AWS Neuron vs vLLM Serving Benchmark Harness with OpenAI Agents SDK
Deploying models safely across heterogeneous cloud environments requires benchmarking AWS Trainium/Inferentia against vLLM GPU targets. You will build an automated benchmarking agent using the OpenAI Agents SDK to execute parallel workload tests, measure throughput (tok/s), and calculate cost per million tokens. The agent must generate an auditable trade-off matrix with less than 3% execution variance.
What you are building
The core problem, expected build, and operating context for this challenge.
Build an automated cross-hardware profiling agent using the OpenAI Agents SDK to evaluate AWS Neuron and vLLM target workloads.
How work is evaluated
Evaluates benchmarking agent's capability to compare AWS Neuron and CUDA endpoints accurately.
Challenge datasets
Public datasets, plus the private datasets you have access to.
How submissions are scored
These dimensions define what the evaluator checks and which criteria separate a passable run from a strong one.
math_correctness
Verifies calculated cost savings percentage strictly matches ground truth ratio
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
analysis_accuracy
Accuracy of cost and throughput comparison calculations • target: 0.99 • range: 0-1
This dimension contributes its full weight only when the submission satisfies the requirement. Partial credit is not awarded.
What you should walk away with
Construct OpenAI Agents SDK function calls to manage AWS Neuron and vLLM benchmarking pipelines
Execute reproducible synthetic token generation benchmarks across target instance types
Collect and normalize latency metrics (TTFT and inter-token latency ITL)
Synthesize an auditable deployment recommendation report
Reference links and supporting material
Latency and throughput measurements recorded across AWS Inf2 and NVIDIA GPU instances.
Challenge setup
Evaluates benchmarking agent's capability to compare AWS Neuron and CUDA endpoints accurately.
Challenge input
JSON containing neuron_bench_data and vllm_bench_data with throughput, TTFT, and hourly_cost
OpenAI Agents SDK
Provides direct tool calling and execution loop orchestration.
Evaluated output
JSON returning optimal_target, cost_savings_pct, ttft_diff_ms, and trade_off_summary
- Verifies calculated cost savings percentage strictly matches ground truth ratio
- Accuracy of cost and throughput comparison calculations • target: 0.99 • range: 0-1
- Benchmark: Cloud-Inference-Cost-Bench
- Analysis Accuracy target: 0.99
- 1 public reference case
- Python execution harness
- Python sandbox (unavailable on Versalist)
Summary shows the first task, up to three tools, and up to three evidence items. Open the technical recipe to view all configured items.
View technical recipe
Configured tools
- OpenAI · Required
- Anthropic · Optional
- Azure OpenAI Service · Optional
Evaluation contract
- math_correctness · Weight 1
- analysis_accuracy · Weight 1
Metrics
- analysis_accuracy · Target 0.99
Benchmarks
- Cloud-Inference-Cost-Bench
- general
Reference cases
- 1 public reference case
Workflow tasks
- Task 1
Input: JSON containing neuron_bench_data and vllm_bench_data with throughput, TTFT, and hourly_cost
Output: JSON returning optimal_target, cost_savings_pct, ttft_diff_ms, and trade_off_summary
Execution harness: Python
Evaluation tools
- OpenAI Agents SDK · Agent Framework
Recipe freeze marker
No recipe freeze time is recorded.