Skip to article frontmatterSkip to article content
Site not loading correctly?

This may be due to an incorrect BASE_URL configuration. See the MyST Documentation for reference.

What each tier of machine can and cannot do, and what things cost.


The tiers

TierExampleGets you through
FreeColab, Kaggle T4Steps 0 to 5
ConsumerOne 24 to 32 GB GPUSteps 0 to 19
WorkstationDGX Spark, 128 GB unifiedEverything, comfortably
RentedA100 or H100 by the hourEverything
Cluster8 GPU nodeEverything, fast

Renting

Rent before you buy. You will learn what you actually need, which is almost never what you first assumed.

Rough costs: a strong single GPU runs about 1 to 3 US dollars per hour on the common providers. Spot and preemptible instances are cheaper if your job can survive being interrupted — and with checkpointing from Step 9, it can.

Budget for the whole book on rented compute: roughly 200 to 600 dollars, if you are disciplined about shutting instances down. Set a billing alert on day one.


Notes for the NVIDIA DGX Spark

The DGX Spark is a desktop machine built around the GB10 Grace Blackwell superchip, with 128 GB of unified memory shared between CPU and GPU, and very high theoretical throughput at low precision.

It is an unusually good fit for the path this book takes, with one real weakness.

Where it is excellent

Anything limited by memory capacity rather than memory speed. 128 GB of unified memory means you can load models that simply will not fit on a 24 GB consumer card, and you can do it without splitting the model across devices.

Continued pretraining and QLoRA fine-tuningStep 11 and Step 12. This is the machine’s sweet spot. A 7B to 14B model with QLoRA runs comfortably, and tools like Unsloth roughly double the speed and halve the memory again.

Local inference of large models. You can serve something substantial without a cloud, which matters if privacy is part of your project’s point.

Iterating without a meter running. The value of not watching a per-hour bill while you experiment is easy to underestimate.

Where it is merely fine

Steps 1 through 10. Your toy model does not need this machine, and you should not wait for one to start.

Where it is genuinely weak

A realistic plan on a DGX Spark

StageTime
Data audit and cleaning (Steps 6 to 7)Days
Tokenizer work (Steps 3 to 5)Days
Vocabulary extension plus QLoRA continued pretraining on a 7B modelHours to a few days
Instruction tuning (Step 12)Hours
Evaluation and a basic chat interfaceDays
Full working prototype1 to 3 weeks of focused work

That is a realistic schedule for one person. It is also why this project is a “few weeks” project rather than a “years of cluster time” project.


What to check before you commit

  1. Does your framework support your hardware properly? Check before buying, not after.

  2. Does the specific quantization method you want work on it?

  3. Can you actually cool it where you plan to put it?

  4. Is your electricity supply adequate?