Quiz: understanding HPC

Reading mode

EngineerDecision-maker

Ten questions to check what you remember from the course. Each answer is marked straight away, with an explanation and a link to the relevant lesson. Nothing is sent anywhere: your best score stays in this browser.

1. Question 1 of 10In a cluster, which component acts as the "air traffic controller", managing the queue, priorities and resource allocation?One answer

   • Parallel storage
   • The scheduler
   • The interconnect
   • The login node

   Check

2. Question 2 of 10Which of these are among the five cluster components presented in the course?Several answers possible

   • Compute nodes
   • A shared spreadsheet for booking machines
   • The interconnect
   • Parallel storage

   Check

3. Question 3 of 10A user wants to run a computation on the cluster. What does she connect to?One answer

   • Directly to the least busy compute node
   • To the parallel storage server
   • To the login node, from which she submits her job to the scheduler

   Check

4. Question 4 of 10A program has 5% of its work that cannot be split. According to Amdahl's law, what is the maximum speedup you can expect by adding processors?One answer

   • An unlimited speedup, proportional to the number of processors
   • A speedup of 5 at most
   • A speedup that will never exceed 20

   Check

5. Question 5 of 10Which parallelism patterns require regular exchanges between machines during the computation?Several answers possible

   • Tight coupling (MPI), for example an atmospheric simulation
   • Data parallelism, for example training an LLM
   • Embarrassingly parallel work, for example converting a batch of images

   Check

6. Question 6 of 10Bridge: organizationYour finance department wants to know which team uses what on the shared cluster. What mechanism can you rely on?One answer

   • A self-reported estimate requested from each team at year end
   • The accounting kept by the scheduler, per project or per user
   • The amount of data sitting on parallel storage
   • The computing centre's electricity bill

   Check

7. Question 7 of 10Bridge: businessDuring an AI training run lasting several days, a single GPU silently degrades and a checkpoint is lost. What consequence does the course describe?One answer

   • None, since the other GPUs compensate automatically
   • A loss limited to a few minutes of computing
   • A loss that can run to hours or even days of computing, on a training run that consumes, in the lesson's illustrative example, 10,752 GPU-hours

   Check

8. Question 8 of 10Bridge: business linesIn which fields does the course cite HPC as a working tool?Several answers possible

   • Finance and risk, for valuing complex portfolios and regulatory stress tests
   • Weather and climate, for global forecasts several times a day
   • Life sciences, for genomics and medical imaging
   • None outside academic research

   Check

9. Question 9 of 10What difference in numerical precision sets AI apart from classic HPC?One answer

   • AI tolerates 16-bit or even 8-bit, while classic HPC often requires double precision (FP64)
   • AI always requires higher precision than simulation
   • There is no difference; both use exactly the same formats

   Check

10. Question 10 of 10Bridge: peopleWhy is observing an AI training run becoming a specialty in its own right for operations teams?One answer

    • Because AI uses neither a scheduler nor parallel storage
    • Because AI adds its own indicators, such as the loss curve, step time or GPU utilisation per rank
    • Because a training run never lasts more than a few minutes

    Check

Start over

Going further

• Course overview
• Glossary
• Bridges
• HPC observability foundations

BridgesWhat this page changes beyond the technical team.

BusinessValue, cost, risk

Be able to explain why an hour of GPU cluster time is expensive and what a training run interrupted for lack of observation costs.

OrganizationWho decides, who owns

Know that the scheduler accounts for usage per project and per user, the basis of fair sharing and chargeback.

PeopleTeams, on-call, skills

Recognise that observing an AI training run is becoming a skill in its own right within operations teams.

Business linesProduct, finance, risk, customer

Be able to work with computing teams as a project manager, buyer or funder, in finance, health, energy or weather.

Why bridges matter →