Community resourceWorksheet
OCR H446 1.1.2 GPU workloads, SIMD and uses
Part 2 of 5 · H446 1.1.2 · Types of processor
OCR H446 1.1.2: GPU workloads, SIMD and uses. Students learn to select CPU or GPU processing from workload shape and explain SIMD benefits through repeated operations over independent data, including suitable machine-learning stages.
Students will:
- classify workloads for CPU or GPU processing
- explain SIMD through operations on independent data
- apply workload reasoning to varied uses, including machine-learning stages
Inside: 5 explanation cells, 1 multiple-choice question, 2 fill-in-the-blanks cells and 4 written answers. 38 marks, about 50 to 65 minutes.
Series: H446 1.1.2 · Types of processor, part 2 of 5.
Shared by Coding PathwayVerified teacher
- 12 cells
- About 60 minutes
- CC BY-SA 4.0
- Shared 31 Aug 2026
- Updated 3 Sept 2026
Preview
The whole resource, exactly as a class sees it. Answers and marking are held back.
GPU workloads, SIMD and uses
A GPU is a specialised processor originally associated with graphics but also useful for non-graphics work. Its value comes from the structure of the workload, not from a slogan that GPUs are always faster.
CPU and GPU optimise for different work
A CPU is a flexible general-purpose processor suited to varied, complex and sequential control work. A GPU contains many specialised processing units and instructions designed for high throughput. In SIMD, one instruction applies the same operation to multiple data items. This works best when many items can be processed independently.
Worked application from two angles
Graphics: applying the same colour/lighting transformation to millions of independent pixels or vertices exposes large amounts of similar work. A GPU can process many items concurrently, supporting higher frame rates, resolution or rendering responsiveness.
Non-graphics modelling: a weather model may perform the same update calculation for many independent grid cells at one stage. That repeated operation can use GPU lanes. Later stages with dependencies, branching or data exchange may stay on the CPU or require coordination.
Machine learning and AI: training or running a model can involve repeated matrix and vector calculations across a large batch of data. Many similar calculations can be carried out concurrently by GPU processing units. However, “AI” does not automatically mean “use a GPU”: a small model, branching control task, sequential stage or frequent transfer between CPU and GPU may limit the benefit.
A complete explanation identifies the repeated operation and independent data, explains how the GPU can process many items concurrently, and then states the useful result in that situation.
Which workload is most naturally suited to a GPU?
- AChoose the next operating-system action through many dependent branches
- BApply the same filter calculation independently to every pixel in a large image
- CRun one short sequential control loop with each result needed by the next step
- DWait for one network response before deciding what to do
- units
- instruction
- independent
- sequential
- optical
Choose CPU, GPU or a mixed CPU+GPU design for each workload. Give one workload-based reason for every choice.
- Render many independent particles.
- Control a lift using many conditional safety checks.
- Train a machine-learning model using repeated matrix calculations across large batches of data.
- Process a sequence in which each result changes the next operation.
- Coordinate a game while also rendering its frames.
Use operation similarity, data independence, branching, sequential dependence and coordination. A large dataset or the label ‘AI’ is not, by itself, a reason to choose a GPU.
Students type their answer here.
Explain two benefits of using a GPU to process and display graphics in an interactive simulation. For each, connect the mechanism to a visible or timing consequence.
Generic ‘more cores’ earns no applied benefit by itself.
Students type their answer here.
Apply the workload model more widely
Use workload shape and SIMD reasoning in non-graphics contexts, then test the model against a large but dependency-heavy counterexample.
Choose one machine-learning task and one scientific-modelling or data-analysis task that could use a GPU. For each task:
- identify a repeated operation;
- identify the data items that can be worked on concurrently;
- explain why part of the wider application may still need a CPU or may gain little from the GPU.
For machine learning, you might consider repeated matrix or vector calculations during training or inference. Focus on workload structure rather than assuming that every AI system benefits from a GPU.
Students type their answer here.
A dataset is enormous, but record 2 cannot be processed until record 1 changes a shared state, and so on. Explain why ‘large data’ alone does not justify a GPU recommendation and state what redesign or condition would be needed.
Separate data volume from available parallelism.
Students type their answer here.
Consolidate your understanding
Check that every GPU choice identifies a repeated operation, many independent data items and why concurrent processing is useful in that situation.