← Course contents · Part 4

· Visual lesson · Session 36 of 56


Why is this job slow?

Before asking why the GPU is slow, ask which part of the job became slower. A timeline gives you somewhere specific to investigate.

Picture the idea

Find the added time

Before · 75 seconds

Prepare 20 · calculate 40 · communicate 10 · write 5

After · 96 seconds

Prepare 20 · calculate 41 · communicate 30 · write 5

Invented sequential phases. Communication accounts for 20 of the 21 added seconds; it is an investigation target, not a confirmed root cause.

Make the comparison fair

Keep input, model, precision, batch size, concurrency, and versions comparable. Separate first-run loading from steady operation.

In the fictional diagram, total time rises from 75 to 96 seconds. Communication accounts for 20 of the 21 added seconds.

That is the first investigation target, not proof of a bad network. A participant arriving late can also make others wait.

Test one explanation

Compare placement, link health, software changes, and phase timings. Write down what evidence would disprove your leading hypothesis.

Preserve logs before intervention. Resets and disruptive diagnostics require approval; they can interrupt work and erase evidence.

Try it yourself

Would a faster GPU explain or repair the communication increase?

Check your answer

No. It might shorten another phase while leaving the regression untouched. After a fix, repeat the original workload and check correctness as well as duration.

Sources and further reading

← All sessions