ML Engineering interview question · Serving and inference · mid level
How does request batching on a GPU trade latency for throughput?
Free to read in full. This is 1 of the 150 questions in the ML Engineering and MLOps interview pack, with 193 diagrams. Get the pack, ₹999 · $20
Why it is asked
Batching is the main lever for GPU serving cost. The interviewer checks that you know why it works, where it stops helping, and how to tune it against a latency target.
The answer
Batching makes each request wait a little longer so the GPU can do far more work per second. A forward pass has a fixed cost (launching GPU kernels, the small programs the GPU runs, and reading the weights from memory) that is paid once per batch, so sharing it across many requests raises throughput, while each request now waits for the whole batch. A simple stated model shows the shape. Assume one forward pass takes 8 ms plus 0.5 ms per item: Going from 1 to 32 gives 11.3 times the throughput for 2.8 times the batch time. Throughput can never pass 2,000 per second (one item every 0.5 ms), so the gains flatten while latency keeps rising. Latency also includes waiting for a batch to form. Dynamic batching, offered by servers such as NVIDIA Triton and TorchServe, collects requests until the batch is full or a short maximum wait, for example 2 to 5 ms, runs out, then runs what it has. At low traffic the timer caps the extra delay; at high traffic batches fill fast. To tune it, fix the latency SLO (service level objective, the target you promise) first, then raise the maximum batch size and wait time while p99 stays under it, and stop when throughput stops rising. For large language models, continuous batching goes further: new requests join a running batch at each generation step instead of waiting for the batch to finish.
- batch 1: 8.5 ms, about 118 predictions per second
- batch 8: 12 ms, about 667 per second
- batch 32: 24 ms, about 1,333 per second
What the interviewer listens for
- Fixed per-batch cost is shared, so throughput rises with batch size.
- Batch time and queue wait rise, so per-request latency grows.
- Gains flatten near the per-item limit (here 2,000 per second).
- Dynamic batching: max batch size plus max wait; tune against p99.
- LLMs use continuous batching at each generation step.
Follow-ups the interviewer may ask
- Why does batching help less on a CPU? Hint: less parallel hardware sits idle at batch 1, so the fixed cost share is smaller; it can still help.
- What limits batch size on a GPU? Hint: memory for activations (and, for LLMs, the KV cache that stores attention values for the text so far) and the latency target.
- Different input lengths in one batch? Hint: padding wastes compute; bucket by length or use continuous batching.
Common mistakes
- Maximizing batch size without a latency target.
- Forgetting the queue wait while a batch fills.
- Expecting throughput to grow forever with batch size.
Written by Roni Das, who builds distributed systems and AI systems for a living.
Learn it properly
This is taught step by step in the course lesson Batching Requests: Does It Help a Small CPU Model?.
One question is not enough for the interview.
This is 1 of the 150 questions in the ML Engineering and MLOps pack
Each question has why it is asked, a full answer, what the interviewer listens for, the follow-ups, the common mistakes and a diagram (193 diagrams in all). A PDF you keep, junior to staff level.
India: UPI, cards or netbanking through Razorpay. Elsewhere: PayPal. Download at once from My packs.
More free questions from this pack
- A model scores 95% accuracy offline but users say it is useless. What do you check first? (mid)
- How do you train on a dataset that does not fit in memory? (junior)
- A long distributed training job dies at hour 20. How do you make training fault tolerant? (senior)
- An online model metric dropped 8% overnight. How do you debug it? (staff)
Free questions from the other packs
- What do temperature and top-p control, and what values would you use for data extraction versus brainstorming? (AI Engineering)
- What is hybrid search, and how do you merge keyword and vector results with Reciprocal Rank Fusion? (AI Engineering)
- Users ask "list all X" questions and RAG returns only five items. Why, and what do you do? (AI Engineering)
- How do you check that an LLM judge agrees with human reviewers? (AI Engineering)
- Cut the cost of an LLM feature serving 10 million requests a day by 60% without hurting quality. Walk through your plan. (AI Engineering)
- Why are LLMs bad at counting letters or reversing a word? (GenAI and LLM)
- Write out scaled dot-product attention with the shape of every tensor. (GenAI and LLM)
- How does speculative sampling keep exactly the target model's distribution? (GenAI and LLM)
- What is GRPO, and how do verifiable rewards train reasoning models? (GenAI and LLM)
- Design a system that answers questions from 300-page contracts accurately. (GenAI and LLM)
- Describe the basic agent loop. (Agentic AI)
- What are pass@k and pass^k, and why do they matter for agents? (Agentic AI)
- How does prompt injection become dangerous once the agent has tools? (Agentic AI)
- Design the security architecture for an agent that can read and send company email. (Agentic AI)
- Design a customer support agent for an e-commerce company. (Agentic AI)
- What happens, step by step, when an LLM server handles one request? (Inference Engineering)
- Derive the KV cache size formula and work it out for Llama 3 70B. (Inference Engineering)
- What is PagedAttention? (Inference Engineering)
- When does speculative decoding stop helping or make things worse? (Inference Engineering)
- Design a service to serve Llama 3 70B at 10 QPS with latency targets. (Inference Engineering)
- You need to show delivery times for 20 cities. Would you use histograms, box plots, violin plots or ECDFs, and why? (Data Science)
- A test is 99% accurate and you test positive. What is the chance you have the disease? (Data Science)
- What share of first-time buyers make a second purchase within 30 days? Write the SQL. (Data Science)
- How would you forecast daily orders for the next four weeks, and how do you know the forecast is any good? (Data Science)
- A coupon A/B test shows a small average lift. How do you find which users the coupon actually works for, and who should get it next time? (Data Science)
- Compute the cosine similarity of a = (3, 4) and b = (4, 3) by hand. What does the result mean? (Maths for ML)
- A batch of 32 MNIST images goes into a dense layer with 128 units. What are the shapes, and what is the rule for when two matrices can be multiplied? (Maths for ML)
- A disease affects 1% of people. A test catches 90% of sick people (sensitivity) and wrongly flags 9% of healthy people (false positive rate). You test positive. What is the chance you are sick? (Maths for ML)
- Why do we not use mean squared error for classification with a sigmoid output? Compare the gradients numerically. (Maths for ML)
- Compute scaled dot-product attention by hand for two tokens with d_k = 2. (Maths for ML)
- A customer says: "We want AI for our support team." What do you do in the first meeting? (Forward Deployed Engineer)
- The customer's admin offers you a global admin account "to save time". What do you ask for instead, and why? (Forward Deployed Engineer)
- Your connector writes tickets into the customer's system. A request timed out, the connector retried, and now there are duplicate tickets. How do you make the writes safe to retry? (Forward Deployed Engineer)
- The customer wants RAG over 10 million documents, all behind their firewall. Size it for me on the whiteboard. (Forward Deployed Engineer)
- One deployment of your AI platform must serve 50 business units of the same enterprise. How do you design multi-tenancy? (Forward Deployed Engineer)