Exam · GPU memory maths for LLMs and SLMs

Take the exam and earn your certificate.

A real, server-scored exam for this course. Score strictly above 80% to pass and get a genuine, verifiable certificate.

Before you start

We ask for your name and email first, the same as a PDF download, so we know you are here and your certificate can carry the right name.

We privately record your name, email and request IP for this exam request. Read the privacy notice and retention policy.

Visit Mickai on LinkedIn

Complete the security check before submitting.

Cloudflare Turnstile helps prevent automated submissions. How this check uses data.

You can retake the exam any time. Each attempt draws a fresh set of questions.

Written practice, ungraded

Write your response in your own notes, then compare it with the sample answer. This practice is separate from the scored exam and does not affect your certificate.

Practice 1

Using the source Model B example, explain why adapters reduce training state and name one memory term they do not automatically remove.

Compare with a sample answer

The base weights are frozen, so they need no gradients or optimiser state. The adapters add far fewer trainable parameters, but activations still matter because gradients flow through the frozen layers. Checkpointing trades extra time for lower activation memory.

The source separates state savings from activation memory and describes checkpointing as a trade-off.

Practice 2

In the Model B table, compare adapters without checkpointing with adapters and checkpointing. State the estimated needs and the trade-off visible in the table.

Compare with a sample answer

The source estimates 10.50 GB without checkpointing and 3.56 GB with checkpointing. The latter uses a lower activation estimate, while checkpointing trades additional computation time for memory. These are workbook estimates, not a prediction for another model.

The table gives both estimates and the course describes checkpointing as a time-memory trade-off.

Practice 3

Describe the source fit-check sequence from an initial estimate to a measured result, including the safety comparison.

Compare with a sample answer

Record free memory and units, gather the model card fields, estimate weights, cache and any training states and activations, then add the allowance. Compare need with 90 percent of capacity or lower free memory. After loading, use the framework peak-memory observation to replace the planning allowance.

The source gives this ordered procedure and treats the allowance as provisional until measurement.

Practice 4

A planning estimate exceeds the usable memory limit. Give two different changes from the source table and describe what each changes or costs.

Compare with a sample answer

Shorter context or fewer simultaneous chats reduce cache but use less text or concurrency. Lower precision changes weight memory but may affect quality. Offloading moves data to system memory and costs speed. A smaller model changes everything but may reduce capability.

Each choice affects a different memory term and carries a corresponding cost in the source table.

Practice 5

Explain why the source does not present a single GPU as a route to production frontier-scale pretraining, even if memory can be split across devices.

Compare with a sample answer

One GPU may run models that fit, fine-tune small models and train toy models, but not pretrain a production frontier-scale model. Splitting divides memory approximately but adds communication and engineering demands. Data collection, cleaning, licensing and reliability work also remain.

The source names computational scale, communication, data work and engineering as separate limits.

Practice 6

What assumptions sit behind the workbook’s training-time illustration, and how should a reader make it more relevant to a particular machine?

Compare with a sample answer

The illustration uses about six operations per parameter per training token, a rough token-to-parameter guide, and a made-up sustained throughput rate. Replace the rate with one measured on the intended machine. The resulting time still excludes data and engineering constraints.

The workbook identifies these as source-paper estimates or assumptions, then advises measuring throughput and notes additional constraints.

About this exam

Is this exam graded?

Yes. Every question is scored on our server, never in your browser, and you need a score strictly above 80% to pass. The questions are a real, reviewed bank for this course, drawn fresh for each attempt.

Is it only multiple choice?

Yes. This is a multiple-choice knowledge assessment, with at least one question from each chapter in every attempt. Written exercises are separate practice and are not automatically scored. Passing does not establish practical professional competence or independently verify identity.

Is there a time limit?

Yes, 30 minutes, shown as a countdown once you start. If time runs out, whatever you have answered is submitted automatically.

Do I need to download the PDF first?

No. This course's full text is on its own page, so you can read it online and come straight here when you feel ready.

What do you do with my name and email?

We privately record your name, email, request IP, course and privacy acknowledgement for this request. Newsletter consent is separate and optional. See /privacy.html for storage and retention. A passed exam creates a server-stored named certificate. An email confirmation is not identity assurance.

Can I retake it if I fail?

Yes, straight away. Each attempt draws a fresh set of questions and shuffles the answer order, so it will not be identical to your last attempt, and an attempt can only be submitted once.

Is the certificate real?

It is a genuine, server-issued certificate with its own verification page and a unique ID, not a self-declared one. It records that you passed this course's exam on this date. It is not an accredited qualification, a professional licence or proof of employment.