Solve the math behind AI. Earn your certificate.
Work through the real calculations a Tensor Core performs, from a single matrix cell to training a large model. Pass a level with 80% or more and download a certificate with your name on it.
Foundations
- Matrix multiply-accumulate
- Counting FMAs and tiles
- Rounding to FP4
- Peak throughput of a GPU
Engineer
- Trace of D = A × B + C
- FP8 throughput and 2:4 sparsity
- Arithmetic intensity
- INT8 quantization and attention
Architect
- GPU-days to train a model
- Roofline and memory limits
- KV cache and mixture-of-experts
- Determinant of A × B
Results
Get your certificate
Almost there
You need 80% to earn the certificate. Read the solutions above, then try again. Every attempt uses new numbers, so the method matters more than the answers.
Check that a certificate is real
Enter the name and certificate ID exactly as printed. Capital letters don't matter.
The rules
New numbers each time
Every problem is generated fresh when you start, so no two attempts are the same.
One try per problem
A correct answer earns 1 point, or half a point if you used the hint. You see the full solution either way.
Small rounding is fine
Decimal answers are accepted within a small margin. You can type commas or scientific notation like 2.5e9.
80% to pass
Pass a level to create a certificate with your name, score, time and a unique ID.
Study these guides before you start
Every problem in the challenge is explained in one of our guides.
What is a Tensor Core? →
Matrix multiply-accumulate, tiles and FMA counts. Covers most of Level 1.
Precision formats →
FP4 rounding, FP8 and INT8 quantization for Levels 1 and 2.
How a GPU works →
Arithmetic intensity and the roofline model for Levels 2 and 3.
Hopper vs Blackwell →
Memory sizing, bandwidth and throughput for Level 3.