Llm Inference Serving medium
You are buying accelerators for a chat workload with short prompts (~300 tokens) and long replies (~1,500 tokens). Option X offers 2x the FLOPs of Option Y at the same memory bandwidth. Option Y offers 1.6x X's memory bandwidth and more HBM. Which fits better?
- A Option Y, since decode dominates and is bandwidth-bound, and extra HBM holds more KV for larger batches.
- B Option X, since tokens per second scale with FLOPs, so doubling compute roughly doubles each stream's speed.
- C Option X, since continuous batching makes decode compute-bound, so the extra FLOPs are used at any batch size.
- D Either one, since serving throughput is set mainly by the scheduler and both run the same batching software.
Answer: A. Option Y, since decode dominates and is bandwidth-bound, and extra HBM holds more KV for larger batches.
This workload is mostly **decode**, which is bound by memory bandwidth, and more HBM lets more KV-heavy sequences batch together. Option C is partly true: very large batches push decode toward compute-bound, but KV reads grow with batch size and latency SLOs cap the batch well before that point. Option X would win for prefill-heavy workloads with long prompts and short outputs.
One system design question a day, with spaced repetition so the ones you miss come back.
Get the appRelated questions
- Your 70B-class model servers each handle about 40 concurrent chats with 4K-token contexts. Product raises the maximum context to 32K tokens, and many users start using it. What is the main capacity consequence?
- A 70B-class model (80 layers, grouped-query attention with 8 KV heads of dimension 128) is served in fp16 on one 8-GPU node with about 640 GB of total HBM. Product wants 256 concurrent sessions, each holding a full 32K-token context. What is the binding constraint, and what does it imply?
- Your server reserves a contiguous KV region sized for max_tokens = 8,192 for every request at admission, but the average sequence ends near 600 tokens. The node reports out-of-KV-memory at about 20 concurrent requests while GPU compute sits at 30%. Which change best raises concurrency without restricting legitimate long outputs?
- You must summarize 10 million documents overnight, and no user is waiting on any single result. How should you configure the serving stack compared with your interactive chat deployment?
- You serve a 7B-parameter model in fp16 (about 14 GB of weights) for a chat app. Users say single-conversation token generation is too slow, and you are choosing new GPU hardware. Which hardware spec most directly increases per-sequence decode speed?
- Your RAG workload sends 8K-token prompts and gets ~150-token answers from a 70B-class model. You plan to move prefill and decode onto separate GPU pools so long prefills stop stalling decode. What new resource must you size carefully?