names of third-party products or companies, if any, are used only to describe them; no affiliation or endorsement is implied. Use at your own risk. This page sets no cookies and runs no scripts or analytics; like any website, its host keeps standard request logs.