- Service code
- BH-RES-03
- Family
- Research
- Cadence
- Per engagement; re-benchmarked on model, traffic or hardware change
- Response
- Benchmark plan agreed within 5 business days of engagement start
Overview
Inference cost is the number that decides whether a product is viable, and it is almost never the number in the vendor deck. Real serving cost depends on your sequence-length distribution, batch composition, cache hit rate, latency target and tail behaviour — interactions that aggregate benchmarks hide. This product measures inference economics on your actual workload and traffic shape, and attributes cost to the levers that move it.
What is covered
- Throughput and latency characterisation per accelerator on your model family and quantisation
- Cost per served token and per request, at target percentile latency rather than mean
- Batching and continuous-batching efficiency: how much of theoretical throughput you actually realise
- Prefix and KV-cache hit-rate impact on cost, with cache-strategy recommendation
- Autoscaling behaviour under diurnal and bursty traffic, including scale-down lag cost
- Model-size and quantisation trade studies against quality targets you define
- Hardware comparison on identical workloads: accelerator, interconnect, memory bandwidth and rack density effects
Deliverables
- Benchmark report: throughput, latency percentiles and cost per token by configuration
- Attribution analysis — which lever moves cost most for your traffic shape
- Configuration recommendation with modelled cost reduction and quality impact
- Hardware comparison matrix on identical workloads
- Capacity model translating traffic forecast into accelerator requirement and cost
Who this is for
Teams serving models at scale whose unit economics do not match their vendor benchmarks.
Specification
- Measurement basis
- Percentile latency at target SLO, not mean throughput
- Workload fidelity
- Your sequence-length and prompt distribution, replayed
- Cost basis
- Per served token, per request, per concurrent session
- Levers analysed
- Batching, caching, quantisation, autoscaling, hardware, topology
- Output
- Benchmark report, attribution, configuration and capacity model
Related services
| Service | Scope |
|---|---|
| GPU Market Pricing Research | What a GPU fleet actually costs, priced line by line, and where the rate card misleads. |
| Workload Characterisation Lab | Establish what your workload actually is, so every downstream sizing decision stops being a guess. |